■ One should not throw out the baby with thebathwater, according to an old aphorism. Somepopular recent positions in AI thinking havedone just this, we suggest, by rejecting the usefulidea of mental representations in their overen-thusiastic zeal to correct some simplificationsand naïveties in the way traditional AI ideas havesometimes been understood. These “situated”perspectives correctly emphasize that agents livein a social world, using their environments tohelp guide their actions without needing toalways plan their futures in detail; but theyincorrectly conclude that the very idea of mentalrepresentation is mistaken. This perspective hasits intellectual roots in parts of recent sociologi-cal thinking which reject the entire fabric ofwestern science. We discuss these ideas and dis-putes in the form of an illustrated fable concern-ing nannies and babies.

A Cast of Characters

Once upon a time there were two hap-py and healthy babies. We will callthem Representation Baby (closely

related to Mind Baby and Person Baby) andScience Baby (closely related to Reality Baby).

These babies were so charming and inspira-tional that for a long time their nanniescared for them very well indeed. During thisperiod it was generally the case that igno-rance was pushed back and human dignityincreased. Nannies used honest, traditionalmethods of baby care which had evolved dur-ing the years. Like many wise old folk, theywere not always able to articulate good justi-fications for their methods, but they worked,and the healthy, happy babies were growingwell and having lots of fun.

Unfortunately, some newer nannieshaven’t been so careful, and the babies are indanger from their zealous ways. We will focuson two nannies who seem to be close friendsand often can be seen together—SituatedNanny (called SitNanny for short) and Radi-cal Social Constructivist Nanny (known toher friends as RadNanny).1

SitNanny Meets RepBabySitNanny is fanatical about a certain kind offirmness. She just hates for things not to befirmly attached to the world. To be fair, thisobsession may be a natural consequence ofher having spent so much time in California.

She believes that traditional ways of bring-ing up baby are far, far too slack in thisregard. Babies shouldn’t think and plan somuch, she believes, but instead should bekept firmly attached to the ground.2

SitNanny declares (in fact, she preaches,loudly and regularly) that all this old-fash-ioned cognitive stuff is just dirty bathwater,and she is going to throw it out. She believesthat it holds babies back, stunting theirdevelopment, and that a nursery based onsuch nonsense is an unhealthy place. She isso fanatical about her views, however, thatshe is quite willing to throw out the babywith the bathwater.

Not long ago, Representation Baby (alsocalled RepBaby) was reflecting on a heretoforehappy childhood when rather suddenly hefound himself threatened by SitNanny, whoadopts just this position.

SitNanny argues passionately with thosewho suggest that keeping the baby might be awise idea, going so far as to deny that there isa baby there at all. She argues further that her

Articles

FALL 1994 15

On Babies and BathwaterA Cautionary Tale

Patrick J. Hayes, Kenneth M. Ford, & Neil Agnew

Copyright © 1994, AAAI. 0738-4602-1994 / $2.00

Articles

16 AI MAGAZINE

Situated Nanny Holding Things Down

task is simply to get the nursery clean inpreparation for other babies.

We think that SitNanny is at heart a goodperson, and many of her suggested reformsare valuable. The intellectual nursery didneed some tidying up. But she has been read-ing political tracts (written by such people asDreyfus and Searle—the Berkeley Brothers)and has become something of a fanatic.

Why would SitNanny want to threatenRepBaby, and what does it mean to be “situ-ated”? There are several possible answers.One emphasizes the fact that agents’ knowl-edge of their world is incomplete, partial andusually incorrect, but that they must never-theless act in that world, often quite prompt-ly. This view leads to a vision that rejects theidea of planning in order to act. Planning, itis said, makes the unjustified assumptionthat the planner has access to all relevantfacts, and it is a time-consuming activity thatignores the ticking of the real clock (Agre &Chapman, 1987). It would be hard to playbaseball, for example, if one had to be alwaysplanning one’s actions. It seems clear that weoften, perhaps usually, simply start on acourse of action in the blithe confidence thatwe will be able to handle any difficultieswhich may arise along the way. We just do it.And we often rely explicitly on the externalworld to guide our actions. If you are drivingin downtown San Francisco and want to getto the airport, the best method is to findHighway 101 south, get in the left-hand lane,and stay in that lane through all subsequentjunctions and turns. It will eventuallybecome the fast lane on the road to the air-port. There is no need to plan a route: thatleft-hand lane will take you there.

This disillusionment with the traditionalAI emphasis on planning combines naturallywith an interest in how simple organisms geton. SitNanny has noticed that while insects,for example, have very impoverished cogni-tive abilities and never seem to plan anythingin the traditional AI sense of the term, thereare an awful lot of them around. Even quitesimple organisms can exhibit surprisinglycomplex behavior when placed in certainsocial and physical settings—a point made byHerb Simon (one of RepBaby’s grandfathers)years ago with his example of the ant travers-ing a beach of pebbles.

Further, it seems insightful to ask howmuch human behavior arises from, or mightbe simulated by, mechanisms that involve lit-tle or no explicit representation of the exter-nal environment. Even if one is concernedwith representations and planning and so

forth, the zero case is of interest. Such a ques-tion certainly leads in a different directionthan the traditional AI work in planning andfocuses interest on different concerns. RepBa-by can live with this. It provides at the veryleast a certain intellectual discipline. Babies,even RepBabies, need to be set firmly on sol-id ground quite often and (one might argue)should learn to walk before they try to playchess.

However, the term “situated” has becomeidentified with a much stronger and moreradical collection of assumptions, in particu-lar an attack on the basic idea of knowledgerepresentation. We believe that these assump-tions are unwarranted and often based onfundamental misunderstandings. In particu-lar, one can agree that AI should pay moreattention to real-time activities that are inti-mately involved with the immediate environ-ment, without feeling a need to reject theentire framework of empirical science,embrace radical social constructivism, orreject the idea of mental representation. Mostof these more radical positions seem to arisefrom reactions against what might be calledstraw babies, versions of classical AI ideasthat in fact are much too simplistic. SitNan-ny seems to think that the AI concept of“physical symbol system” amounts to a claimthat programs, mental representations, andexternal texts are indistinguishable (e.g.,Clancey, 1993) and that the AI vision ofknowledge is of something rigid, unchange-able, and context-insensitive, a kind of “men-tal toolkit” (Lave, 1988). In our view, all thesepositions misunderstand the essential con-tent of the physical symbol system hypothe-sis as it is most widely understood in AI. Par-ticular AI systems (often, ironically, thosethat SitNanny was working on before she gother new perspective) may make such simpli-fying assumptions, but there is no reasonwhy the general idea of symbolic representa-tion must be so limited, and indeed it tradi-tionally has not been thought of in such alimited way.

The situationalists are attacking the veryidea of knowledge representation—thenotion that cognitive agents think abouttheir environments, in large part, by manipu-lating internal representations of the worldsthey inhabit. Let us be frank: we think therepresentational hypothesis is a great idea.The reasons for being so positive are welldocumented, but we have two main justifica-tions for our enthusiasm. First, it accounts formuch that is otherwise completely puzzlingabout how cognition could happen in the

The intellectualnursery didneed sometidying up.

Articles

FALL 1994 17

Articles

18 AI MAGAZINE

Representation Baby Is Worried about His Future

physical world; second, it allows experimentsand makes empirical predictions, which haveso far largely been confirmed. But one needsto understand the key word “representation”in a sufficiently broad fashion.

Internal representations might not be con-sciously available to introspection, might uti-lize ontological frameworks that are deter-mined by social or other contexts, might beinvolved with (and have their use involvedin) social practices or any other kind ofhuman activity, and might be involved inperceptual or motor skills at any cognitivelevel. None of these are in any way at vari-ance with the representationalist hypothesis.RepBaby can play happily with these toys.

What a representation means (not what“representation” means, notice) is a complexquestion, but a meaning might be relative tosuch things as goals, attitudes or purposes,and it need not presuppose that the environ-ment is uniquely determined. In fact, avail-able semantic accounts of representation lan-guages—model theories—emphasize theextent to which such relativity is probablyinevitable.

SitNanny speaks with many voices, but aprocess of reconstruction and guesswork sug-gests to us that SitNanny understands repre-sentation in a different sense from that inwhich it is used in the framework she rejectsso vehemently. In fact, several of the newcritics (some having noisily jumped ship)take representation to mean something like atext or a picture, an object consciouslymanipulated by talking to oneself or visualiz-ing internal imagery, and identical in natureto external representations such as writingsor diagrams. For example, Clancey (1993)assumes that the use of a representation isalways conscious and deliberate, and if a pro-cess is “pre-linguistic” then representationsare not involved in it. But RepBaby was notbrought up this way. It is at variance with theusual meaning of the term as used in the rep-resentationalist position in cognitive science,which hypothesizes internally representedknowledge in a much broader sense, includ-ing the computational modeling of uncon-scious cognitive processes and processes ofsocial interaction. One might believe thatthis broader sense is somehow inappropriateor incoherent, but such a position needs tobe argued, not simply asserted. Several of thecritics simply assume that calling somethinga representation entails that it has this exter-nal, perceived character. We will call this mis-take the “textual fallacy.”

Clancey (correctly) emphasizes that post

hoc justifications for action are not to beidentified with reasons for acting, apparentlybelieving that this confusion—which mayonce have been a common one in the expertsystem field in which he worked—is endemicto the representationalist perspective; but it isnot. John Searle makes a similar mistakewhen he concludes that anything represent-ed must be equated with a consciousthought, and any proposed mechanism ofunconscious activity must therefore notinvolve representations. Lucy Suchman(1987), a more subtle critic, also commits thismistake by drawing a sharp contrast betweencognitive science’s rigid, mechanical view ofplans and the weak resource view endorsedby situated action theory. However these arenot different views of a single idea, but twodistinct ideas: plans as programs or datastructures, and plans as external maps ortexts.

SitNanny has been accused of this errorbefore (Vera & Simon, 1993) and has protest-ed that she does not mean to throw away allnotions of symbols or mental representation,but here she misses the point. What makesRepBaby so useful is that he makes physicalsymbols provide an explanation of mentalphenomena, and this explanatory role iswhat SitNanny rejects. Her new clean bath-water will be scented with all manner ofthings: Gibsonian ideas of direct perception;concepts from fringe neurology, sociology,ethnomethodology, and political theory; pre-computational psychological theory; andGod knows what else. But the central idea ofphysical symbol system will be purged fromit (Clancey, 1993; Greeno & Moore, 1993).

Where’s the Beef?When listening to SitNanny tell us that repre-sentations are not in the head, we often won-der, “Where’s the beef?” According to SitNan-ny, not only are there no mentalrepresentations, but individuals do not pos-sess knowledge. SitNanny is unhappy withboth parts of the term “knowledge represen-tation.” Not only is the idea of representationseen as misleading and false, but even thenotion of knowledge itself, as we understandit, seems deeply suspect. Knowledge andmeaning are seen as extra-personal and locat-ed in the community rather than in the head.We need to get our terminology clear. Let’stake a simple example. A plumber and a cus-tomer who knows nothing about plumbingare together in a kitchen. Where is theknowledge of plumbing? RepBaby has an

SitNanny is unhappywith bothparts of the term “knowledgerepresen-tation.”

Articles

FALL 1994 19

and places the meanings not, as it were, dis-tributed among the heads of the members ofthe society, but in the society itself. Goingone step beyond Jung, we have to imaginesomething like an invisible collective con-sciousness—what Harry Collins calls a “col-lectivity”—in which the meanings of ourindividual beliefs are located. The old AI ideaof common sense might fit easily here: whatmore natural place to locate the sense com-mon to the members of a society than in thatsociety itself?

We are quite sympathetic to this view,properly understood. But notice the differ-ence between this idea and the earlier one.This idea says that meanings of mental repre-sentations are socially determined; the otherclaims that the representational tokens them-selves aren’t in the head or that representa-tional tokens can have only an external,social existence, or even that there isn’t anyrepresentation at all. This conflict is like thedifference between insisting that RepBabywear fashionable clothes and claiming thatthere isn’t anyone there to be dressed.

Suppose the plumber in the kitchen real-izes he has forgotten his flux and knows thata soldered joint might therefore be unreli-able, so he contemplates using a compressionfitting in spite of its extra cost. Perhapsindeed a proper account of what these termsmean might involve a description of theplumbing community and his relationship toit. Any experience, including this one, maychange and enrich his knowledge of joiningpipes. But these are mental tools that heuses—or, more properly, mental mechanismsout of which he is formed—and when heleaves the kitchen, they go with him.

In summary, the central concept of mentalrepresentation is more robust than realizedby SitNanny. We believe that SitNanny’sattack on RepBaby is unwarranted and oftenbased on fundamental misunderstandings.RepBaby is worth saving.

“Hey, Hey, Ho, Ho,Science Baby Must Go”

Now we come to Science Baby—an extremelyrobust and healthy lad. SitNanny doesn’tbear Science Baby much harm. But RadNan-ny, one of SitNanny’s mentors, is much moredangerous—RadNanny belongs to a cult.Remarkably, even this most useful baby isunder attack by RadNanny and the membersof her cult.

Like an anthropologist studying a primi-tive tribe, RadNanny (and her band of sociol-

almost childishly simple answer to this ques-tion: It’s in the plumber’s head! Is RepBabyright? Well, let us try to converge on thelocus of plumbing knowledge. If someonewere to ask a question about hot-water cylin-ders, it would be correct to tell him that theanswer could be found in the kitchen, so itseems correct to say that the knowledge is inthe kitchen.

Where in the kitchen is it, then? Perform asimple experiment by removing the plumberfrom the kitchen. Now nobody in thekitchen knows much about plumbing, and itwould be wrong to say that that was wherethe answer could be found. It seems clearthat the plumber’s knowledge of plumbing isfirmly attached to the plumber. Perhaps somekind of invisible mental leash would explainits attachment to him, but we suggest that asimpler and more plausible theory of thisphenomenon might be that the plumber’sknowledge of plumbing is in his head andtherefore moves with him, much as his kid-neys and his toenails do.

Memories provide an even more vivid illus-tration of beliefs being in the head. We allhave memories that no one else shares. Sure-ly, these are not located anywhere but in ourown heads (or at most our own bodies). Amemory can be damaged by banging a head,but never by stubbing a toe or digging a hole.

Notice this is not to claim that memoriesare veridical or that they are anything as sim-ple as a replaying of earlier recordings. Theymight be the result of some complex and cre-ative process of reconstruction with which weall constantly rewrite our past. And it might becorrect to say that we are almost never doingjust one thing, that our actions are alwaysembedded in some social context; althoughwe think this fact is less significant than Sit-Nanny does. But whatever is going on whenwe think or when we are recalling somethingfrom our past, it is happening inside us, notsomehow out there in the society.

There is, however, a rather different way tounderstand SitNanny’s social-meaning claim.One might argue that although there arethings in the head we call representations,any account of the meaning of these things(and hence any account of why they shouldbe regarded as representations) must beinherently social. Our beliefs have meaningonly by virtue of their role in a society ofwhich we are part. Wittgenstein argued thatthe idea of a language spoken by a single per-son is incoherent: languages are systems ofcommunication between agents. The socio-logical stance takes this idea rather further

We believethat

SitNanny’sattack on

RepBaby isunwarranted

and oftenbased on

fundamentalmisunder-standings.

RepBaby is

worth saving.

Articles

20 AI MAGAZINE

Articles

FALL 1994 21

Science Baby Doing His Sums

Articles

22 AI MAGAZINE

The Notorious RadNanny Looking for Babies

ogists) has moved into scientific laboratoriesto peer over the shoulders of scientists. Rad-Nanny has noticed that the behavior of realscientists typically does not follow theabstract ideal of disinterested pursuit of truth.However RadNanny draws the startling andextreme conclusion that all science is arbi-trary and that reality is merely a constructionof a social game.

This anti-epistemic and also anti-scientificstance is clear in statements such as “the nat-ural world has a small or nonexistent role inthe construction of scientific knowledge”(Collins 1981, p. 3) and “the fashioning ofnormative models of thinking from particu-lar, ‘scientific,’ culturally valued, named bod-ies of knowledge is a cultural act” (Lave 1988,p. 172). RadNanny claims that since we haveno direct access to “reality,” science (like reli-gion, politics, and literature) merely reflectsthe myth-making tendencies of the humantribe, with one myth being no more reality-based than another.

On this view, science should not be grant-ed any kind of relative authority or have anyparticular success acknowledged. It is just onecultural tradition: Zen Buddhism, medievalChristianity, or even the semi-coherent cyni-cism of Beavis and Butthead are all perfectlyvalid alternatives. In fact, sometimes scientif-ic rationality is seen as morally or politicallyinferior to Beavis and Butthead: “Logic, asMarx has it, is the money of the mind, andno matter how dialectical, it always expressesa reified and alienated mediation of man andreality” (Warren, 1984 p. 50, cited in Lave,1988 p. 173). We have criticized this culturalrelativism elsewhere (Agnew, Ford, & Hayes,1994), so here we will be brief.

By any reasonable measure, Science Babyhas been spectacularly successful. Perhaps nosingle scientist operates according to the ide-al rules, as RadNanny delights in telling us,but the approximate global effect is to pro-duce knowledge that reality (the hidden-hand editor) has had an opportunity to gripas firmly as it can be made to.

RadNanny claims that the natural worldplays little or no part in the beliefs of scien-tists, and that logic is only a tool used by oneclass to enslave others; but what can be thebasis for her beliefs other than empiricalobservation of the natural world (which, ofcourse, includes the scientists she studies),and what logic does she use to explain herideas? Are RadNanny’s sociological friendssomehow issued special glasses that allowthem to see more clearly than other scien-tists?

SitNanny often enthusiastically endorsesthis radical social constructivism because itcan be seen as locating knowledge in thecommunity rather than in the head and socan be taken to support a position that deniesa central role to internal representations. Wesometimes wonder if SitNanny really under-stands the full import of RadNanny’s doc-trines.

It is important to notice that it is not justScience Baby that is under attack here, butthe whole notion of an objective reality. Rad-Nanny’s ideas are a sort of universal baby sol-vent. However, we are not too worried aboutReality Baby. Even RadNanny will duck if onethrows a brick at her.

ConclusionThese new nannies have a lot of valuablereforms to make. We aren’t trying to arguesimply for conservatism. Sometimes the old-fashioned nannies were (and still are) usingdirty bathwater. One does need to keep Rep-Baby under control—if you let him, he willtry to get hold of everything.

SitNanny is right to emphasize that a lotcan be done without involving detailed,explicit knowledge of it. Sometimes the bestplan is just to follow the road. But thismusn’t be taken too far. It is really an obser-vation about what kind of informationshould be represented rather than a rejectionof the idea of knowledge representation itself.In any case, the question of how much plan-ning goes on is essentially an empirical one.Something clever happens when a good base-ball fielder begins to run toward a catchbefore the bat has made contact with theball. Maybe this isn’t planning in a very sim-plistic sense, but it isn’t just following the leftlane either.

Even RadNanny has some things of valueto give us. A certain kind of naïve realism iswrong, and human scientists sometimes doall sorts of strange human things. But if theradical constructivists would just take thetime to actually look at what the knowledgerepresentation idea means, they would seethat it has already passed that level ofnaïveté. If someone’s perceptions are repre-sentations, then of course they aren’t indirect touch with the world (whatever thatmight mean). RepBaby is a pragmatic con-structivist, but he gets on quite happily withReality Baby. Science Baby likes both of them.RadNanny, SitNanny, and others of height-ened social awareness have been doing somenew and exciting work that can rinse away a

Articles

FALL 1994 23

Articles

24 AI MAGAZINE

Reality Baby in a Relativist Stew

Articles

FALL 1994 25

Examining our Assumptions

papers in these areas. Hayes organized one ofthe first interdisciplinary cognitive scienceprograms, at the University of Rochester. Hehas held several academic and industrialpositions in aspects of AI. Hayes has been sec-retary of AISB, a chairman and trustee ofIJCAI, governor of the Cognitive ScienceSociety, and President of AAAI.

Ken Ford is currently Direc-tor of the Institute for theInterdisciplinary Study ofHuman and Machine Cogni-tion at the University of WestFlorida. His research interestsinclude the philosophicalfoundations of AI, knowl-

edge-based systems, issues of representation,and constructivist approaches to cognitivescience and AI. He has written more than ahundred scientific papers, and is the authorand editor of three books. He received hisPh.D. in computer science from Tulane Uni-versity. Ford is Editor-in-Chief of the AAAIPress, Executive Editor of the InternationalJournal of Expert Systems, Associate Editor ofthe Journal of Experimental and TheoreticalArtificial Intelligence, and is a Behavioral andBrain Sciences (BBS) Associate. He is past presi-dent of Florida Artificial Intelligence ResearchSociety (FLAIRS).

Neil Agnew is a professor ofpsychology at York University,Toronto Canada. His academ-ic interests include decisionmaking under uncertainty,models of individual andinstitutional change, and thetheory of science. He has pub-

lished extensively in these fields and has justpublished the sixth edition of his text TheScience Game. Neil has occupied various postsin both professional and academic associa-tions, and is a fellow of the Canadian Psycho-logical Association.

certain kind of conceptual grime that some-times clouds our thinking. It is healthy to beobliged to examine our own assumptionsfrom time to time.

But we must not overreact and throw awayeverything. Such absolute rejections reflectthinking that is just as rigid and dogmatic asthe simplistic naïveté they purport to over-come. People sometimes forget that the onlyreason for having nannies at all is to lookafter the babies.

Notes(1) There are those who say that these are not newnannies at all, but old nannies in new dresses. Forexample, it is sometimes said that SitNanny oftenlooks like an older nanny who was fired for cause,the notorious Nanny Skinner.

(2) Like insects, in fact (Brooks, 1991).

ReferencesAgnew, N.; Ford, K. M.; and Hayes, P. J. 1994.Expertise in Context: Personally Constructed,Socially Selected and Reality Relevant? InternationalJournal of Expert Systems 7(1): 65–88.

Agre, P., and Chapman, D. 1987. PENGI: An Imple-mentation of a Theory of Activity. In Proceedingsof the Seventh National Conference on ArtificialIntelligence, 268–272. Menlo Park, Calif.: AmericanAssociation for Artificial Intelligence.

Brooks, R. A. 1991. Intelligence without Represen-tation. Artificial Intelligence 47:139–159.

Clancey, W. 1993. Situated Action: A Neuropsycho-logical Interpretation. Cognitive Science 17:87–116.

Collins, H. M. 1981. Stages in the Empirical Pro-gramme of Relativism. Social Studies of Science11:3–10.

Greeno, J. G., and Moore, J. 1993. Situativity andSymbols. Cognitive Science 17:49–59.

Lave, J. 1988. Cognition in Practice: Mind, Mathemat-ics, Culture in Everyday Life. New York: CambridgeUniversity Press.

Suchman, L. 1987. Plans and Situated Actions: TheProblem of Human-Computer Communication. NewYork: Cambridge University Press.

Vera, A., and Simon, A. H. 1993. Situated Action: ASymbolic Interpretation. Cognitive Science 17:7–48.

Warren, S. 1984. The Emergence of Dialectical Theory.Chicago: University of Chicago Press.

Pat Hayes is a professor ofcomputer science and philos-ophy at the University ofIllinois at Urbana. Hayes’sresearch interests includeknowledge representation,machine inference, and thephilosophical foundations of

AI. He has published several influential

Articles

26 AI MAGAZINE