Full Terms & Conditions of access and use can be found athttp://www.tandfonline.com/action/journalInformation?journalCode=psns20

Social Neuroscience

ISSN: 1747-0919 (Print) 1747-0927 (Online) Journal homepage: http://www.tandfonline.com/loi/psns20

Detection of deception using fMRI: Better thanchance, but well below perfection

George T. Monteleone , K. Luan Phan , Howard C. Nusbaum , DanielFitzgerald , John-Stockton Irick , Stephen E. Fienberg & John T. Cacioppo

To cite this article: George T. Monteleone , K. Luan Phan , Howard C. Nusbaum , DanielFitzgerald , John-Stockton Irick , Stephen E. Fienberg & John T. Cacioppo (2009) Detection ofdeception using fMRI: Better than chance, but well below perfection, Social Neuroscience, 4:6,528-538, DOI: 10.1080/17470910801903530

To link to this article: https://doi.org/10.1080/17470910801903530

Published online: 25 Sep 2009.

Submit your article to this journal

Article views: 740

View related articles

Citing articles: 21 View citing articles

Detection of deception using fMRI: Better than chance,but well below perfection

George T. Monteleone, K. Luan Phan, Howard C. Nusbaum, Daniel Fitzgerald,and John-Stockton Irick

University of Chicago, Chicago, IL, USA

Stephen E. Fienberg

Carnegie-Mellon University, Pittsburgh, PA, USA

John T. Cacioppo

University of Chicago, Chicago, IL, USA

Functional brain imaging has been considered a new and better technique for the detection of deception.The reasoning is that there is a neural locus or circuit for lying that is sensitive, specific, generalizable acrossindividuals and measurement contexts, and robust to countermeasures. To determine the extent to whichthe group results predicted lying at the level of the individual, we reanalyzed data on 14 participants from astudy that had previously identified regions involved in lying (thus satisfying the criterion for sensitivity).We assessed the efficacy of functionally determined brain regions based on the lie�truth contrast for N � 1participants to detect deception in the Nth individual. Results showed that no region could be used tocorrectly detect deception across all individuals. The best results were obtained for medial prefrontalcortex (mPFC), correctly identifying 71% of participants as lying with no false alarms. Lowering thethreshold for a response increased hits and false alarms. The results suggest that although brain imaging is amore direct index of cognition than the traditional polygraph, it is subject to many of the same caveats andthus neuroimaging does not appear to reveal processes that are necessarily unique to deception.

Historian Ken Alder (2007) has described thepolygraph and lie detection as an ‘‘Americanobsession,’’ a pursuit steeped more in an Amer-ican cultural fantasy than reliable evidence.Traditional polygraphy, using an array of physio-logical measures such as heart-rate, blood pres-sure, and electrodermal response, faced severalfundamental problems: the sensitivity of the test,the effectiveness of the test in measuring what itpurports to measure (specificity), the general-izability of the test across measurement contexts,

and its robustness to countermeasures. In 2003,the Committee to Review the Scientific Evidenceon the Polygraph concluded:

Notwithstanding the limitations of the qualityof the empirical research and the limited abilityto generalize to real-world settings, we con-clude that in populations of examinees such asthose represented in the polygraph researchliterature, untrained in countermeasures, spe-cific-incident polygraph tests can discriminate

# 2008 Psychology Press, an imprint of the Taylor & Francis Group, an Informa business

Correspondence should be addressed to: George T. Monteleone, Center for Cognitive and Social Neuroscience, University of

Chicago, 5848 S. University Avenue, Chicago, IL 60637, USA

We thank Alex Japour for his contributions to this research. Support for this research was provided by the Department of Army

Research Grant #DAAD19-03-1-0042 and the MacArthur Law & Neuroscience Project.

SOCIAL NEUROSCIENCE, 2009, 4 (6), 528�538

www.psypress.com/socialneuroscience DOI:10.1080/17470910801903530

lying from truth telling at rates well abovechance, though well below perfection. (Com-mittee to Review the Scientific Evidence on thePolygraph, 2003, p. 4)

New areas of research have since taken on thechallenge of detecting deception by moving toneurophysiological measures, including the phy-siological detection of deception using functionalmagnetic resonance imaging (fMRI). The notionis that by imaging the brain processes involved indeception, the problems of traditional polygraphycan be avoided.

Several fMRI studies of lie detection havebeen conducted using binary-choice paradigmsand nomothetic (group-level) analyses to deter-mine whether it was possible to discriminate lyingand truthful responding. Spence et al. (2001)reported a study in which participants eitherlied or told the truth about acts they hadperformed during the course of that day. A cuepresented along with the question signaled tothe participant whether to lie or tell the truth.Blood-oxygen-level dependent (BOLD) contrastsbetween lie and truth conditions revealed signifi-cantly greater activation during lying in thebilateral ventral prefrontal cortex (VLPFC), dor-solateral prefrontal cortex (DLPFC), medial pre-frontal cortex (mPFC), and the left inferiorparietal cortex. Spence et al. (2001) interpretedthe results in terms of learning and motorresponse inhibition. Using an in-scanner adapta-tion of the Guilty Knowledge Test (GKT) withplaying cards as stimuli, Langleben et al. (2002)instructed participants to either lie or tell thetruth about cards they had in their possession.Results revealed a greater activation in theDLPFC and in the medial frontal gyrus extendingto the anterior cingulate cortex (ACC) duringlying than truth-telling. Also using the GKT, Phanet al. (2005) found lying to be associated withgreater activity in the mPFC, bilateral DLPFCand VLPFC, and bilateral superior temporal sulci(STS). Activation in the mPFC and DLPFC wereactivated differentially by lying and truth-tellingin all three studies.

Brain imaging has also been performed usingexperimental paradigms to simulate more realis-tic lying situations. Lee et al. (2002) used a forced-choice memory paradigm in which participantsfeigned memory impairments on two memorytests (digit memory task and autobiographicalmemory task) in a simulated assessment ofmalingering. The participants were requested to

devise their own mental strategies for the lie,thereby increasing the external validity of themanipulation of lying. The activation for feignedmemory impairment versus accurate recall con-ditions was similar for the two memory tasks andincluded anterior frontal regions (BA 9 and 10),bilateral parietal and temporal cortex, and sub-cortical regions including the caudate. The pre-frontal region result suggested a role ofinformation manipulation and executive control,with anterior regions (BA 10) participating insimultaneous processing of primary and second-ary goals, and the DLPFC region processinganticipation of performance and working mem-ory. The authors suggested that the activation ofthe parietal region was part of a network involvedin calculated responses during lying, the temporalregions as participating in the cognitive manip-ulation of visual information, and the subcorticalregions as involved in self-monitoring and inhibi-tion of learned rules. A study by Nunez, Casey,Egner, Hare, and Hirsch (2005) investigated theneural correlates of false versus true responses toautobiographical or non-autobiographical infor-mation. The investigators reported a contrasteffect for autobiographical items in the mPFC,DLPFC, VLPFC, ACC, BA 9/10, and caudate. Atthe chosen levels of significance, however, nodifferences were found between lying andtruth for non-autobiographical items, suggestingdifferences in neural response for lying dependingon the content being tested. In an experimentalstudy to assess the effects of the type of lie told,Ganis, Kosslyn, Stose, Thompson, and Yurgelun-Todd (2003) investigated lies that were rehearsedand spontaneous lies. Lies generally were asso-ciated with greater activation in BA 10, fusiformgyrus, and visual cortex whereas activation for lieswas greater than truth-telling in mPFC andDLPFC only for spontaneous lies.

Across these studies, activity in the mPFC andDLPFC was greater for lying than truthfulresponding in all except the conditions in Ganiset al. (2003) in which the lies were autobiogra-phical and well-rehearsed. It is not clear whetherthese regions are indicative of lying or areindicative of the increased self-monitoring andload on working memory that are often involvedwhen one is lying. These studies have also focusedon nomothetic analyses of the BOLD response asa function of lying versus truth-telling, wherelying and truth-telling are not only known but arecontrolled by the experimenter. The criticalquestion, which has limited the scientific validity

DETECTION OF DECEPTION USING FMRI 529

of the traditional polygraph, is whether lying canbe discriminated from truth-telling on an indivi-dual level based on the observed physiologicalresponse when one is answering the criticalquestion.

The first fMRI study endeavoring to discrimi-nate lying from truth-telling at the idiographiclevel was reported by Kozel et al. (2005). BOLDresponses were used to develop a mathematicalmodel weighting regions of interest (ROIs) todistinguish lying from truth-telling, and a secondgroup’s BOLD responses to the task were used toclassify them as lying or truth-telling. The authorsreported classification at 90% accuracy using acombination of analysis in the middle frontalgyrus, inferior frontal gyrus, and anterior cingu-late. This is only slightly better than the classifica-tion accuracy found for traditional polygraphy.No single region showed classification at betterthan 85% accuracy, and each cluster had a falsealarm rate of nearly 20%. The combined analysisyielded false alarms for the 10% who were notsuccessfully classified.

In a similar analysis based on a set of grouppredictor clusters, Langleben et al. (2005) re-ported single-cluster predictions at no better than75% correct classification using the area underthe curve of the receiver operator characteristicfunction, which plots probability of a false posi-tive (specificity) against probability of correctlypredicting a lie (sensitivity.) The area under thecurve was described as the probability of accu-rately classifying a pair of observations, with thebest single region (left inferior parietal lobule)showing an area under the curve of 75.1%. Usinga stepwise logistic regression model across 14ROIs, the authors reported a final model areaunder the curve of 84.7%.

These initial attempts at individual classifica-tion of lies using fMRI were comparable to theresults of the traditional polygraph: well abovechance, but well below perfection (Committee toReview the Scientific Evidence on the Polygraph,2003). Indeed, the value of the area under theROC curve reported by Langleben et al. (2005)lies in the middle of the accuracy results fromtraditional polygraphy summarized in the reportof the Committee to Review the ScientificEvidence on the Polygraph (2003).

More recently, Spence, Kaylor-Hughes, Brook,Lankappa, and Wilkinson (in press) scanned awoman who had been convicted of poisoning achild but who continued to profess her innocence.Results revealed longer response latencies and

greater activation of the VLPFC and ACC whenshe endorsed her accuser’s version of events thanwhen she endorsed her version of events; theauthors concluded that: ‘‘While we have not‘proven’ that this subject is innocent, we demon-strate that her behavioural and functional anato-mical parameters behave as if she were.’’ Such aconclusion based on an individual’s responses inan fMRI study can be questioned on logicalgrounds because lying is not the only cognitiveprocess that has been associated with increasedactivation of the VLPFC and ACC, and little isknown about how general the association isbetween lying and activation in these regions.

The purpose of the current study is to examinedata from a typical fMRI study of lying (Phan etal., 2005) to determine how well lying could bedifferentiated from truth-telling at the idiographiclevel, with an emphasis on sensitivity, specificity,and generalizability across individuals. For in-stance, because the decision threshold can affectthe sensitivity of measurement and the classifica-tion results, we used both a liberal and aconservative threshold for identifying significantdifferences in the BOLD response. In addition,prior studies examining idiographic classificationshave restricted their analyses to the predictorclusters that significantly differentiated lying fromtruth-telling in the prediction sample. Such aprocedure does not discriminate between theindividual whose scan shows only that the ROIsidentified in the predictor clusters were moreactive during lying than truth-telling*as set forthby the predictor model*and the individual whosefMRI scan shows that many if not most areas ofthe brain are more active during lying than truth-telling. Whether the inclusion of pseudo-ROIsincreases or decreases specificity has not beenexamined previously. Consequently, we also per-formed analyses to determine the effects of theapplication of an ROI control region.

METHODS

A detailed account of experimental methods canbe found in Phan et al (2005). Fourteen partici-pants were given a modified version of the GKTin which they were asked to either tell the truth orlie about the possession of a playing card (e.g., 2of hearts or 5 of clubs). Data were collectedon a 4T Siemens platform, and preprocessedusing Statistical Parametric Mapping (SPM), in-cluding slice-time correction, realignment, spatial

530 MONTELEONE ET AL.

normalization to the Montreal Neurological In-stitute (MNI) template, resampling of functionalimages to 2-mm isotropic voxels, and spatialsmoothing with a 6 mm full-width�half maximumGaussian kernel. Preprocessed data were con-verted from ANALYZE (spm99) to a format foruse with the analysis software AFNI (Cox, 1996).A canonical hemodynamic response function wasconvolved with the experimental conditions usingthe AFNI tool WAVER, and this model wasregressed against the experimental data at eachvoxel to provide within-subjects statistical mapsof the responses for each condition, as well as amap of the lie vs. truth contrast.

Nomothetic analyses

A group response map was generated by carryingout a whole-brain voxelwise one-sample t-test ofthe Lie�Truth contrast value across all 14 parti-cipants at pB.01 with a cluster size extent thresh-old of 119 contiguous voxels based on corner-to-corner connectivity. Phan et al. (2005) reportedresults only in regions where they had a priorihypotheses, namely ACC, mPFC, DLPFC, andVLPFC. In this case, a whole-brain voxelwiseresult was generated to identify all possible regionsthat differentiated lies from true statements.

Idiographic analyses

Idiographic analyses were carried out at twosignificance levels to assess the accurate detectionof lies on an individual by individual basis using aconservative (pB.01) and a liberal (pB.05) cri-terion for Type I errors. To develop the predictivemodel for the idiographic analysis, the contrastimages for 13 of the 14 participants were enteredinto the second stage of a random-effects analysis(one-sample t-test, two-tailed; t�3.057, df�12,pB.01; t�2.179, df�12, pB.05). This processwas performed for each of the 14 participants,resulting in 14 group response maps, each of whichwas then used as the template for the neuralsignature of lying when examining each individualcontrast map. That is, after group analysis from 13participants yielded predictor clusters, the remain-ing 14th participant’s data was thresholded at thesame individual voxel extent levels as the groupdata, yielding a set of target clusters for theindividual. Specifically, each remaining indivi-dual’s contrast map was submitted to the same

threshold based on the coefficient of the least-squares estimate of the empirical data to themodel (pB.01, t�2.585 and pB.05, t�1.964).Significant regions were determined by applyingan individual voxel probability threshold of pB.01 with a cluster volume of 952 ml (119 contiguousvoxels) and again with a voxelwise threshold ofpB.05 with a cluster volume of 5160 ml (645voxels) based on corner-to-corner connectivity in3D space, which was equivalent to a connectivityradius of 3.46 mm. The cluster volume was chosenas the means to correct for multiple comparisons(aB.05). Cluster volume threshold was deter-mined with a Monte Carlo simulation for whichthe input parameters modeled the analysis (voxelsize 2�2�2 mm, connectivity radius 3.46 mm,individual voxel p�.01 or pB.05, GaussianFWHM filter width�6 mm) executed within amask of the entire brain (231,766 voxels) for 5000iterations using the AFNI program AlphaSim.The Monte Carlo simulation randomly simulates‘‘active’’ voxels within the mask according theprobability and spatial parameters for the speci-fied number of iterations, ultimately calculatingthe probability that a cluster of size X would occurby chance. The volume X is then used as aselection criterion on the experimental data toobtain activity clusters that meet the corrected alevel (Forman et al., 1995; Xiong, Gao, Lancaster,and Fox, 1995).

For each participant, individual target clustermaps were overlaid on group predictor maps toassess signal detection. Significant voxels fromthe individual maps that fell within the nomo-thetic predictor map were tallied as hits if thevalences from individual results matched those ofthe group result. No predictor maps yieldedsignificant True�Lie clusters; however, sometarget maps yielded True�Lie clusters. If aTrue�Lie target cluster overlapped with aLie�True predictor cluster, it was counted as afalse alarm. If no individual results were found fora predictor cluster, it was tallied as a miss for thatcluster. If hit and false-alarm target voxels weredetected in the same predictor cluster, the clusterwas classified according to the result that had thelarger volume.

A threshold was set for the size of clusters thatoverlapped between the nomothetic predictormap and individual signal map for each subject.Overlap cluster size was determined by executinga Monte Carlo simulation within the mask of thepredictor map voxels at the appropriate voxelwisethreshold (pB.01 or pB.05) to determine the

DETECTION OF DECEPTION USING FMRI 531

cluster size corresponding to aB.05. The clustersize for overlap regions at the voxelwise pB.01level was 37 contiguous voxels, and at pB.05 itwas 203 voxels.

The pseudo-ROI specificity control

Finally, we sought to discriminate between theindividuals whose fMRI scan shows only that thepredictor clusters were more active during lyingthan truth-telling and those who showed greateractivation during lying than truth-telling acrossmany areas, among them one or more of thosespecified in the predictor clusters. To do so, weformed a set of pseudo-ROIs that served ascontrol regions for comparison with the predictorclusters. Assuming that the predictor clusters arein fact part of a reliable neural circuit, hit ratesinside predictor clusters were hypothesized to besignificantly higher than hit rates in the randomlysampled pseudo-ROIs. Given that results at theliberal threshold of pB.05 were expected tocontain a considerably higher rate of false alarmsthan at pB.01, the pB.05 results were examinedto determine whether the rate of false alarmswould be decreased.

To form the pseudo-ROIs, the predictor ROIswere subtracted from a Talairach atlas map of allcortical grey matter, and for each predictor ROI apseudo-ROI of equivalent volume was con-structed from randomly sampled voxels in theremaining regions. Pseudo-ROIs were tested forsignal detection in the same manner as grouppredictor ROIs, with a tally of significant voxelsfound at the individual level. A chi-square test(positive/negative�pseudo-ROI/actual-ROI)was used to compare frequency distributions ofhits and false alarms in the pseudo-ROIs andpredictor ROIs. Actual ROIs were expected tohave a significantly higher hit rate than pseudo-ROIs, and predictor clusters that did not meetthis criterion at pB.05 were discarded on groundsof non-specificity.

RESULTS

Nomothetic analyses

Whole-brain voxelwise group analysis showedthat group Lie�Truth contrasts revealed signifi-cant activation in the medial prefrontal cortex(mPFC) in the region of the superior medial

frontal gyrus, L inferior frontal gyrus (VLPFC),anterior cingulate, bilateral Brodmann’s areas 9and 10 (frontopolar region), bilateral temporo-parietal junction (TPJ), bilateral middle frontalgyrus (DLPFC), bilateral middle temporal gyurs,bilateral caudate, bilateral lingual gyrus, and themedial cingulate gyrus at a voxelwise pB.01 witha cluster minimum of 952 ml (119 contiguousvoxels). Not all subjects’ group analyses yieldedthe same set of regions; however, the majority ofthese regions (mPFC, VLPFC, ACC, BA 9.10,TPJ, R middle temporal, and L caudate) werefound in all 14 group template maps. Results froma representative participant are illustrated inFigure 1 and detailed in Table 1.

Idiographic analyses

Figure 2 depicts the number of participants whoshowed activation in each ROI identified usingdata from the remaining 13 participants at twolevels of detection threshold. First, more lenientdetection thresholds are generally associated withmore hits, fewer misses, and more false alarms.This was found in the present study: at pB.01,43% of clusters were hits, 57% were misses, and0% were false alarms, whereas at the morelenient detection threshold of pB.05, 53% werehits, 43% misses, and 4% of detected signals werefalse alarms. As would be expected from signaldetection theory, decreasing the decision thresh-old increased the hit rate at the expense of alsoincreasing the false alarm rate. Finally, no con-trast was found in which truthful responsesproduced greater activation than lying.

The best classification at pB.01, with a classi-fication success of 71%, was found in the mPFCin the region of the medial frontal gyrus (seeFigure 2, top panel). No other regions predictedmore than 7 of 14 participants. Four out of fivesubjects were classified in L Middle Temporalgyrus, but this region achieved significance and,thus, served as an ROI, in only 36% of thepredictor maps. No false positives were observedin any of these regions.

Lowering the detection threshold to pB.05increased the number of template maps inwhich the areas achieved significance and thepercentage of correct classifications in someregions. For instance, 78% of participants wereclassified successfully as lying in mPFC and BA9/10, and 64% in VLPFC, ACC, and Left TPJ(see Figure 2, middle panel). However, these

532 MONTELEONE ET AL.

increased levels of detection came at the cost ofincreased false positives.

The effects of applying a pseudo-ROI controlare illustrated in the bottom panel of Figure 2. Ifactivation in a predicted ROI was accompanied byactivation in the matched pseudo-ROI*that is, ifthe activation was nonspecific*then activation inthe predicted ROI was treated as nondiagnostic.The results from four participants were affectedby the implementation of pseudo-ROIs. Theregions excluded were not entirely consistent

across the participants, however. For participant#8, mPFC, BA 9/10, left TPJ, and left caudateregions were excluded, and for participant #14,right lingual gyrus was excluded. All five exclu-sions would otherwise have been considered hits.

DISCUSSION

Recent reports tout fMRI lie-detection as a meansto spot liars (Peplow, 2004), detect terrorists (Wild,

Figure 1. An illustration of the signal detection process from a representative subject. The Lie�True group predictor map (top

left) was determined using 13 subjects’ data, and was used to mask the individual Lie�True map (bottom left). Both maps were

subjected to a voxelwise threshold of pB.01 with a minimum cluster size of 119 contiguous voxels as determined by a whole-brain

Monte Carlo simulation at aB.05. The overlap map was cluster-limited by another Monte Carlo simulation carried out in the

predictor map, resulting in a cluster minimum of 37 voxels.

DETECTION OF DECEPTION USING FMRI 533

TABLE 1

Lie�Truth regions for the representative subject pictured in Figure 1: The top table shows the predictor ROIs; the bottom table shows detected overlap regions

Volume (ml) Hemisphere Region(s) TT atlas region(s) BA X Y Z Mean B-value

1 9104 R/L mPFC Superior frontal gyrus 6, 9, 10, 32 4 47 26 .904

R/L BA 9,10 Medial frontal gyrus

L Anterior cingulate Anterior cingulate

2 3544 L VLPFC Inferior frontal gyrus 47 �53 37 �18 .816

3 2552 L Temporoparietal Inferior parietal lobule 13, 40 �42 �36 26 .531

4 2120 R BA 9/10 Medial frontal gyrus 10 14 61 2 .503

R Anterior cingulate Anterior cingulate

5 2040 L DLPFC Middle frontal gyrus 6 �42 23 31 .654

6 1664 L Caudate Caudate 25, 47 �15 29 �14 .548

7 1520 R Temporoparietal Superior temporal gyrus 42 60 �33 17 .660

8 1296 R Middle temporal Middle temporal gyrus 20 53 �25 �18 .855

1 1120 R/L mPFC Medial frontal gyrus 32 2 37 32 *

2 728 L VLPFC Inferior frontal gyrus 47 �53 37 4 *

3 672 L BA 9/10 Superior frontal gyrus 9, 10 �24 68 7 *

4 480 L DLPFC Middle frontal gyrus 6 �22 27 38 *

53

4M

ON

TE

LE

ON

EE

TA

L.

Lie > True Detection by Region, P<.01

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

CFPm

GFI( CFPLV L

) etalugniC roiretnA

01 ,9 AB LB RA

01 ,9

eT Lm

teiraporop

al

ropmeT R

o

lateirap

diM L

d

latnorF el

/

FPLD

C

RMid

atnorF eld

l / D

FPLC

M Li

pmeT eldd

oarl

lddiM R

eT e

laropm aC L

uetad

etaduaC R

L

laugniL

uryG

s

L Ri

yG laugn

r susury

G etalugniC laideM

# o

f S

s S

ho

win

g R

egio

n in

Pre

dic

tor

Map

misshit

Lie < True Detection by Region, P<.05

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

m

CFP

LLVPFC

GFI()

retnA

i roC

etalugniL

AB9,

1 0 01 ,9 AB R

eT Lm

poorp

lateira

ReTmpo

rpo

ira

late

L

elddiM

orFn

at l

CFPLD /

RM

i dd

F elro

n atl /

LDP

CFlarop

meT elddiM L R

diM

dle

opmeT

arl

L

uaC

etadtaduaC R

e

LL i

ugnal

uryG

syG laugniL R

rsu

Med

laiCi gn

etalu

G yr

su

# o

f S

s S

ho

win

g R

egio

n in

Pre

dic

tor

Map

famisshit

Lie < True Detection by Region, P<.05 with Control Region Adjustment

0123456789

1011121314

CFPm

L

PLVFC

I( FG

)

Aniretor

C in

ugla

te

LAB

9,10

RBA

9,10

LTe

ropm

oap

ir eta

l

RTe

opm

orap

ir etal

Li

Mdd

leF

norta

l /

FPLD

C

Ri

Mdd

lF ero

ntal

/PLDFC

Li

Mdd

leTe

pm

arol

Ri

Mdd

lT e

meop

arl

L

uaCdat

e

R

uaCd

eta

LLi

nugal

yGru

s

RLi

nugal

yGru

s

eM

laidC

gniul

ate

uryG

s# o

f S

s S

ho

win

g R

egio

n in

Pre

dic

tor

Map

fa

miss

hit

Figure 2. (Top panel) The number of hits and misses in each ROI at a significance threshold of pB.01 with a minimum overlap

cluster size of 37 voxels. The total bar height indicates the number of subjects for which a response was predicted in each region.

Some regions (R DLPFC, L MTG, R caudate, L and R lingual gyri, and medial cingulate) were predicted for only a fraction of

subjects. No false alarms (True�Lie in a region predicting Lie�True) were detected at this significance level. Some regions (R

DLPFC, L middle temporal, R caudate, L and R lingual, and cingulate gyrus) were significantly active in only a portion of subjects’

predictor maps. (Middle panel) The number of hits and misses in each ROI at a significance threshold of pB.05 with a minimum

overlap cluster size of 203 voxels. (Bottom panel) The number of hits, misses, and false alarms in each ROI at a significance threshold

of pB.05 after the application of the pseudo-ROI control region (see text for an explanation of pseudo-ROIs).

DETECTION OF DECEPTION USING FMRI 535

2005), and reveal secret intentions in the brain(Haynes et al., 2007). Efforts to privatize newdeception detection technologies have broughtforth ethical concerns (Pearson, 2006), issues ofpersonal privacy (Olson, 2005) and legal concernsabout fourth amendment rights and personalliberty (Boire, 2005; Greely & Illes, 2007). Forensicapplications and reports of effective ‘‘brain finger-printing’’ (Farwell & Smith, 2001) and fMRI(Spence et al., in press) have been accompaniedby efforts on the part of private interests to applybrain imaging for commercial gain. The presentresults raise several questions about the claim thatthere is an invariant neural locus unique to lyingthat is sensitive, specific, and consistent acrossindividuals.

First, using a variant of the GKT, lying andtruth-telling in the current study were associatedwith differential activation in as many as 16different brain regions in some template maps.The classification of an individual as lying or notbased on ROIs specified by the template mapsproved to be well above chance, but no betterthan the levels typically observed using tradi-tional polygraphy (Committee to Review theScientific Evidence on the Polygraph, 2003). Thebest classification accuracy found for any ROIusing a stringent classification threshold was 71%(with no false positives), found using the mPFC.Using a less stringent criterion for identifyingareas of brain activation led to more hits, withactivation in two ROIs in the frontal cortexproducing correct classifications in 78% of cases,but false-alarm rates also increased across theareas that were identified.

Even though activation of the mPFC providedbetter than chance classification of individuals aslying in the present study, it does not mean thatwe detected a lie response per se. The mPFC hasbeen associated with various cognitive processesincluding self-awareness and self-referent proces-sing, mentalizing, executive functioning, and the-ory of mind (e.g., Frith & Frith, 2003; Krendl &Heatherton, in press; Saxe, Carey, & Kanwisher,2004). Accurate classifications in the currentstudy could have resulted from one or more ofthese cognitive processes occurring more often inthe lying than in the truthful response conditions.It is obvious that knowing that lying leads to askin conductance response does not logicallymean that a skin conductance response indicatesa person is lying, because lying is not the onlyprocess that produces a skin conductance re-sponse. Similarly, knowledge that lying leads to

regional brain activation (e.g., mPFC) does notimply logically that activation of that regionmarks a person who is lying.

Brain imaging is a more direct index ofcognition than the traditional polygraph, but itmay be subject to many of the same caveats. Forinstance, we found that adopting a more lenientthreshold increased hits, but at the expense ofincreasing false positives, just as has been foundin studies of traditional polygraphy. False posi-tives can be particularly pernicious in legal andemployment applications, and even low falsepositive rates can lead to large numbers oftruthful individuals being classified as lyingwhen the base rate for lying is low. For instance,if the fMRI detection procedure has an accuracyindex of .90 in a hypothetical population of 10,000examinees that includes 10 liars, a detectionthreshold set to detect 80% of the liars will miss2 of the 10 liars and will falsely classify 1,598truth-telling individuals as lying (Committee toReview the Scientific Evidence on the Polygraph,2003).

Second, our experimental paradigm made itpossible to know when participants were tellingthe truth or lying, thereby making it possible todetermine regions of brain activation that differ-entiated lying and truth-telling (i.e., a template).Typically, it is not possible to know whether aparticipant is telling the truth or lying. Theintersubject variability in brain response to lyingobserved in the current study means that thetemplate identified using a nomothetic approachmay fail to characterize accurately many indivi-duals’ brain response to lying. This was evident inthe present study, where nomothetic maps per-mitted lies to be accurately classified in the caseof most but not all individuals. Templates de-signed at the idiographic rather than the nomo-thetic level may be worth investigating todiscriminate lying from truth-telling, but themerit of such an approach has yet to be proven.To develop such a template, for instance, indivi-duals would need to lie and to tell the truth inknown sequences in response to multiple ques-tions. This procedure may be limited if partici-pants are not cooperative or if different kinds oflies are associated with different patterns of brainactivation.

Third, the empirical support to date has raisedquestions about the notion that there is aninvariant neural signature for lying. The cardversion of the GKT is a simple task that maylimit individual variability in type of lie that is

536 MONTELEONE ET AL.

emitted. Although a card-based guilty knowledgetest serves as a model for lying, it may be anoverly simplified paradigm for testing legitimatelie-detection, and does not address the issue ofdifferences in response according to differenttypes of lie or the type of information involved.Thus, any level of success of lie classificationbased on this paradigm is but an early steptowards more ecologically valid scenarios. In-deed, the research by Ganis et al. (2003) suggeststhat different types of lies may be associated withcompletely different patterns of brain activation.This seems plausible, as different types of lies arethe product of one or more distinguishablecomponent processes, which should be reflectedin differences in neural activation.

Fourth, the brain imaging research on lying todate has not considered what detection problemsmight be introduced by participants who usephysical or mental countermeasures. If a partici-pant were to perform serial subtraction or tofocus on suppressing an irrelevant response whenresponding truthfully and lying, for instance,would the pattern of neural activation thatdifferentiates lying from truth-telling still beevident? Or, in light of the nomothetic templatefound in the present study, would thinking aboutthe mental state of the examiner and oneself asthe object of interrogation when respondingtruthfully act as a cognitive countermeasure tothe detection of deception? Such questions will beimportant to address before the applied value oflie detection using fMRI can be determined.

Finally, social, cultural, and linguistic differ-ences across participants in the understanding ofquestions and the meaning and appropriatenessof deception may contribute to differences inregional activation patterns observed during lyingversus truth-telling. In the absence of an invariantneural signature for lying, successful detection ofdeception using fMRI may require a betterappreciation of possible sociocultural effects.

In sum, fMRI as a method for the detection ofdeception should carry the same burden of proofas demanded of any other method. As in priorstudies, fMRI analyses permitted the differentia-tion of deceptive and truthful responding at thegroup level. The classification of lying based oncontrast maps, however, was better than chancebut far from perfect. Two different reasons formisclassifications were identified. In some cases,the activation of a predicted ROI did not reflectspecific activation in this region but insteadreflected activation of a very large area that

included part of the predicted ROI. The use ofpseudo-ROIs proved useful in identifying suchcases. The second and more common reason formisclassifications was that one or more of theregions found to differentiate lying from truthfulresponding in the template map were not asso-ciated with significant differences in activation inthe test participant. Using more lenient thresh-olds for detecting activation only nominally im-proved classification accuracy and came at theexpense of increased false alarm rates. Closerinspection of these individuals’ contrast maps (lie� true) suggested that a few subjects showed lyingassociated with very different regions of activa-tion than most individuals. Although this isspeculative, lying for these individuals may havebeen achieved by equally atypical means. To-gether, these results suggest that, although fMRImay permit investigation of the neural correlatesof lying, at the moment it does not appear toprovide a very accurate marker of lying that canbe generalized across individuals or even perhapsacross types of lies by the same individuals.

REFERENCES

Alder, K. (2007). The lie detectors: The history of anAmerican obsession. New York: The Free Press.

Blasi, G., Goldberg, T., Weickert, T., Das, S., Kohn, P.,Zoltick, B., Bertolino, A., Callicott, J., Weinberger,D. R., & Mattay, V. S. (2006). Brain regions under-lying response inhibition and interference monitor-ing and suppression. European Journal ofNeuroscience, 23(6), 1658�1664.

Blumenfeld, R. S., & Ranganath, C. (2006). Dorsolat-eral prefrontal cortex promotes long-term memoryformation through its role in working memoryorganization. Journal of Neuroscience, 26(3), 916�925.

Boire, R. G. (2005). Searching the brain: the fourthamendment implications of brain-based deceptiondetection devices. American Journal of Bioethics,5(2), 62�63.

Committee to Review the Scientific Evidence on thePolygraph (2003). The polygraph and lie detection.Washington, DC: National Academy Press.

Cox, R. W. (1996). AFNI: Software for analysis andvisualization of functional magnetic resonance neu-roimages. Computers and Biomedical Research, 29,162�173.

Farwell, L. A., & Smith, S. S. (2001). Using brainMERMER testing to detect knowledge despiteefforts to conceal. Journal of Forensic Sciences,46(1), 135�143.

Forman, S. D, Cohen, J. D., Fitzgerald, M., Eddy, W. F.,Mintun, M. A., & Noll, D. C. (1995). Improvedassessment of significant activation in functionalmagnetic resonance imaging (fMRI): Use of a

DETECTION OF DECEPTION USING FMRI 537

cluster-size threshold. Magnetic Resonance Medi-cine, 33, 636�647.

Frith, D. D., & Frith, U. (2003). Interacting minds: Abiological basis. Science, 286, 1692�1695.

Ganis, G., Kosslyn, S. M., Stose, S., Thompson, W. L., &Yurgelun-Todd, D. A. (2003). Neural correlates ofdifferent types of deception: an fMRI investigation.Cerebral Cortex, 13, 830�836.

Greely, H. T., & Illes, J. (2007). Neuroscience-based liedetection: the urgent need for regulation. AmericanJournal of Law & Medicine, 33, 377�431.

Haynes, J. D., Sakai, K., Rees, G., Gilbert, S., Frith, C.,& Passingham, D. (2007). Reading hidden intentionsin the human brain. Current Biology, 17, 323�328.

Kozel, F. A., Johnson, K. A., Mu, Q., Grenesko, E. L.,Laken, S. J., & George, M. S. (2005). Detectingdeception using functional magnetic resonanceimaging. Biological Psychiatry, 58(8), 605�613.

Kozel, F. A., Padgett, T. M., & George, M. S. (2004a). Areplication study of the neural correlates of decep-tion. Behavioral Neuroscience, 118, 852�856.

Kozel, F. A., Revell, L. J., Lorberbaum, J. P., Shastri, A.,Elhai, J. D., Horner, M. D., Smith, A., Nahas, Z.,Bohning, D. E., & George, M. S. (2004b). A pilotstudy of functional magnetic resonance imagingbrain correlates of deception in healthy youngmen. Journal of Neuropsychiatry and Clinical Neu-rosciences, 16, 295�305.

Krendl, A. C., & Heatherton, T. F. (in press). Compo-nents of the social brain. In G. G. Berntson & J. T.Cacioppo (Eds.), Handbook of neuroscience for thebehavioral sciences. New York: Wiley.

Langleben, D. D., Loughead, J. W., Bilker, W. B.,Ruparel, K., Choldress, A. R., Busch, S. I., & Gur,R. C. (2005). Telling truth from lie in individualsubjects with fast event-related fMRI. Human BrainMapping, 26, 262�272.

Langleben, D. D., Schroeder, L., Maldjian, J. A., Gur,R. C., McDonald, S., Ragland, J. D., O’Brien, C. P.,& Childress, A. R. (2002). Brain activity duringsimulated deception: An event-related functionalmagnetic resonance study. NeuroImage, 15, 727�732.

Lee, T. M., Liu, H. L., Tan, L. H., Chan, C. C. H.,Mahankali, S., Feng, C. M., Hou, J., Fox, P. T., &Gao, J. H. (2002). Lie detection by functionalmagnetic resonance imaging. Human Brain Map-ping, 15, 157�164.

Nunez, J. M., Casey, B. J., Egner, T., Hare, T., & Hirsch,J. (2005). Intentional false responding shares neural

substrates with response conflict and cognitivecontrol. NeuroImage, 25, 267�277.

Olson, S. (2005). Brain scans raise privacy concerns.Science, 307, 1548�1550.

Pearson, H. (2006). Lure of lie detectors spooksethicists. Nature, 441, 918�919.

Peplow, M. (2004). Brain imaging could spot liars.Nature News. Retrieved January 15, 2008 fromhttp://www.bioedonline.org/news/news.cfm?art�1409

Phan, K. L., Magalhaes, A., Ziemlewicz, T. J., Fitzger-ald, D. A., Green, C., & Smith, W. (2005). Neuralcorrelates of telling lies: a functional magneticresonance imaging study at 4 tesla. AcademicRadiology, 12, 164�172.

Rahm, B., Opwis, K., Kaller, C. P., Spreer, J., Schwarz-wald, R., Seifritz, E., Halsband, U., & Unterrainer,J.M. (2006). Tracking the subprocesses of decision-based action in the human frontal lobes. Neuro-Image, 30(2), 656�667.

Saxe, R., Carey, S., & Kanwisher, N. (2004). Under-standing other minds: linking developmental psy-chology and functional neuroimaging. AnnualReview of Psychology, 55, 87�124.

Spence, S.A., Farrow, T.F., Herford, A.E., Wilkinson,I.D., Zheng, Y., & Woodruff, P.W. (2001). Behavioraland functional anatomical correlates of deception inhumans. NeuroReport, 12, 2849�2853.

Spence, S. A., Kaylor-Hughes, C. J., Brook, M. L.,Lankappa, S. T., & Wilkinson, I. D. (in press).‘Munchausen’s syndrome by proxy’ or a ‘miscar-riage of justice’? An initial application of functionalneuroimaging to the question of guilt or innocence.European Psychiatry.

Wild, J. (2005). Brain imaging ready to detect terrorists,say neuroscientists. Nature, 437, 457.

Wirsing, B. (2007). Revealing secret intentions in thebrain (Press Release). Munich: Max Planck Societyfor the Advancement of Science.

Wolpe, P. R., Foster, K. R., & Langleben, D. D. (2005).Emerging neurotechnologies for lie detection:Promises and perils. The American Journal ofBioethics, 5(2), 39�49.

Xiong, J., Gao, J.-H., Lancaster, J. L., & Fox, P. (1995).Clustered pixels analysis for functional MRI activa-tion studies of the human brain. Human BrainMapping, 3, 287�301.

538 MONTELEONE ET AL.