PHILOSOPHICAL PSYCHOLOGY VOL 13 NO 3 2000

Natural epistemology or evolvedmetaphysics Developmental evidencefor early-developed intuitivecategory-speci c incomplete andstubborn metaphysical presumptions

PASCAL BOYER

ABSTRACT Cognitive developmental evidence is sometimes conscripted to support ldquonaturalizedepistemologyrdquo arguments to the effect that a general epistemic stance leads children to build theory-likeaccounts of underlying properties of kinds A review of the evidence suggests that what promptsconceptual acquisition is not a general epistemic stance but a series of category-speci c intuitiveprinciples that constitute an evolved ldquonatural metaphysicsrdquo This consists in a system of categoriesand category-speci c inferential processes founded on de nite biases in prototype formation Evidencefor this system provides a better understanding of the limited ldquoplasticityrdquo of ontological commitmentsas well as a computationally plausible account of their initial state avoiding ambiguities aboutinnateness This may provide a starting point for a ldquonaturalized epistemologyrdquo that takes intoaccount evolved properties of human conceptual structures

An ontology speci es kinds of stuff in the world There is now considerablepsychological evidence that mentally represented conceptual knowledge includesontological commitments The latter not only specify large categories of objects (egldquoartefactrdquo ldquoanimalrdquo ldquopersonrdquo ldquoeventrdquo or ldquonumberrdquo) but also produce spontaneousinferences from perceived states and events as well as predictions of future possiblestates and events These ontological commitments expectations and inferenceengines are crucial in the acquisition of new basic-level kinds (eg ldquotelephonerdquoldquogirafferdquo ldquochildrdquo ldquodozenrdquo) and they are now the object of ne-grained develop-mental accounts (see Hirschfeld amp Gelman 1994 for illustrations) How thesebroad ontological concepts and their associated inferential principles are acquiredand mentally represented is a crucial question for epistemology Indeed some recent

Pascal Boyer Centre National de la Recherche Scienti que France Institut des Sciences de lrsquoHomme14 Avenue Berthelot 69363 Lyon France tel 1 33 47 272 64 60 fax 1 33 47 272 65 90email pboyerish-lyoncnrsfr until 31 July 2000 ISBER 2201 North Hall University ofCalifornia Santa Barbara CA 93106 USA tel 1 1ndash805ndash893ndash7749 fax 1 1-805-893-7995email pboyerisberucsbedu

ISSN 0951-5089 (print) ISSN 1465-394X (online)00030277ndash21 Oacute 2000 Taylor amp Francis Ltd

278 P BOYER

arguments in ldquonaturalized epistemologyrdquo take their illustrations and derive some oftheir arguments from the experimental literature on conceptual development How-ever this use of the evidence is often misleading I will argue here that a moreprecise description of the developmental evidence suggests a more speci c picture ofearly ontology (and of its consequences for adult conceptual structure) This mayhelp construct a natural epistemology supported by computationally tractableempirically based accounts of our cognitive dispositions rather than by a prioriassumptions about what they should be It may also illuminate traditional debateson processes of concept acquisition on the conceptual continuity between infantsand adults and on the putative ldquoinnatenessrdquo of particular ontological commitments

Epistemic stance vs natural metaphysics

That our conceptual sensitivities are dramatically off target can be in principleexcluded on evolutionary grounds (Quine 1969) However this leaves open thepossibility of a vast range of sub-optimal designs whereby organisms survive on thebasis of false but innocuous beliefs A ldquosoundrdquo evolved cognition in this view wouldhave two main characteristics First it would be able to recognize relevant featuresof objects and identify their correlations Second it would go beyond surfacecorrelations and identify underlying structures that cause them

Developmental psychological evidence seems to bear good news for this versionof the naturalized epistemology project Studies of pre-schoolersrsquo induction andcategorization behaviour seem to show that children are certainly not driven by apure sensitivity to correlations of external properties in objects Summing updevelopmental research by Gelman and Wellman and others Kornblith argues thatpre-schoolers show ldquoan understanding of the relevance of underlying properties ofnatural objects to their kind and function even in cases where they do not knowwhat those underlying properties arerdquo (Kornblith 1993 p 70) This results from anepistemic stance that appears very early ldquoeven extremely young children look forregularities beyond super cial characteristics in their classi catory schemesrdquo (Korn-blith 1993 p 78) According to Kornblith developing minds are essentially on theright epistemic track given the kinds to be discovered in nature

This kind of developmental evidence however is compatible with two ratherdifferent interpretations which I will call the epistemic stance and natural metaphys-ics views respectively

The epistemic stance view is that children attend to underlying properties ofobjects as a consequence of a general principled cognitive strategy They have ageneral disposition to attend to underlying properties causally connected propertiesin particular as well as correlated surface properties of objects

The alternative natural metaphysics view is that children are led to attend tounderlying properties in speci c ways for each ontological domain In this viewthey for instance stipulate common internal functional processes in members of ananimal species not because of a general essentialistic stance but because ofparticular domain-speci c presumptions about objects in the ANIMAL domain

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 279

These presumptions in this view would not be activated when dealing withobjects in another ontological domain such as artefacts for instance

Kornblith himself does not choose between these two interpretations Bycontrast R Millikan presents a view of early concept development that stronglysuggests an ldquoepistemic stancerdquo interpretation Millikan rst argues that a singleldquosubstancerdquo notion underpins the development of concepts traditionally viewed asdifferent concepts of kinds (TIGER TELEPHONE) of ldquostuffsrdquo (MILK) and ofpersons (BILL HILLARY) We should view concepts as cognitive ldquopracticalskillsrdquo that allow us to recognize substances rather than descriptions of thesesubstances Now many properties of objects though not necessarily hidden are infact undetectable in each particular context So a natural cognitive system needs ageneral ldquoconceptual trackingrdquo capacity This conceptual analogue of perceptualtracking ldquomakes it possible for the organism to store away knowledge about thething collected on earlier encounters for use on later occasionsrdquo (Millikan 1998p 62) Conceptual tracking includes the disposition to go beyond observableencounters and to assume non-obvious stable properties in objects of asame substance This is construed as a general strategy that applies acrossdomains indeed allows the child gradually to acquire representations of the differ-ences between domains For Millikan the concept of a ldquosubstancerdquo is moregeneral than that of KIND or STUFF or INDIVIDUAL and probably comes rstin conceptual development ldquoPutting it Quinersquos way the childrsquos rst recogni-tions hellip must be merely of more Mama more milk and more mouserdquo (Millikan1998 p 58 my emphasis) This then develops into domain-speci c knowledgeldquoTracking in this property-blind way would make it possible to observe for variouskinds of objects what sorts of things tend to remain the same and what maychange within a short periodrdquo (Millikan 1998 p 63) All this seems to suggestthat early cognitions are indeed the outcome of a single cross-domain epistemicstance In the alternative ldquonatural metaphysicsrdquo view the child would be describedas developing rst a set of domain-speci c understandings of underlying proper-ties

On the whole naturalizing epistemology arguments are compatible with eitherMillikanrsquos general ldquoepistemic stancerdquo interpretation or the ldquonatural metaphysicsrdquointerpretation However choosing between these twomdashthat is evaluating which ofthem is better supported by the developmental evidencemdashdoes make a differenceas regards the initial form and the plasticity of ontological commitments

If we construe early development as driven by one general epistemic attitudewe should expect that the infantrsquos conceptual baggage consists not in speci cpresumptions about what kinds of objects there are in the world but of a set ofgood recipes about how to get sound evidence beyond misleading surface featuresThis in turn implies that ontological commitments should be rather exibledepending on what kinds of evidence children are presented with Some aspects ofconceptual development seem to support this view For instance evidencefor early ldquoessentialismrdquo shows that the childrsquos ontological understandings areintrinsically theoretical underdetermined by available evidence (Gopnik ampWellmann 1994 Wellmann amp Gelman 1992) Now if cognitive processes

280 P BOYER

are intrinsically theoretical they are also defeasible as Churchland points out(Churchland 1989 p 188 see also Churchland 1979 p 63ff)

If on the other hand we interpret conceptual development as the consequenceof natural metaphysics the infant is described as starting with a set of de nitepresumptions about broad ontological categories This in turn suggests that develop-ing ontological commitments should be less exible as they are partly informed bythese prior ontological presumptions

A more precise consideration of the evidence does not just help choose betweenthese interpretations of early development It also suggests a precise computation-ally tractable description of what ldquoprior ontological presumptionsrdquo may consist ofand of their connection to evolved computational properties of human brains

Psychological evidence categories and inferences

That ontological categories are real psychological structures is not really in doubtConcepts such as ANIMAL PERSON or ARTEFACT are activated by speci ccues in the environment and allow speci c inferences from these cues Presentationof new instances of these different categories triggers different neural activation (seeeg Martin et al 1996) and some neural pathologies sometimes result in selectiveimpairment of one of the categories (see eg Sartori amp Job 1988 Damasio 1990)Developmental evidence demonstrates the early emergence of a system of fastspontaneous identi cations categorizations and inferences at the level of broadontological categories This occurs at an age where these processes cannot be muchaffected by conscious deliberation or by explicitly available knowledge

There is evidence for the presence and salience of categories like HUMANARTEFACT ANIMAL from infancy especially from spontaneous sorting behav-iour (Younger 1985) and serial touching (Starkey 1981) That infants possess thecapacities underlying complex categorization activities is not really controversial(Nelson 1973 Younger 1990) Infants use those general capacities in the spon-taneous representation of their environment Obviously human infants can likemost animals distinguish con-speci cs from other types of beings on the basis ofcomplex perceptual cues (Morton amp Johnson 1991) that trigger particular forms ofinteraction (see Meltzoff 1994 for instance) But less obvious ontological distinc-tions appear very early like a distinction between living and non-living things(Bullock 1985 Gelman et al 1983 Richards amp Siegler 1986) that is grounded inan early sensitivity to the perceptual difference between self- and non-self-generatedmovement in physical objects (Massey amp Gelman 1988 Premack 1990 Premackamp James-Premack 1995) The salience of the category ANIMAL is also demon-strated in habituation tests (Eimas 1994 Faulkender 1974 Ross 1980) Mandlerand her colleagues have demonstrated the existence of a variety of ontologicalcategories such as ANIMAL and ARTEFACT in 18-month-old children (Mandleret al 1991)

To say that ontological categories include perceptual cues does not entail thatthey are the direct outcome of simple feature distinctions Indeed ontologicalidenti cation often overrides super cial similarity Although young children focus

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 281

on shape similarity when trying to constrain the reference of a novel noun (Landauet al 1988) the salience of shape is dependent on prior ontological identi cation[1] Ontological categories allow children to make a clear intuitive distinctionbetween super cial similarity and same kind membership

Ontological categories are intimately associated with domain-speci c inferenceengines also called ldquomodes of construalrdquo (Keil 1994) or ldquoskeletal principlesrdquo(Gelman 1990) in the literature Inference engines focus on particular aspects of theobjects encountered They are activated by the identi cation of an object as amember of one of the ontological categories They deliver inferences about variousnon-obvious aspects of these objects They can be concurrently activated in therepresentation of a single object To take a simple example seeing someone walktowards you and anticipating what will happen next simultaneously involves (1)physical predictions eg how fast this personrsquos current motion will get her closer toyou (2) biological expectations eg that this person is propelled by some internalsource of energy and (3) psychological inferences eg that this person may want totalk to you There is developmental evidence for a variety of such aspect-basedprinciples of which I will only mention a few (for a general survey see Atran 1989Gelman 1990 and the essays collected in Hirschfeld amp Gelman 1994)

Number

The domain of number provides a good illustration of speci city of informationalcues and inference principles Experiments by Wynn demonstrated that infants aresensitive to the pure ldquonumerosityrdquo of displays (Wynn 1992) ldquoImpossiblerdquo changesin the number of items displayed (eg 1 1 1 5 1 2ndash1 5 2) surprise infants indishabituation experiments (see also Antell amp Keating 1983 Starkey et al 1990)Early sensitivity to numerical aspects of displays underlies the subsequent ease withwhich children acquire counting systems (Gallistel amp Gelman 1992)

Physics of solid objects

Many of the principles underlying adult intuitive expectations in the domain ofldquointuitive physicsrdquo appear early in infancy eg ldquocontinuityrdquo (objects move incontinuous paths) ldquosolidityrdquo (objects do not coincide in space) ldquosupportrdquo (unsup-ported objects fall downwards) (Baillargeon 1987 Baillargeon amp Hanko-Summers1990 Spelke 1990) or a rule of ldquono action at a distancerdquo that excludes non-contiguous physical causation (Leslie 1979 1988)

Features of animacy

Identi cation of an object as an ANIMAL or HUMAN triggers inferences aboutagency For instance pre-schoolers assume that movement is externally generatedfor artefacts and internally generated for animals This even in a case where bothkinds of items are unfamiliar and even in a case where the child can actually see ahuman hand moving the animal (Gelman et al 1994) Children are reluctant to

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

278 P BOYER

arguments in ldquonaturalized epistemologyrdquo take their illustrations and derive some oftheir arguments from the experimental literature on conceptual development How-ever this use of the evidence is often misleading I will argue here that a moreprecise description of the developmental evidence suggests a more speci c picture ofearly ontology (and of its consequences for adult conceptual structure) This mayhelp construct a natural epistemology supported by computationally tractableempirically based accounts of our cognitive dispositions rather than by a prioriassumptions about what they should be It may also illuminate traditional debateson processes of concept acquisition on the conceptual continuity between infantsand adults and on the putative ldquoinnatenessrdquo of particular ontological commitments

Epistemic stance vs natural metaphysics

That our conceptual sensitivities are dramatically off target can be in principleexcluded on evolutionary grounds (Quine 1969) However this leaves open thepossibility of a vast range of sub-optimal designs whereby organisms survive on thebasis of false but innocuous beliefs A ldquosoundrdquo evolved cognition in this view wouldhave two main characteristics First it would be able to recognize relevant featuresof objects and identify their correlations Second it would go beyond surfacecorrelations and identify underlying structures that cause them

Developmental psychological evidence seems to bear good news for this versionof the naturalized epistemology project Studies of pre-schoolersrsquo induction andcategorization behaviour seem to show that children are certainly not driven by apure sensitivity to correlations of external properties in objects Summing updevelopmental research by Gelman and Wellman and others Kornblith argues thatpre-schoolers show ldquoan understanding of the relevance of underlying properties ofnatural objects to their kind and function even in cases where they do not knowwhat those underlying properties arerdquo (Kornblith 1993 p 70) This results from anepistemic stance that appears very early ldquoeven extremely young children look forregularities beyond super cial characteristics in their classi catory schemesrdquo (Korn-blith 1993 p 78) According to Kornblith developing minds are essentially on theright epistemic track given the kinds to be discovered in nature

This kind of developmental evidence however is compatible with two ratherdifferent interpretations which I will call the epistemic stance and natural metaphys-ics views respectively

The epistemic stance view is that children attend to underlying properties ofobjects as a consequence of a general principled cognitive strategy They have ageneral disposition to attend to underlying properties causally connected propertiesin particular as well as correlated surface properties of objects

The alternative natural metaphysics view is that children are led to attend tounderlying properties in speci c ways for each ontological domain In this viewthey for instance stipulate common internal functional processes in members of ananimal species not because of a general essentialistic stance but because ofparticular domain-speci c presumptions about objects in the ANIMAL domain

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 279

These presumptions in this view would not be activated when dealing withobjects in another ontological domain such as artefacts for instance

Kornblith himself does not choose between these two interpretations Bycontrast R Millikan presents a view of early concept development that stronglysuggests an ldquoepistemic stancerdquo interpretation Millikan rst argues that a singleldquosubstancerdquo notion underpins the development of concepts traditionally viewed asdifferent concepts of kinds (TIGER TELEPHONE) of ldquostuffsrdquo (MILK) and ofpersons (BILL HILLARY) We should view concepts as cognitive ldquopracticalskillsrdquo that allow us to recognize substances rather than descriptions of thesesubstances Now many properties of objects though not necessarily hidden are infact undetectable in each particular context So a natural cognitive system needs ageneral ldquoconceptual trackingrdquo capacity This conceptual analogue of perceptualtracking ldquomakes it possible for the organism to store away knowledge about thething collected on earlier encounters for use on later occasionsrdquo (Millikan 1998p 62) Conceptual tracking includes the disposition to go beyond observableencounters and to assume non-obvious stable properties in objects of asame substance This is construed as a general strategy that applies acrossdomains indeed allows the child gradually to acquire representations of the differ-ences between domains For Millikan the concept of a ldquosubstancerdquo is moregeneral than that of KIND or STUFF or INDIVIDUAL and probably comes rstin conceptual development ldquoPutting it Quinersquos way the childrsquos rst recogni-tions hellip must be merely of more Mama more milk and more mouserdquo (Millikan1998 p 58 my emphasis) This then develops into domain-speci c knowledgeldquoTracking in this property-blind way would make it possible to observe for variouskinds of objects what sorts of things tend to remain the same and what maychange within a short periodrdquo (Millikan 1998 p 63) All this seems to suggestthat early cognitions are indeed the outcome of a single cross-domain epistemicstance In the alternative ldquonatural metaphysicsrdquo view the child would be describedas developing rst a set of domain-speci c understandings of underlying proper-ties

On the whole naturalizing epistemology arguments are compatible with eitherMillikanrsquos general ldquoepistemic stancerdquo interpretation or the ldquonatural metaphysicsrdquointerpretation However choosing between these twomdashthat is evaluating which ofthem is better supported by the developmental evidencemdashdoes make a differenceas regards the initial form and the plasticity of ontological commitments

If we construe early development as driven by one general epistemic attitudewe should expect that the infantrsquos conceptual baggage consists not in speci cpresumptions about what kinds of objects there are in the world but of a set ofgood recipes about how to get sound evidence beyond misleading surface featuresThis in turn implies that ontological commitments should be rather exibledepending on what kinds of evidence children are presented with Some aspects ofconceptual development seem to support this view For instance evidencefor early ldquoessentialismrdquo shows that the childrsquos ontological understandings areintrinsically theoretical underdetermined by available evidence (Gopnik ampWellmann 1994 Wellmann amp Gelman 1992) Now if cognitive processes

280 P BOYER

are intrinsically theoretical they are also defeasible as Churchland points out(Churchland 1989 p 188 see also Churchland 1979 p 63ff)

If on the other hand we interpret conceptual development as the consequenceof natural metaphysics the infant is described as starting with a set of de nitepresumptions about broad ontological categories This in turn suggests that develop-ing ontological commitments should be less exible as they are partly informed bythese prior ontological presumptions

A more precise consideration of the evidence does not just help choose betweenthese interpretations of early development It also suggests a precise computation-ally tractable description of what ldquoprior ontological presumptionsrdquo may consist ofand of their connection to evolved computational properties of human brains

Psychological evidence categories and inferences

That ontological categories are real psychological structures is not really in doubtConcepts such as ANIMAL PERSON or ARTEFACT are activated by speci ccues in the environment and allow speci c inferences from these cues Presentationof new instances of these different categories triggers different neural activation (seeeg Martin et al 1996) and some neural pathologies sometimes result in selectiveimpairment of one of the categories (see eg Sartori amp Job 1988 Damasio 1990)Developmental evidence demonstrates the early emergence of a system of fastspontaneous identi cations categorizations and inferences at the level of broadontological categories This occurs at an age where these processes cannot be muchaffected by conscious deliberation or by explicitly available knowledge

There is evidence for the presence and salience of categories like HUMANARTEFACT ANIMAL from infancy especially from spontaneous sorting behav-iour (Younger 1985) and serial touching (Starkey 1981) That infants possess thecapacities underlying complex categorization activities is not really controversial(Nelson 1973 Younger 1990) Infants use those general capacities in the spon-taneous representation of their environment Obviously human infants can likemost animals distinguish con-speci cs from other types of beings on the basis ofcomplex perceptual cues (Morton amp Johnson 1991) that trigger particular forms ofinteraction (see Meltzoff 1994 for instance) But less obvious ontological distinc-tions appear very early like a distinction between living and non-living things(Bullock 1985 Gelman et al 1983 Richards amp Siegler 1986) that is grounded inan early sensitivity to the perceptual difference between self- and non-self-generatedmovement in physical objects (Massey amp Gelman 1988 Premack 1990 Premackamp James-Premack 1995) The salience of the category ANIMAL is also demon-strated in habituation tests (Eimas 1994 Faulkender 1974 Ross 1980) Mandlerand her colleagues have demonstrated the existence of a variety of ontologicalcategories such as ANIMAL and ARTEFACT in 18-month-old children (Mandleret al 1991)

To say that ontological categories include perceptual cues does not entail thatthey are the direct outcome of simple feature distinctions Indeed ontologicalidenti cation often overrides super cial similarity Although young children focus

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 281

on shape similarity when trying to constrain the reference of a novel noun (Landauet al 1988) the salience of shape is dependent on prior ontological identi cation[1] Ontological categories allow children to make a clear intuitive distinctionbetween super cial similarity and same kind membership

Ontological categories are intimately associated with domain-speci c inferenceengines also called ldquomodes of construalrdquo (Keil 1994) or ldquoskeletal principlesrdquo(Gelman 1990) in the literature Inference engines focus on particular aspects of theobjects encountered They are activated by the identi cation of an object as amember of one of the ontological categories They deliver inferences about variousnon-obvious aspects of these objects They can be concurrently activated in therepresentation of a single object To take a simple example seeing someone walktowards you and anticipating what will happen next simultaneously involves (1)physical predictions eg how fast this personrsquos current motion will get her closer toyou (2) biological expectations eg that this person is propelled by some internalsource of energy and (3) psychological inferences eg that this person may want totalk to you There is developmental evidence for a variety of such aspect-basedprinciples of which I will only mention a few (for a general survey see Atran 1989Gelman 1990 and the essays collected in Hirschfeld amp Gelman 1994)

Number

The domain of number provides a good illustration of speci city of informationalcues and inference principles Experiments by Wynn demonstrated that infants aresensitive to the pure ldquonumerosityrdquo of displays (Wynn 1992) ldquoImpossiblerdquo changesin the number of items displayed (eg 1 1 1 5 1 2ndash1 5 2) surprise infants indishabituation experiments (see also Antell amp Keating 1983 Starkey et al 1990)Early sensitivity to numerical aspects of displays underlies the subsequent ease withwhich children acquire counting systems (Gallistel amp Gelman 1992)

Physics of solid objects

Many of the principles underlying adult intuitive expectations in the domain ofldquointuitive physicsrdquo appear early in infancy eg ldquocontinuityrdquo (objects move incontinuous paths) ldquosolidityrdquo (objects do not coincide in space) ldquosupportrdquo (unsup-ported objects fall downwards) (Baillargeon 1987 Baillargeon amp Hanko-Summers1990 Spelke 1990) or a rule of ldquono action at a distancerdquo that excludes non-contiguous physical causation (Leslie 1979 1988)

Features of animacy

Identi cation of an object as an ANIMAL or HUMAN triggers inferences aboutagency For instance pre-schoolers assume that movement is externally generatedfor artefacts and internally generated for animals This even in a case where bothkinds of items are unfamiliar and even in a case where the child can actually see ahuman hand moving the animal (Gelman et al 1994) Children are reluctant to

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 279

These presumptions in this view would not be activated when dealing withobjects in another ontological domain such as artefacts for instance

Kornblith himself does not choose between these two interpretations Bycontrast R Millikan presents a view of early concept development that stronglysuggests an ldquoepistemic stancerdquo interpretation Millikan rst argues that a singleldquosubstancerdquo notion underpins the development of concepts traditionally viewed asdifferent concepts of kinds (TIGER TELEPHONE) of ldquostuffsrdquo (MILK) and ofpersons (BILL HILLARY) We should view concepts as cognitive ldquopracticalskillsrdquo that allow us to recognize substances rather than descriptions of thesesubstances Now many properties of objects though not necessarily hidden are infact undetectable in each particular context So a natural cognitive system needs ageneral ldquoconceptual trackingrdquo capacity This conceptual analogue of perceptualtracking ldquomakes it possible for the organism to store away knowledge about thething collected on earlier encounters for use on later occasionsrdquo (Millikan 1998p 62) Conceptual tracking includes the disposition to go beyond observableencounters and to assume non-obvious stable properties in objects of asame substance This is construed as a general strategy that applies acrossdomains indeed allows the child gradually to acquire representations of the differ-ences between domains For Millikan the concept of a ldquosubstancerdquo is moregeneral than that of KIND or STUFF or INDIVIDUAL and probably comes rstin conceptual development ldquoPutting it Quinersquos way the childrsquos rst recogni-tions hellip must be merely of more Mama more milk and more mouserdquo (Millikan1998 p 58 my emphasis) This then develops into domain-speci c knowledgeldquoTracking in this property-blind way would make it possible to observe for variouskinds of objects what sorts of things tend to remain the same and what maychange within a short periodrdquo (Millikan 1998 p 63) All this seems to suggestthat early cognitions are indeed the outcome of a single cross-domain epistemicstance In the alternative ldquonatural metaphysicsrdquo view the child would be describedas developing rst a set of domain-speci c understandings of underlying proper-ties

On the whole naturalizing epistemology arguments are compatible with eitherMillikanrsquos general ldquoepistemic stancerdquo interpretation or the ldquonatural metaphysicsrdquointerpretation However choosing between these twomdashthat is evaluating which ofthem is better supported by the developmental evidencemdashdoes make a differenceas regards the initial form and the plasticity of ontological commitments

If we construe early development as driven by one general epistemic attitudewe should expect that the infantrsquos conceptual baggage consists not in speci cpresumptions about what kinds of objects there are in the world but of a set ofgood recipes about how to get sound evidence beyond misleading surface featuresThis in turn implies that ontological commitments should be rather exibledepending on what kinds of evidence children are presented with Some aspects ofconceptual development seem to support this view For instance evidencefor early ldquoessentialismrdquo shows that the childrsquos ontological understandings areintrinsically theoretical underdetermined by available evidence (Gopnik ampWellmann 1994 Wellmann amp Gelman 1992) Now if cognitive processes

280 P BOYER

are intrinsically theoretical they are also defeasible as Churchland points out(Churchland 1989 p 188 see also Churchland 1979 p 63ff)

If on the other hand we interpret conceptual development as the consequenceof natural metaphysics the infant is described as starting with a set of de nitepresumptions about broad ontological categories This in turn suggests that develop-ing ontological commitments should be less exible as they are partly informed bythese prior ontological presumptions

A more precise consideration of the evidence does not just help choose betweenthese interpretations of early development It also suggests a precise computation-ally tractable description of what ldquoprior ontological presumptionsrdquo may consist ofand of their connection to evolved computational properties of human brains

Psychological evidence categories and inferences

That ontological categories are real psychological structures is not really in doubtConcepts such as ANIMAL PERSON or ARTEFACT are activated by speci ccues in the environment and allow speci c inferences from these cues Presentationof new instances of these different categories triggers different neural activation (seeeg Martin et al 1996) and some neural pathologies sometimes result in selectiveimpairment of one of the categories (see eg Sartori amp Job 1988 Damasio 1990)Developmental evidence demonstrates the early emergence of a system of fastspontaneous identi cations categorizations and inferences at the level of broadontological categories This occurs at an age where these processes cannot be muchaffected by conscious deliberation or by explicitly available knowledge

There is evidence for the presence and salience of categories like HUMANARTEFACT ANIMAL from infancy especially from spontaneous sorting behav-iour (Younger 1985) and serial touching (Starkey 1981) That infants possess thecapacities underlying complex categorization activities is not really controversial(Nelson 1973 Younger 1990) Infants use those general capacities in the spon-taneous representation of their environment Obviously human infants can likemost animals distinguish con-speci cs from other types of beings on the basis ofcomplex perceptual cues (Morton amp Johnson 1991) that trigger particular forms ofinteraction (see Meltzoff 1994 for instance) But less obvious ontological distinc-tions appear very early like a distinction between living and non-living things(Bullock 1985 Gelman et al 1983 Richards amp Siegler 1986) that is grounded inan early sensitivity to the perceptual difference between self- and non-self-generatedmovement in physical objects (Massey amp Gelman 1988 Premack 1990 Premackamp James-Premack 1995) The salience of the category ANIMAL is also demon-strated in habituation tests (Eimas 1994 Faulkender 1974 Ross 1980) Mandlerand her colleagues have demonstrated the existence of a variety of ontologicalcategories such as ANIMAL and ARTEFACT in 18-month-old children (Mandleret al 1991)

To say that ontological categories include perceptual cues does not entail thatthey are the direct outcome of simple feature distinctions Indeed ontologicalidenti cation often overrides super cial similarity Although young children focus

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 281

on shape similarity when trying to constrain the reference of a novel noun (Landauet al 1988) the salience of shape is dependent on prior ontological identi cation[1] Ontological categories allow children to make a clear intuitive distinctionbetween super cial similarity and same kind membership

Ontological categories are intimately associated with domain-speci c inferenceengines also called ldquomodes of construalrdquo (Keil 1994) or ldquoskeletal principlesrdquo(Gelman 1990) in the literature Inference engines focus on particular aspects of theobjects encountered They are activated by the identi cation of an object as amember of one of the ontological categories They deliver inferences about variousnon-obvious aspects of these objects They can be concurrently activated in therepresentation of a single object To take a simple example seeing someone walktowards you and anticipating what will happen next simultaneously involves (1)physical predictions eg how fast this personrsquos current motion will get her closer toyou (2) biological expectations eg that this person is propelled by some internalsource of energy and (3) psychological inferences eg that this person may want totalk to you There is developmental evidence for a variety of such aspect-basedprinciples of which I will only mention a few (for a general survey see Atran 1989Gelman 1990 and the essays collected in Hirschfeld amp Gelman 1994)

Number

The domain of number provides a good illustration of speci city of informationalcues and inference principles Experiments by Wynn demonstrated that infants aresensitive to the pure ldquonumerosityrdquo of displays (Wynn 1992) ldquoImpossiblerdquo changesin the number of items displayed (eg 1 1 1 5 1 2ndash1 5 2) surprise infants indishabituation experiments (see also Antell amp Keating 1983 Starkey et al 1990)Early sensitivity to numerical aspects of displays underlies the subsequent ease withwhich children acquire counting systems (Gallistel amp Gelman 1992)

Physics of solid objects

Many of the principles underlying adult intuitive expectations in the domain ofldquointuitive physicsrdquo appear early in infancy eg ldquocontinuityrdquo (objects move incontinuous paths) ldquosolidityrdquo (objects do not coincide in space) ldquosupportrdquo (unsup-ported objects fall downwards) (Baillargeon 1987 Baillargeon amp Hanko-Summers1990 Spelke 1990) or a rule of ldquono action at a distancerdquo that excludes non-contiguous physical causation (Leslie 1979 1988)

Features of animacy

Identi cation of an object as an ANIMAL or HUMAN triggers inferences aboutagency For instance pre-schoolers assume that movement is externally generatedfor artefacts and internally generated for animals This even in a case where bothkinds of items are unfamiliar and even in a case where the child can actually see ahuman hand moving the animal (Gelman et al 1994) Children are reluctant to

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

280 P BOYER

are intrinsically theoretical they are also defeasible as Churchland points out(Churchland 1989 p 188 see also Churchland 1979 p 63ff)

If on the other hand we interpret conceptual development as the consequenceof natural metaphysics the infant is described as starting with a set of de nitepresumptions about broad ontological categories This in turn suggests that develop-ing ontological commitments should be less exible as they are partly informed bythese prior ontological presumptions

A more precise consideration of the evidence does not just help choose betweenthese interpretations of early development It also suggests a precise computation-ally tractable description of what ldquoprior ontological presumptionsrdquo may consist ofand of their connection to evolved computational properties of human brains

Psychological evidence categories and inferences

That ontological categories are real psychological structures is not really in doubtConcepts such as ANIMAL PERSON or ARTEFACT are activated by speci ccues in the environment and allow speci c inferences from these cues Presentationof new instances of these different categories triggers different neural activation (seeeg Martin et al 1996) and some neural pathologies sometimes result in selectiveimpairment of one of the categories (see eg Sartori amp Job 1988 Damasio 1990)Developmental evidence demonstrates the early emergence of a system of fastspontaneous identi cations categorizations and inferences at the level of broadontological categories This occurs at an age where these processes cannot be muchaffected by conscious deliberation or by explicitly available knowledge

There is evidence for the presence and salience of categories like HUMANARTEFACT ANIMAL from infancy especially from spontaneous sorting behav-iour (Younger 1985) and serial touching (Starkey 1981) That infants possess thecapacities underlying complex categorization activities is not really controversial(Nelson 1973 Younger 1990) Infants use those general capacities in the spon-taneous representation of their environment Obviously human infants can likemost animals distinguish con-speci cs from other types of beings on the basis ofcomplex perceptual cues (Morton amp Johnson 1991) that trigger particular forms ofinteraction (see Meltzoff 1994 for instance) But less obvious ontological distinc-tions appear very early like a distinction between living and non-living things(Bullock 1985 Gelman et al 1983 Richards amp Siegler 1986) that is grounded inan early sensitivity to the perceptual difference between self- and non-self-generatedmovement in physical objects (Massey amp Gelman 1988 Premack 1990 Premackamp James-Premack 1995) The salience of the category ANIMAL is also demon-strated in habituation tests (Eimas 1994 Faulkender 1974 Ross 1980) Mandlerand her colleagues have demonstrated the existence of a variety of ontologicalcategories such as ANIMAL and ARTEFACT in 18-month-old children (Mandleret al 1991)

To say that ontological categories include perceptual cues does not entail thatthey are the direct outcome of simple feature distinctions Indeed ontologicalidenti cation often overrides super cial similarity Although young children focus

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 281

on shape similarity when trying to constrain the reference of a novel noun (Landauet al 1988) the salience of shape is dependent on prior ontological identi cation[1] Ontological categories allow children to make a clear intuitive distinctionbetween super cial similarity and same kind membership

Ontological categories are intimately associated with domain-speci c inferenceengines also called ldquomodes of construalrdquo (Keil 1994) or ldquoskeletal principlesrdquo(Gelman 1990) in the literature Inference engines focus on particular aspects of theobjects encountered They are activated by the identi cation of an object as amember of one of the ontological categories They deliver inferences about variousnon-obvious aspects of these objects They can be concurrently activated in therepresentation of a single object To take a simple example seeing someone walktowards you and anticipating what will happen next simultaneously involves (1)physical predictions eg how fast this personrsquos current motion will get her closer toyou (2) biological expectations eg that this person is propelled by some internalsource of energy and (3) psychological inferences eg that this person may want totalk to you There is developmental evidence for a variety of such aspect-basedprinciples of which I will only mention a few (for a general survey see Atran 1989Gelman 1990 and the essays collected in Hirschfeld amp Gelman 1994)

Number

The domain of number provides a good illustration of speci city of informationalcues and inference principles Experiments by Wynn demonstrated that infants aresensitive to the pure ldquonumerosityrdquo of displays (Wynn 1992) ldquoImpossiblerdquo changesin the number of items displayed (eg 1 1 1 5 1 2ndash1 5 2) surprise infants indishabituation experiments (see also Antell amp Keating 1983 Starkey et al 1990)Early sensitivity to numerical aspects of displays underlies the subsequent ease withwhich children acquire counting systems (Gallistel amp Gelman 1992)

Physics of solid objects

Many of the principles underlying adult intuitive expectations in the domain ofldquointuitive physicsrdquo appear early in infancy eg ldquocontinuityrdquo (objects move incontinuous paths) ldquosolidityrdquo (objects do not coincide in space) ldquosupportrdquo (unsup-ported objects fall downwards) (Baillargeon 1987 Baillargeon amp Hanko-Summers1990 Spelke 1990) or a rule of ldquono action at a distancerdquo that excludes non-contiguous physical causation (Leslie 1979 1988)

Features of animacy

Identi cation of an object as an ANIMAL or HUMAN triggers inferences aboutagency For instance pre-schoolers assume that movement is externally generatedfor artefacts and internally generated for animals This even in a case where bothkinds of items are unfamiliar and even in a case where the child can actually see ahuman hand moving the animal (Gelman et al 1994) Children are reluctant to

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 281

on shape similarity when trying to constrain the reference of a novel noun (Landauet al 1988) the salience of shape is dependent on prior ontological identi cation[1] Ontological categories allow children to make a clear intuitive distinctionbetween super cial similarity and same kind membership

Ontological categories are intimately associated with domain-speci c inferenceengines also called ldquomodes of construalrdquo (Keil 1994) or ldquoskeletal principlesrdquo(Gelman 1990) in the literature Inference engines focus on particular aspects of theobjects encountered They are activated by the identi cation of an object as amember of one of the ontological categories They deliver inferences about variousnon-obvious aspects of these objects They can be concurrently activated in therepresentation of a single object To take a simple example seeing someone walktowards you and anticipating what will happen next simultaneously involves (1)physical predictions eg how fast this personrsquos current motion will get her closer toyou (2) biological expectations eg that this person is propelled by some internalsource of energy and (3) psychological inferences eg that this person may want totalk to you There is developmental evidence for a variety of such aspect-basedprinciples of which I will only mention a few (for a general survey see Atran 1989Gelman 1990 and the essays collected in Hirschfeld amp Gelman 1994)

Number

The domain of number provides a good illustration of speci city of informationalcues and inference principles Experiments by Wynn demonstrated that infants aresensitive to the pure ldquonumerosityrdquo of displays (Wynn 1992) ldquoImpossiblerdquo changesin the number of items displayed (eg 1 1 1 5 1 2ndash1 5 2) surprise infants indishabituation experiments (see also Antell amp Keating 1983 Starkey et al 1990)Early sensitivity to numerical aspects of displays underlies the subsequent ease withwhich children acquire counting systems (Gallistel amp Gelman 1992)

Physics of solid objects

Many of the principles underlying adult intuitive expectations in the domain ofldquointuitive physicsrdquo appear early in infancy eg ldquocontinuityrdquo (objects move incontinuous paths) ldquosolidityrdquo (objects do not coincide in space) ldquosupportrdquo (unsup-ported objects fall downwards) (Baillargeon 1987 Baillargeon amp Hanko-Summers1990 Spelke 1990) or a rule of ldquono action at a distancerdquo that excludes non-contiguous physical causation (Leslie 1979 1988)

Features of animacy

Identi cation of an object as an ANIMAL or HUMAN triggers inferences aboutagency For instance pre-schoolers assume that movement is externally generatedfor artefacts and internally generated for animals This even in a case where bothkinds of items are unfamiliar and even in a case where the child can actually see ahuman hand moving the animal (Gelman et al 1994) Children are reluctant to

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

282 P BOYER

judge artefacts capable of self-propelled motion and resist extending animate proper-ties to inanimate objects such as dolls and stuffed animals although they judge themvery similar to real people and animals (Gelman et al 1983)

Structure from function inferences

These are based on the ldquoteleologicalrdquo principle that the typical function of an object(or part of an object) is an explanation for its structure Young children seem toadopt this teleological stance both for artefacts and for some aspects of the structureof living organisms (Keil 1994) However in both domains they are very selectiveand often wrong in their choice of features For instance they may identify artefactsin terms of function yet fail to identify crucial functional affordances (Tversky ampHemenway 1984)

Biological processes

Speci c principles include the assumption that members of a living kind have someinternal similarity beyond their common external features and sometimes in con ictwith observable features (Becker amp Ward 1991 Gelman amp Markman 1986 Keil1986 Massey amp Gelman 1988) Also living things are construed as belonging toa set of taxonomic classes (jointly exhaustive of the domain and mutually exclusive)such that taxonomic proximity predicts non-obvious resemblance These principleshowever constitute intuitive expectations not an integrated and explicit ldquonaivebiologyrdquo [2]

Mentalistic psychology

From the earliest stages of cognitive development children readily interpret thebehaviour of animate beings particularly persons as caused by unobservable beliefsand intentions Children from the age of three assume that thoughts and intentionsare immaterial They also assume that actual states of affairs cause perceptions thatcause beliefs that cause intentions From the age of four they can appreciate thepossibility of false beliefs (see eg Gopnik 1993 Perner 1991 Wellmann 1990)

Intuitive ontology is not a static system It only emerges gradually in the courseof conceptual development I chose to stop at what is typical of a ve- or six-year-oldbecause the categories and principles will to a large extent be found in a similar format the end-point of cognitive development On the whole intuitive ontologicalconcepts as demonstrated in studies of adult categories of neural correlates and ofneural pathologies show that further developments re ne this ontological system andelaborate upon it but generally do so in accordance with the fundamental principlesdescribed here

This very short survey highlights the functional complexity of early-developedontology Intuitive ontological categorization requires a set of speci c perceptual cuesThese lead to identi cation of an object both in terms of a kind concept and in termsof a set of ontological categories The latter activate inference engines that specialize inparticular aspects of the object and (1) stipulate non-obvious properties (2)

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 283

stipulate causal connections and (3) produce speci c expectations Obviouslycategories and aspect-speci c theoretical inference engines are related Some infer-ence engines to do with biological properties for instance are preferentiallyactivated about an object when that object is identi ed as a member of the ANIMALor HUMAN categories if something is identi ed as an ARTEFACT neitherpsychological nor biological inferences are activated in normal circumstances How-ever categorical identi cation and aspect-speci c inferencing are functionally separ-ate operations Some inference engines apply in the same way to different ontologicaldomains eg intuitive physical constraints to animals and plants and humans Alsosome situations require that one switch from one inference engine to anotherwithout changing the ontological identity of the object considered For instance thesame animal species can be alternatively considered in terms of a species-speci cessence (in situations like predation) and in terms of individual psychological features(eg pets hunting dogs)

Intuitive ontology constrains the development of kind concepts Conceptualdevelopment is more than the gradual acquisition or enrichment of foundationaltheories for ontological domains During the same developmental period childrenacquire a vast repertoire of ldquobasic-levelrdquo concepts that underpin the pre-schoolerrsquoslexical explosion It is plausible that developing minds acquire a number of particularconcepts by xing the values of variables postulated by intuitive ontology Withinontological categories the operation of inference engines introduces differentiationby specifying the range of features to be expected or stored or compared therebyproviding ldquorespects for similarityrdquo ie constraints on similarity metrics Take forinstance concepts of animal kinds Within the perceptually based identi cation ofANIMAL inference engines specify that each exemplar belongs to one exclusivetaxonomic class of ANIMAL that this results in particular ldquointernalrdquo properties thatthese are normally linked to a particular visual prototype that inter-exemplardifferences are irrelevant to behavioural dispositions and so on (Atran 1996) Thisis suf cient to generate all the seemingly ldquotheoreticalrdquo inferences that count aspossessing eg the concept ldquotigerrdquo The point applies to artefact concepts as wellIn both domains the operation of ontological categories and inference enginesprovides a set of variables whose particular values identify a basic category Obvi-ously variable- xing is not all there is to acquiring new conceptual structures A greatdeal of associative information is stored in an explicit format This requires ameta-representational capacity to represent new conceptual slots for complex combi-nations of concepts A vast amount of knowledge is represented in this way (Sperber1994) This makes it possible not only to represent explicitly the output of intuitiveontology but also to create concepts that go beyond intuitive expectations or violatethem (Boyer 1998a) Formal and informal instruction result in the constantenrichment of the conceptual base predominantly though not exclusively throughmeta-representational mechanism

General epistemic stance or domain-speci c presumptions

Early-developed presumptions about ldquounderlying structurerdquo of kinds might result

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

284 P BOYER

from a general epistemic stance or from particular presumptions about each of theontological domains In the general stance interpretation for instanceldquopsychological essentialismrdquo is characteristic of human conceptual developmentacross domains

The psychological evidence suggests multiple domain-speci c ldquoessentialist in-ferencesrdquo much more restricted in scope than a general epistemic stance Essential-ism in the child and the adult is not one general assumption but a variety ofinferential mechanisms not necessarily applied together or to all domains Considerfor instance the difference between PLANTS and ANIMALS In both domains weintuitively expect singular instances to be members of classes that are jointlyexhaustive and mutually exclusive (Atran amp Medin 1999) However plants andanimals are not construed in the same way at all For instance children intuitivelyassume (1) that members of an animal species have the same ldquoinnardsrdquo (2) thatinner structure is more stable than observable features and (3) that inner featurescause external ones Now these aspects of essentialist assumptions are less readilyapplied to plants than to animals Even in domains where children do applystraightforward full-blown essentialist assumptions they do it with remarkableselectivity To take but one example children are more willing to use ldquoinnatepotentialrdquo as an explanation for dynamic behaviours (eg that an animal makes aparticular noise) than for static features (eg that an animal has a straight tail)There are many other examples of this selectivity and of the childrsquos generaltendency to assume that only some features count as essentially caused regardless oftheir frequency in actual instances Note also that children may use essentialistinferences in domains where they do not actually focus on ldquokindsrdquo but rather onsocially de ned ldquopseudo-kindsrdquo like racial and ethnic categories (Hirschfeld 1996)as well as non-kind collections such as kinship-based social groups (Hirschfeld1989) All this leads Hirschfeld and Gelman to conclude that essentialism is neitheran exclusively biological phenomenon nor a pervasive cognitive bias (Hirschfeld ampGelman 1999) It is an inference mechanism that is selectively applied to particularaspects of the ANIMAL PERSON and even SUBSTANCE domains and in eachdomain only to particular features of the objects considered Children refrain fromessentializing certain aspects of natural kinds even when there is supportingevidence

This also seems to cast doubt on the notion of a general ldquoconceptual trackingrdquocapacity Millikan argued that this general capacity comes rst and domain-speci cpresumptions second The ldquopractical ability to re-identify a substance hellip has to becomplemented with another equally important ability hellip One hellip must have some graspof what kinds of things can be learnedrdquo about the substance (Millikan 1998 p 58my emphasis) There is however no positive evidence for this two-stage processInasmuch as we have evidence for ldquoconceptual trackingrdquo that is evidence forparticular kinds of tracking applied to particular aspects of particular categories ofobjects Moreover there is no evidence that this selectivity appears as a result oftabulating which features are more stable than others Face-recognition capacitiesare in place from an early stage before the child can perceive that faces change lessthan clothes (Morton amp Johnson 1991) The link between the facial expressions the

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 285

child detects and the ones she produces appears before the child can detect the effectof her own emotional displays on others (Meltzoff 1994 Meltzoff amp Moore 1983)The infant detects some phonological features in her language before she can tabulatewhich features are actually more critical than others (Mehler et al 1988) Ingeneral we have no evidence for general cross-domain same-substance detectioncapacities before the emergence of domain-speci c similarity-with-respect-to-X ca-pacities That ldquoconceptual trackingrdquo is performed in parallel by several independentcapacities is supported by the available evidence rather than by the notion thatconceptual development ldquomustrdquo proceed in a particular way

To sum up then both Millikanrsquos ldquosame-substance trackingrdquo and Kornblithrsquosldquoessentialist stancerdquo are on the right track in the sense that the childrsquos performanceis compatible with ldquoessentialismrdquo and ldquoconceptual trackingrdquo However that thechildrsquos cognitions are compatible with a given epistemological stance does not entailthat they are caused by a cognitively implemented version of that stance Theevidence suggests a more complex system of intuitive domain-speci c expectationsthat is a natural metaphysics rather than a natural epistemology

Does development suggest ontological plasticity

Churchland has long maintained that conceptual structures display inherentldquoplasticityrdquo This notion has become the focus of an intense debate On one sideJerry Fodor has maintained that activation and functioning of modular structures isimpervious to higher-level in uence from knowledge or expertise (Fodor 1988)Churchland takes the opposite stance and observes that major revisions occur inknowledge but also in perception (Churchland 1988) [3] If Churchland is rightabout cognition in general we should expect ontological concepts and inferences tobe ldquoplasticrdquo too as a result of changing cultural circumstances In this section I listdifferent types of arguments or evidence concerning the potential plasticity ofhuman cognition in general and examine to what extent they apply to intuitiveontology in particular

Does ontology change in childhood

Developmental psychologists have been arguing for some time between ldquocontinuityrdquo(Eimas 1994 Spelke 1988) and ldquoconceptual changerdquo (Carey 1985) interpretationsof development The continuity thesis states that development boils down to agradual enrichment of fundamental intuitive principles By contrast Carey andother authors have demonstrated important changes in such fundamental propertiesas WEIGHT in intuitive physics and LIVE BEING in biology and consider theseas every bit as radical as scienti c revisions of ontology Here we must distinguishbetween two kinds of change One is the emergence of new principles that cannotbe accounted for in terms of previous principles For example children around theage of four begin to show some awareness of the behavioural consequences of falsebelief which was not previously available A second type of change is the acquisitionof generic knowledge that can override some intuitive principles or refocus their

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

286 P BOYER

application Children gradually learn that some aspects of animal behaviour areproperly understood in biological rather than psychological terms The principlesof biological understandings are not changed in this process but their domain ofapplication certainly is Asking whether there is ldquogenuine conceptual changerdquo incognitive development in general blurs this important distinction There is nodisagreement in developmental psychology on the fact that categories and principlesof intuitive ontology do mature However once these categories and principles arein place further knowledge does not seem to revise or displace them

Does training or instruction change intuitive ontology

As Churchland points out it often seems possible to retrain onersquos intuition oracquire what would appear to be new intuitive responses as a result of high-levelknowledge Obviously the most salient example of conceptual training is the system-atic acquisition of scienti c knowledge Science challenges the concepts that couldbe built by straightforward extensions of intuitive expectations Science has no placefor intuitive notions of FORCE in physics or ESSENCE in biology Scienti cpsychology challenges intuitive understandings of mental functioning and may oneday demonstrate that concepts of ldquobeliefrdquo and ldquointentionrdquo are incoherent (Church-land 1981) So this is where we should see major revisions of intuitive ontology

As far as evidence can tell acquiring scienti c theories that con ict withintuitive understandings does not usually result in a straightforward replacement ofthose understandings Knowledge of a theory does not create the kind of intuitiveexpectations that would be consistent with that theory Indeed this lack of congru-ence was the initial trigger for studies of ldquointuitiverdquo or ldquofolk-theoriesrdquo of suchdomains as physics statistics logic or biology Students trained in logic still tend togive ldquowrongrdquo (but intuitively salient) answers to simple tests and show robust biasesin their evaluation of the con rmation of hypotheses (Wason 1983 Wason ampJohnson-Laird 1972) Intuitive understandings also provide us with frequency-based assumptions about the probability of events evaluating the probability of asingle event is a particularly dif cult task for humans (Gigerenzer 1991 Gigerenzeramp Hoffrage 1995 Koehler 1996) which may explain why even ldquoexpertsrdquo some-times fail to use Bayesian principles (Tversky amp Kahnemann 1982 see a discussionin Gigerenzer amp Murray 1987 pp 163ndash167) In the domain of physics subjectsrsquoexpectations about ballistics for instance are not really congruent with Newtonianmechanics and tend to relate force to speed rather than to acceleration (Kaiser et al1986 McCloskey 1983) [4]

Intuitive understandings and scienti c ones do not seem to con ict simplybecause they do not really meet This is partly caused by the fact that scienti cconcepts and intuitive ontology develop through very different routes In particularscienti c concepts and propositions are invariably acquired in the form of meta-representational beliefs Darwinians for instance are not people who just happen tobelieve that ldquoevolution is caused by blind variation and selective retentionrdquo butpeople who believe that they have that belief and that it is true Otherwise it wouldnot be possible for them to argue for that view of evolution evaluate it in the light

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 287

of evidence assess its explanatory power and so on All these operations requirethat the belief in variation and retention be explicitly represented in other wordsthat the representations in question are meta-representational

Are there historicalndashcultural differences

That different cultures or periods have a different take on broad ontologicalcategories and inference engines would constitute straightforward evidence for exibility At rst sight the anthropological record seems to provide evidencefor such differences ldquoOtherrdquo people are described as believing that rocks y in theair mountains have blood circulation trees have a soul spirits y right throughphysical obstacles and so on (Here ldquoother peoplerdquo in fact refers to anyone outsidea tiny population of scienti c rationalists) While such representations are oftentaken very seriously by the people concerned there is no evidence that intuitiveontology differs from one location to another Developmental and adult evidencewhen available shows categories and developmental schedules similar to those ofWestern subjects (Walker 1985) Moreover it is probably the case that suchsupernatural representations are parasitic upon intuitive ontology which (1) pro-vides a background against which violations of intuitive expectations are attention-grabbing and (2) supplements those violations with inferences derived fromnon-violated principles For instance rituals explicitly directed at superhumanagents are founded on tacit intuitive understandings of agency (Lawson amp Mc-Cauley 1990) There is now experimental evidence for these processes (Barrett ampKeil 1996)

Turning to less fancy beliefs there is no evidence for major cultural differencesin the categorical distinctions or inferential engines described above Considerintuitive biology although we do nd cultural differences in the way biologicalphenomena are explained the major features of biological essentialism as describedhere are stable across cultures (Atran 1996) In the domain of intuitive psycholog-ical inferences (ldquotheory of mindrdquo) we nd a similar pattern There are culturalvariations in explicit constructs like a local theory of personality of the connectionsbetween mind and body of how minds develop (Heelas amp Lock 1981) Theseexplicit cultural theories consist in an enrichment never in a revision of the intuitiveprinciples described above Developmentally too one nds unexpected culturalsimilarity for instance in the emergence of ldquotheory of mindrdquo in Western and Pygmychildren (Avis amp Harris 1991) Differences of developmental schedule remainwithin a narrow envelope and lead to the same end-point Despite otherwise vastlydifferent cultural environments people seem to use categories and inference enginesthat are substantially similar

This also applies to science as a cultural phenomenon Churchland (1979) hasprovided a vivid description of a possible future culture based on scienti c under-standings rather than sloppy and ultimately false common-sense concepts It is ofcourse dif cult to predict the course of cultural evolution but the past recordinasmuch as it provides a sound basis for induction suggests a rather differentpicture The diffusion of scienti c knowledge certainly has effects on generic

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

288 P BOYER

knowledge in a population But as I said above thatrsquos not the same as changingintuitive understandings Despite more than three centuries of Newtonian physicsthe theory fails to govern intuitive expectations Even people who actually use thetheory seem to produce routine spontaneous physical inferences on the basis of thesame intuitive principles as untutored subjects the world over

To sum up then curing human history and across cultures people have beenand are faced with very diverse environments (social and natural) as well as differentculturally transmitted explicit accounts of what makes solid things solid what makesliving things live what constitutes a mind and so on These differences providesubstantial evidence for the ldquoplasticityrdquo of mind understood as the capacity toentertain a wide variety of concepts and beliefs However it is also striking that we nd no corresponding differences in intuitive ontology Cultural input seems tobuild upon intuitive ontology to challenge it sometimes even to build a morepowerful ontology in the case of science But it does not eliminate the intuitiveexpectations

ldquoInnate conceptsrdquo symbolic and computational descriptions

A crucial problem for developmental psychology is to account for the early emerg-ence of non-trivial ontological distinctions For instance infants make distinctionsbetween linguistic intonation and other sounds between faces and other displaysbetween inanimate and animate motion between physical and intentional causationThis is often interpreted as evidence for a rich conceptual repertoire at the initialstage of development (Spelke 1988) A traditional argument here stems from thepoverty of the stimulus There is not enough information in the auditory environ-ment to tell the infant what is language and what is noise Similar arguments areinvoked to explain early intuitive physics face-recognition or expectations of socialinteraction However it is dif cult to evaluate the plausibility of such arguments inthe absence of any account however speculative of their implementation AsChurchland points out the best naturalistic account of what it is for a cognitivesystem to have a theory is not a classical sentential model but the ldquostatendashspacesemanticsrdquo of neural networks (Churchland 1995) Now networks are not justpassive information-storage devices They go beyond the information given in thatthe effects of input vectors depend on past activation now stored in the implicit formof synaptic weight patterns So input poverty arguments may grossly underestimatethe amounts of information complex networks can extract from a set of noisystimuli That such networks are capable of producing quasi-theoretical inferences iscomputationally coherent (Rumelhart amp Todd 1992) empirically plausible as amodel of development (Elman et al 1996 Quinn amp Johnson 1997 Thelenamp Smith 1995) and consistent with neuro-physiological development (Quartz ampSejnowski 1997) Network and dynamic system models would suggest that it is byno means necessary that infants are equipped with rich conceptual structures

Resolution of this ldquodebaterdquo is hindered by two conceptual ambiguitiesFirst discussions of whether certain ontological concepts are or not present at

the initial stage are by and large informed by a ldquodescriptionistrdquo account of concepts

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 289

But understanding developmental change may require that we take a leaf out ofMillikanrsquos book and construe most concepts including ontological categories as(specialized) sameness tracking skills rather than descriptive structures (Millikan1998)

Second contending accounts of development differ along not just one but twodimensions One question is whether it is plausible that infantsrsquo cognitive capacitiesinclude distinctions at the ontological level The other is whether developmentalchange is better captured by high-level symbolic or low-level computational descrip-tions As it happens it seems that answers to these two questions are oftencorrelated in the literature The resulting situation can be summarized (caricaturedperhaps) as this matrix of options

No prior ontological distinctions Some prior ontologicaldistinctions

1 Classical empiricist models of 2 Most models ofSymbolic description cognitive development domain-speci cityComputational 3 Most connectionist models of 4 description cognitive development

The pitfalls of classical empiricist accounts are familiar to most developmentalpsychologists As a result the only developmental accounts available either takelow-level computational aspects into consideration and assume no initial stageontological distinctions or else accept such distinctions and ignore issues of compu-tational implementation But there is nothing necessary in these conjunctionsIndeed the evidence seems to require a domain-speci c computational account

To illustrate these two issues consider for instance the evidence for animacydetection Infants detect the difference between animate and inanimate motion aspreferential looking and dishabituation protocols have shown To most developmen-tal psychologists it would seem contrived to suggest that the distinction betweenphysically and intentionally caused motion is extracted from other informationThere seems to be no available lower-level information that would correlate with thisdifference between types of motion and orient a cognitive system toward the motioncues

Considering rst the issue of concepts-as-descriptions is this evidence for thefact that the infant possesses the ldquoanimate agentrsquo concept The infantrsquos distinctiondoes amount to an ontological distinction in that it picks outmdashand applies differentinferences tomdashtwo different kinds of stuff in the world The kinematic cues used byinfants are among those used by adults to distinguish between agents and non-agentsin their environment The rather crude cues used by infants really are evidence forvery early biases but they will undergo continuous complexi cation with develop-ment If we adopt a descriptionist account of concepts we may wonder at whatpoint the childrsquos ontological distinctions actually amount to possessing the ldquoanimateagentrdquo concept If we construe concepts as skills this question does not make muchsense Concepts mature in the sense that a cognitive system becomes better skilledat detecting sameness and avoiding false positives As it happens this calibration

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

290 P BOYER

process is precisely documented in the domain of animacy detection (1) infants aresensitive to actual kinematic differences (2) pre-schoolers are apt to make inferencesabout possible motion from static cues (Gelman et al 1983) and (3) older childrenand adults use all this plus information about internal structure and biologicalfunctioning to decide what can and what cannot move intentionally (Carey 1988)Whether infants do or do not have the concept of ldquoanimate agentrdquo may well be a redherring introduced by descriptionist accounts of concepts

Turning to the second ambiguity this example also shows that accepting initialontological presumptions does not con ne developmental accounts to symbolicmodels that ignore implementation issues Speci c neural capacities underpin thedistinction between animate and inanimate causation (or if one prefers a lessmetaphysical description the typical kinematic difference between these two kindsof events) We now have a precise description of the relevant computational aspectsof perception (Ahlstrom et al 1997 Bellefeuille amp Faubert 1998 Heptulla-Chatterjee et al 1996) as well as of the neural structures involved (Bonda et al1996) A neuro-computational account of developmental processes for this ontolog-ical distinction may be available very soon (see Johnston et al 1999 for an attempt)

One could make similar points for the domain of face-recognition and own-lan-guage distinction Infants distinguish faces from other kinds of visual displays on thebasis of speci c cues the infantrsquos visual system is biased to detect This has promptedthe question whether these cues are evidence for a ldquoface conceptrdquo (see Morton etal 1990 Kleiner 1993 for different interpretations of the evidence) Again thequestion makes sense only in a descriptionist account of the ldquofacerdquo concept Thatthe bias is mostly driven by particular low-level visual features that happen to bepresent in typical faces does not rule out that it constitutes a (primitive) facedetection skill Also taking these low-level biases into consideration takes us closerto a computational account of initial dispositions for face-recognition Finallyinfants seem predisposed to detect their mother tongue (literally their motherrsquoslanguage) But they do this on the basis of very rough characteristics of linguisticinformation that are certainly insuf cient to make accurate distinctions betweenlanguages (Jusczyk et al 1993) Again the infantrsquos skills in the domain are at aprimitive stage This does not entail that they are not specialized skills

In general most early ontological distinctions seem to have the followingcharacteristics (1) they are continuous with some adult distinctions (which is whythey are often described as initial evidence for the adult concept) yet (2) they aremuch less ef cient than more mature concepts in terms of sameness tracking (3)they allow rich information pick-up given an initial bias and (4) this bias can beeasily modelled in terms of network dynamics This last point is particularlyimportant for biologically plausible developmental models A strong argumentagainst attributing a rich conceptual repertoire to infants is that such attributions arein general computationally intractable (Thelen amp Smith 1995) That is there is noway to determine what computational processes would underpin the precisebehavioural differences observed By contrast we see here that ontological distinc-tions may be founded on low-level detection that may be easily modelled in termsof network initial topology and biases This interpretation of early-developed cate-

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 291

gories and inference engines is consistent with Churchlandrsquos notion that suchstructures are ldquotheoreticalrdquo and reduce to dynamic prototype activation However italso suggests that some biases in the dynamics of prototype building are soentrenched that we should not be surprised that their long-term effects in terms ofconceptual dispositions are extremely dif cult to override

Intuitive ontology as an evolved natural metaphysics

Developmental evidence suggests a mentally represented ontological repertoire withthe following characteristics (1) it consists of metaphysical presumptions aboutparticular domains of reality rather than a generalized epistemic stance (2) thesepresumptions are intuitive (best accessed via implicit measures) and theoretical(computing non-trivial inferences) (3) some of the presumptions are available atvery early stages of cognitive development indeed in infancy (4) differences incultural environment or training do not seem to effect profound changes in intu-itively available categories and principles and (5) intuitive ontology is far fromoptimal Intuitive ontology just does not deliver any fast intuitive inferential re-sponses for some categories of objects and aspects of experience [5] and it oftenincludes ad hoc ldquobuck-stoppingrdquo assumptions or causal inferences without a descrip-tion of causal mechanism [6]

These features are not too surprising if we consider intuitive ontologicaldistinctions and inference engines as part of the cognitive equipment typical of thespecies as a result of evolution by natural selection (Boyer 1998ab Cosmides ampTooby 1994) This is not the place to discuss evolutionary constraints on this orthat particular capacity However it is relevant to ldquonaturalized epistemologyrdquoarguments to consider some general properties one might expect from a biologicallyimplemented evolutionarily constrained cognitive system

First it seems plausible from an evolutionary standpoint that ontologicaldispositions put developing minds on the ldquorightrdquo track by providing them with a(relatively) sound ontology rather than a (relatively) sound epistemology Having theright ontology as long as your epistemology is not dramatically unsound puts youon the right track without much cognitive expense Having a feeble ontology and theright epistemology on the other hand could be disastrous An organism equippedin this way would waste much cognitive power discovering valid generalizationsabout its environment that are not of any consequence for gene transmission

Second there are obvious computational advantages to a set of fastidenti cation and inference processes that deliver intuitive expectations Indepen-dent evolutionary considerations support the notion of cognitive evolution as anaccretion and complexi cation of specialized problem-solving algorithms rather thanas the development of an integrated general intelligence This is supported both byinter-speci c comparisons and by the computational cost and negative consequencesof developing an unbiased domain-general correlation-detector (Cosmides ampTooby 1994) [7] Also the presence of such biases makes it less surprising thatintuitive ontology is sub-optimal being only as good as it takes to provide anadaptive edge over previous versions

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

292 P BOYER

Third empirically substantiated evolutionary considerations may better explainwhy ontological presumptions develop early and initially appear as rough biases thatonly partially map onto adult conceptual distinctions The fact that a great amountof developmentally crucial information is stored in a speciesrsquo typical environmentimplies that minimal but de nite biases are suf cient to trigger adequate develop-ment This may help us better understand the relevance of neural plasticity at thelevel of neural structures and neural connectivity which is sometimes considered asevidence for the general plasticity of cognitive functioning However it is a generalprinciple of ontogenetic development that low-level plasticity is recruited in theservice of higher-level predictable design Indeed it makes such predictable formmore likely through calibration

Fourth an evolutionary framework accounts for the stubborn aspects of intu-itive ontology without implying that they are present in full-blown form at birth orinevitable in human development The fact that de nite biases lead to a particularontology is a reliable feature of development under normal conditions Childrenbrought up in zero gravity or on some planet with no animals may well developdifferent ontological concepts and a different intuitive physics or intuitive biologyWhat makes some conditions ldquonormalrdquo in this account is not that they are inevitableor even common It is only that such conditions prevailed in the environment wheregenes that control for neural structure and development were selected

To conclude the evidence suggests that concept acquisition most likely doesnot require a sound epistemic attitude but rather some robust metaphysical preju-dices supporting quasi-theoretical inference processes that are in principle defeasi-ble but in practice extremely stubborn If such prejudices are a plausible outcome ofevolution through natural selection ldquonatural metaphysicsrdquo is the outcome of aspeciesrsquo particular history and of the species-typical needs of its members and istherefore natural only relative to that particular species This certainly bodes ill foran extreme version of the ldquonaturalizing epistemologyrdquo project for we cannot deducemetaphysical validity from what is adaptive for a particular kind of organism withparticular requirements On the other hand cognitive and evolutionary consider-ations may lead to a more ldquonaturalrdquo version of the project providing an empiricallymotivated account of the relative ldquo trdquo between our categories and some naturalregularities

Notes

[1] This can be seen as the consequence of an implicit principle that nouns probably refer to kinds andldquoshape is often though not always a good source of information about object kindrdquo (Soja et al1992 p 102) As a consequence children are more likely to override shape similarity for livingkinds than for other ontological categories (Becker amp Ward 1991)

[2] The property of being ldquoaliverdquo is poorly understood before the age of 10 and then mostly on thebasis of culturally transmitted information (Carey 1985) Also young children explain somecharacteristics of animals either in terms of psychological dispositions (Inagaki amp Hatano 1987) ormechanistically (Au amp Romo 1998)

[3] A trained musician can ldquohearrdquo in auditory imagery the change induced by adding a B- at sound to

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 293

a C-major triad for instance (Churchland 1995) astronomical competence can make one ldquoseerdquothe planets orbiting on the ecliptic plane (Churchland 1979 p 32ff)

[4] It is perhaps the case that the so-called ldquoexpertsrdquo who participate in such experiments are nottrained enough perhaps ldquorealrdquo experts would perform differently This however only strengthensthe point that reeducating intuitions if it happens at all seems to require massive instructionaleffort

[5] This is clear in childrenrsquos expectations about plants for instance It is only gradually that the childassimilates PLANTS in a broad domain of LIVING THINGS Such ldquoholesrdquo in the ontologicallandscape can be found in the adult too Problems to do with the shape of the Earth are not theobject of any intuitive expectation They are handled by explicit socially transmitted meta-repre-sented knowledge

[6] People can routinely explain the behaviour of animals in terms of essential properties withoutnecessarily having a description of what these inner properties consist of or how they governbehaviour A species has its ldquoessencerdquo and thatrsquos that Scienti c theories typically challenge thiskind of ldquobuck-stoppingrdquo and try to describe the processes whereby underlying entities like FORCEor ESSENCE cause observable properties Doing this science often shows that the postulates werevacuous or entirely ad hoc

[7] This does not mean that an evolved architecture excludes general capacities but that the latterprobably ride piggyback on a host of specialized capacities and often handle meta-representationsof what intuitive ontology delivers (Mithen 1996 Rozin 1976)

References

AHLSTROM V BLAKE R amp AHLSTROM U (1997) Perception of biological motion Perception 261539ndash1548

ANTELL E amp KEATING DP (1983) Perception of numerical invariance in neonates Child Development54 695ndash701

ATRAN S (1989) Basic conceptual domains Mind and Language 4 5ndash16ATRAN S (1996) Modes of thinking about living kinds science symbolism and common sense

In D OLSON amp N TERRANCE (Eds) Modes of thought explorations in culture and cognition New YorkCambridge University Press

ATRAN S amp MEDIN D (Eds) (1999) Folkbiology Cambridge MA MIT PressAU TKF amp ROMO L (1998) Mechanical causality in childrenrsquos folkbiology In S ATRAN amp D MEDIN

(Eds) Folkbiology Cambridge MA MIT PressAVIS M amp HARRIS P (1991) Beliefndashdesire reasoning among Baka children evidence for a universal

conception of mind Child Development 62 460ndash467BAILLARGEON R (1987) Young infantsrsquo reasoning about the physical and spatial characteristics of a

hidden object Cognitive Development 2 179ndash200BAILLARGEON R amp HANKO-SUMMERS S (1990) Is the top object adequately supported by the bottom

object Young infantsrsquo understanding of support relations Cognitive Development 5 29ndash53BARRETT JL amp KEIL FC (1996) Conceptualizing a non-natural entity anthropomorphism in God

concepts Cognitive Psychology 31 219ndash247BECKER AH amp WARD TB (1991) Childrenrsquos use of shape in extending novel labels to animate

objects identity versus postural change Developmental Psychology 6 3ndash16BELLEFEUILLE A amp FAUBERT J (1998) Independence of contour and biological-motion cues for

motion-de ned animal shapes Perception 27 225ndash235BONDA E PETRIDES M OSTRY D amp EVANS A (1996) Speci c involvement of human parietal

systems and the amygdala in the perception of biological motion Journal of Neuroscience 163737ndash3744

BOYER P (1998a) Cognitive tracks of cultural inheritance how evolved intuitive ontology governscultural transmission American Anthropologist 100 876ndash889

294 P BOYER

BOYER P (1998b) If ldquotrackingrdquo is category-speci c a ldquocommon structurerdquo may be redundantcommentary on R Millikanrsquos article Behavioral and Brain Sciences 21 67ndash68

BULLOCK M (1985) Animism in childhood thinking a new look at an old question DevelopmentalPsychology 21 217ndash225

CAREY S (1985) Conceptual change in childhood Cambridge MA MIT PressCAREY S (1988) Conceptual differences between children and adults Mind and Language 3 167ndash181CHURCHLAND PM (1979) Scienti c realism and the plasticity of mind Cambridge Cambridge

University PressCHURCHLAND PM (1981) Eliminative materialism and the propositional attitudes Journal of

Philosophy 78 67ndash90CHURCHLAND PM (1988) Perceptual plasticity and theoretical neutrality Philosophy of Science 55

167ndash187CHURCHLAND PM (1989) The neurocomputational perspective Cambridge MA MIT PressCHURCHLAND PM (1995) The engine of reason the seat of the soul Cambridge MA MIT

PressCOSMIDES L amp TOOBY J (1994) Origins of domain-speci city the evolution of functional organiza-

tion In LA HIRSCHFELD amp S GELMAN (Eds) Mapping the mind domain-speci city in culture andcognition New York Cambridge University Press

DAMASIO AR (1990) Category-related recognition defects as a clue to the neural substrates ofknowledge Trends in Neuroscience 3 95ndash98

EIMAS PD (1994) Categorization in early infancy and the continuity of development Cognition 5083ndash93

ELMAN JL BATES EA JOHNSON MH KARMILOFF-SMITH A PARISI D amp PLUNKETT T (1996)Rethinking innateness a connectionist perspective

FAULKENDER PJ (1974) Genelized habituation of concept stimuli in toddlers Child Development 451002ndash1010

FODOR JA (1988) A reply to Churchlandrsquos ldquoPerceptual plasticity and theoretical neutralityrsquo Philosophyof Science 55 188ndash198

GALLISTEL CR amp GELMAN R (1992) Preverbal and verbal counting and computation Cognition 4479ndash106

GELMAN R (1990) First principles organize attention and learning about relevant data number andthe animatendashinanimate distinction as examples Cognitive Science 14 79ndash106

GELMAN R SPELKE E amp MECK E (1983) What pre-schoolers know about animate and inanimateobjects In DS ROGERS (Ed) The acquisition of symbolic skills London Plenum

GELMAN S amp MARKMAN E (1986) Categories and induction in young children Cognition 23183ndash209

GELMAN SA GOTTFRIED GM amp COLEY J (1994) Essentialist beliefs in children the acquisitionof concepts and theories In LA HIRSCHFELD amp SA GELMAN (Eds) Mapping the minddomain-speci city in cognition and culture Cambridge Cambridge University Press

GIGERENZER G (1991) How to make cognitive illusions disappear beyond ldquoheuristics and biasesrdquoEuropean Review of Social Psychology 2 83ndash115

GIGERENZER G amp HOFFRAGE U (1995) How to improve Bayesian reasoning without instructionfrequency formats Psychological Review 102 684ndash704

GIGERENZER G amp MURRAY D (1987) Cognition as intuitive statistics Hillsdale NJ ErlbaumGOPNIK A (1993) How we know our minds the illusion of rst-person knowledge of intentionality

Brain and Behavioral Sciences 16 1ndash14GOPNIK A amp WELLMANN H (1994) The theory theory In LA HIRSCHFELD amp SA GELMAN (Eds)

Mapping the mind domain-speci city in cognition and culture New York Cambridge University PressHEELAS P amp LOCK A (1981) Indigenous psychologies The anthropology of the self New York Academic

PressHEPTULLA-CHATTERJEE S FREYD JJ amp SHIFFRAR M (1996) Con gural processing in the perception

of apparent biological motion Journal of Experimental Psychology Human Perception andPerformance 22 916ndash929

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 295

HIRSCHFELD LA (1989) Rethinking the acquisition of kinship terms International Journal of BehavioralDevelopment 12 541ndash68

HIRSCHFELD LA (1996) Race in the making cognition culture and the childrsquos construction of human kindsCambridge MA MIT Press

HIRSCHFELD LA amp GELMAN SA (Eds) (1994) Mapping the mind domain-speci city in culture andcognition New York Cambridge University Press

HIRSCHFELD LA amp GELMAN SA (1999) Essentialism In S ATRAN amp M D (Eds) Folk-biologyCambridge MA MIT Press

INAGAKI K amp HATANO G (1987) Young childrenrsquos spontaneous personi cation as analogy ChildDevelopment 58 1013ndash1020

JOHNSTON A MCOWAN PW amp BENTON CP (1999) Robust velocity computation from a biologi-cally motivated model of motion perception Proceedings of the Royal Society of London B BiologicalSciences 266 509ndash518

JUSCZYK PW FRIEDERICI AD WESSELS J SVENKERUD VY amp JUSCZYK AM (1993) Infantsrsquosrecognition of foreign versus native language words Journal of Memory and Language 32 402ndash420

KAISER MK JONIDES J amp ALEXANDER J (1986) Intuitive reasoning about abstract and familiarphysics problems Memory and Cognition 14 308ndash312

KEIL FC (1986) The acquisition of natural kind and artefact terms In AWD MARRAR (Ed)Conceptual change Norwood NJ Ablex

KEIL FC (1994) The birth and nurturance of concepts by domains the origins of concepts of livingthings In LA HIRSCHFELD amp SA GELMAN (Eds) Mapping the mind domain-speci city in cognitionand culture New York Cambridge University Press

KLEINER KA (1993) Speci c vs non-speci c face recognition device In B DE BOYSSON-BARDIES SDE SCHONEN P JUSCZYK P MACNEILAGE amp J MORTON (Eds) Developmental neuro-cognition speechand face processing in the rst year of life Dordrecht Kluwer Academic Press

KOEHLER JJ (1996) The base rate fallacy reconsidered descriptive normative and methodologicalchallenges Behavioral and Brain Sciences 19 1ndash53

KORNBLITH H (1993) Inductive inference and its natural ground Cambridge MA MIT PressLANDAU B SMITH LB amp JONES SS (1988) The importance of shape in early lexical learning

Cognitive Development 3 299ndash321LAWSON ET amp MCCAULEY RN (1990) Rethinking religion connecting cognition and culture

Cambridge Cambridge University PressLESLIE A (1979) The representation of perceived causal connection Unpublished PhD Philosophy thesis

Oxford UniversityLESLIE A (1988) The necessity of illusion perception and thought in infancy In L WEIZKRANTZ (Ed)

Thought without language Oxford Clarendon PressMANDLER J BAUER P amp MCDONOUGH L (1991) Separating the sheep from the goats differentiating

global categories Cognitive Psychology 23 263ndash298MARTIN A WIGGS CL UNGERLEIDER LG amp HAXBY JV (1996) Neural correlates of category-

speci c knowledge Nature 379 649ndash652MASSEY C amp GELMAN R (1988) Pre-schoolersrsquo ability to decide whether pictured unfamiliar objects

can move themselves Developmental Psychology 24 307ndash317MCCLOSKEY M (1983) Naive theories of motion In D GENTNER amp AL STEVENS (Eds) Mental

models Hillsdale NJ Lawrence ErlbaumMEHLER J JUSCZYK P LAMBERTZ G HALSTED N BERTONCINI J amp AMIEL-TISON C (1988) A

precursor of language acquisition in young infants Cognition 29 143ndash178MELTZOFF A (1994) Imitation memory and the representation of persons Infant Behavior and

Development 17 83ndash99MELTZOFF A amp MOORE MK (1983) Newborn infants imitate adult facial gestures Child Development

54 702ndash709MILLIKAN RG (1998) A common structure for concepts of individuals stuffs and real kinds more

mama more milk more mouse Behavioral and Brain Sciences 21 55ndash100MITHEN S (1996) The prehistory of the mind London Thames amp Hudson

296 P BOYER

MORTON J amp JOHNSON M (1991) CONSPEC and CONLERN a two-process theory of infantface-recognition Psychological Review 98 164ndash181

MORTON J JOHNSON MH amp MAURER D (1990) On the reasons for newbornrsquos responses to facesInfant Behavioral Development 13 99ndash103

NELSON K (1973) Some evidence for the cognitive primacy of categorization and its functional basisMerrillndashPalmer Quarterly 19 21ndash40

PERNER J (1991) Understanding the representational mind Cambridge MA MIT PressPREMACK D (1990) The infantrsquos theory of self-propelled objects Cognition 36 1ndash16PREMACK D amp JAMES-PREMACK A (1995) Intention as psychological cause In D SPERBER

D PREMACK amp A JAMES-PREMACK (Eds) Causal cognition A multidisciplinary debate OxfordClarendon Press

QUARTZ SR amp SEJNOWSKI TJ (1997) The neural basis of cognitive development a constructivistmanifesto Behavioral and Brain Sciences 20 537ndash596

QUINE WVO (1969) Ontological relativity and other essays New York Columbia University PressQUINN PC amp JOHNSON MH (1997) The emergence of perceptual category representations in young

infants a connectionist analysis Journal of Experimental Child Psychology 66 236ndash263RICHARDS DD amp SIEGLER RS (1986) Childenrsquos understanding of the attributes of life Journal of

Experimental Child Psychology 42 1ndash22ROSS G (1980) Categorization in 1- to 2-year olds Developmental Psychology 16 391ndash396ROZIN P (1976) The evolution of intelligence and access to the cognitive unconscious In JM

SPRAGUE amp AN EPSTEIN (Eds) Progress in psychobiology and physiological psychology New YorkAcademic Press

RUMELHART DE amp TODD PM (1992) Learning and connectionist representations In DE MEYER

amp S KORNBLUM (Eds) Attention and performance XIV synergies in experimental psychology arti cialintelligence and cognitive neuroscience Cambridge MA MIT Press

SARTORI G amp JOB R (1988) The oyster with four legs a neuropsychological study on the interactionof visual and semantic information Cognitive Neuropsychology 5 105ndash132

SOJA NN CAREY S amp SPELKE ES (1992) Perception ontology and word-meaning Cognition 45101ndash107

SPELKE ES (1988) The origins of physical knowledge In L WEIZKRANTZ (Ed) Thought withoutlanguage Oxford Oxford University Press

SPELKE ES (1990) Principles of object perception Cognitive Science 14 29ndash56SPERBER D (1994) The modularity of thought and the epidemiology of representations In LA

HIRSCHFELD amp SA GELMAN (Eds) Mapping the mind domain-speci city in cognition and culture NewYork Cambridge University Press

STARKEY D (1981) The origins of concept-formation object sorting and object preference in earlyinfancy Child Development 52 489ndash497

STARKEY P SPELKE ES amp GELMAN R (1990) Numerical abstraction by human infants Cognition36 97ndash127

THELEN E amp SMITH LB (1995) A dynamic systems approach to the development of cognition and actionCambridge MA MIT Press

TVERSKY A amp KAHNEMANN D (1982) Evidential impact of base rates In D KAHNEMANN P SLOVIC

amp A TVERSKY (Eds) Jugdments under uncertainty heuristics and biases Cambridge CambridgeUniversity Press

TVERSKY B amp HEMENWAY K (1984) Objects parts and categories Journal of Experimental PsychologyGeneral 113 169ndash193

WALKER SJ (1985) Atimodemo Semantic conceptual development among the Yoruba Unpublished PhDdissertation Cornell University

WASON PC (1983) Realism and rationality in the selection task In JSBT EVANS (Ed) Thinking andreasoning psychological approaches London Routledge Chapman amp Hall

WASON PC amp JOHNSON-LAIRD PN (1972) The psychology of reasoning structure and content LondonBatsford

WELLMANN H (1990) The childrsquos theory of mind Cambridge MA MIT Press

NATURAL EPISTEMOLOGY OR EVOLVED METAPHYSICS 297

WELLMANN H amp GELMAN SA (1992) cognitive development foundational theories of core domainsAnnual Review of Psychology 43 337ndash375

WYNN K (1992) Evidence against empiricist accounts of the origins of numerical knowledge Mindand Language 7 315ndash332

YOUNGER B (1985) The segregation of items into categories by ten-month-old infants Child Develop-ment 56 1574ndash1583

YOUNGER B (1990) Infant categorization memory for category-level and speci c item informationJournal of Experimental Child Psychology 50 131ndash155