Columbia Law School Scholarship Archive

Faculty Scholarship Faculty Publications

2016 fMRI and Lie Detection

Anthony D. Wagner [email protected]

Richard J. Bonnie [email protected]

BJ Casey [email protected]

Andre Davis [email protected]

David L. Faigman

See next page for additional authors

Follow this and additional works at: https://scholarship.law.columbia.edu/faculty_scholarship

Part of the Criminal Law Commons, and the Science and Technology Law Commons

Recommended Citation Anthony D. Wagner, Richard J. Bonnie, BJ Casey, Andre Davis, David L. Faigman, Morris B. Hoffman, Owen D. Jones, Read Montague, Stephen J. Morse, Marcus E. Raichle, Jennifer A. Richeson, Elizabeth S. Scott, Laurence Steinberg, Kim Taylor-Thompson & Gideon Yaffe, fMRI and Lie Detection, MACARTHUR FOUNDATION RESEARCH NETWORK ON LAW & NEUROSCIENCE, FEBRUARY 2016; VANDERBILT LAW RESEARCH PAPER NO. 17-10 (2016). Available at: https://scholarship.law.columbia.edu/faculty_scholarship/2015

This Working Paper is brought to you for free and open access by the Faculty Publications at Scholarship Archive. It has been accepted for inclusion in Faculty Scholarship by an authorized administrator of Scholarship Archive. For more information, please contact [email protected]. Authors Anthony D. Wagner, Richard J. Bonnie, BJ Casey, Andre Davis, David L. Faigman, Morris B. Hoffman, Owen D. Jones, Read Montague, Stephen J. Morse, Marcus E. Raichle, Jennifer A. Richeson, Elizabeth S. Scott, Laurence Steinberg, Kim Taylor-Thompson, and Gideon Yaffe

This working paper is available at Scholarship Archive: https://scholarship.law.columbia.edu/faculty_scholarship/ 2015

Attached hereto:

fMRI and Lie Detection A Knowledge Brief of the MacArthur Foundation Research Network on Law and Neuroscience

Cite as:

Anthony Wagner, Richard J. Bonnie, BJ Casey, Andre Davis, David L. Faigman, Morris B. Hoffman, Owen D. Jones, Read Montague, Stephen J. Morse, Marcus E. Raichle, Jennifer A. Richeson, Elizabeth S. Scott, Laurence Steinberg, Kim Taylor-Thompson, Gideon Yaffe, fMRI and Lie Detection: A Knowledge Brief of the MacArthur Foundation Research Network on Law and Neuroscience (2016).

Owen D. Jones Director, MacArthur Foundation Research Network on Law and Neuroscience New York Alumni Chancellor's Chair in Law & Professor of Biological Sciences, Vanderbilt University

Richard J. Bonnie Stephen J. Morse Harrison Foundation Professor of Medicine and Law Ferdinand Wakeman Hubbell Professor of Law Professor of Psychiatry and Neurobehavioral Sciences Professor of Psychology and Law in Psychiatry Director, Institute of Law, Psychiatry and Public Policy University of Pennsylvania Professor of Public Policy Frank Batten School of Leadership and Public Policy Marcus E. Raichle University of Virginia Professor of Radiology, Neurology, Neurobiology, Biomedical Engineering, and Psychology, Washington University

BJ Casey Jennifer A. Richeson Professor of Psychology, Yale University Philip R. Allen Professor of Psychology, Yale University Adjunct Professor, Weill Cornell Medical College, New York City Director, Social Perception and Communication Lab Director, Fundamentals of the Adolescent Brain Lab Elizabeth S Scott Andre Davis Harold R. Medina Professor of Law, Columbia University Judge, United States Court of Appeals for the Fourth Circuit Laurence Steinberg David L. Faigman Distinguished University Professor John F. Digardi Distinguished Professor of Law Laura H. Carnell Professor of Psychology, Temple University Director of the UCSF/UC Hastings Consortium on Law Science & Health Policy Kim A. Taylor-Thompson University of California, Hastings Professor of Clinical Law, New York University

Morris B. Hoffman Anthony Wagner District Judge, Second Judicial District, State of Colorado Professor of Psychology Director, Stanford Memory Lab Read Montague Associate Director of the Cognitive & Neurobiological Imaging Center Professor of Physics Stanford University Professor of Psychiatry and Behavioral Medicine Director, Human Neuroimaging Laboratory Gideon Yaffe Director, Computational Psychiatry Unit, Professor of Law Virginia Tech & University College, London Professor of Philosophy & Psychology, Yale University

The MacArthur Foundation Research Network on Law and Neuroscience Vanderbilt Law School, 131 21st Avenue South, Nashville, TN 37203

Electronic copy available at: https://ssrn.com/abstract=2881586 Law + fMRI AND Neuro LIE DETECTION THE MACARTHUR FOUNDATION RESEARCH NETWORK ON LAW AND NEUROSCIENCE

In September 2012, the Sixth Circuit Court of Appeals, citing Federal Rule of Evidence 702 and Rule 403, agreed with the trial court’s exclusion of fMRI-based lie detection evidence in the fraud case of United States v Semrau.

A scant month earlier, Judge Eric M. Johnson of the Maryland Sixth Judicial Circuit, Montgomery County, had refused to admit potentially exculpatory fMRI lie detection evidence in the murder trial of State v Gary Smith. Citing the Frye standard, Johnson wrote, “It is clear to the Court that the use of fMRI to detect deception and verify truth in an individual’s brain has not achieved general acceptance in the scientific community.”

While research on fMRI-based lie detection has continued, the general consensus in the scientific community regarding its probative value remains the same. This brief explores why.

WHAT IS fMRI? The typical experimental paradigm involves “instructed” lies: a subject is given detailed instructions about how For more than a decade now, scientists have been explor- and when to lie, then placed in a scanner. Does conscien- ing the potential of functional magnetic resonance imaging, tiously following those instructions constitute lying? Many or fMRI, to assess increased activity in brain regions asso- researchers worry that the answer is no, rendering the ciated with the cognitive processes required for lying. experimental results irrelevant. fMRI does not measure neural activity directly. Instead, A distinct but related question arises from the poorly it measures small and variable changes in the ratio of defined nature of the real-world lie. Two equally false oxygenated to deoxygenated blood in the brain when a statements—“Of course I remember you” and “No, I particular task is performed or stimulus presented—the didn’t kill him”—may be as distinct neurally as they are so-called BOLD, or blood oxygen level-dependent, re- morally. Similarly, an often-repeated lie or one first told sponse. Firing neurons, like working muscles, require many years ago might look markedly different from an oxygen; follow the trail of oxygenated hemoglobin, and unpracticed or recent lie. you find neural activity.

A statement based on faulty memory (“I never said that”) LIES, DAMNED LIES, AND BEING COOPERATIVE may not trigger any neural activity associated with decep- The most fundamental question scientists raise when tion at all. There is some evidence to suggest that fMRI reviewing fMRI lie detection research is this: Do these scanning will detect the subject’s belief, even if that belief experiments actually examine lies?

Electronic copy available The MacArthur at: httpsFoundation://ssrn.com Research Network/abstract on Law and=2881586 Neuroscience / fMRI AND LIE DETECTION / 02.23.16 1 PERSPECTIVES

Lateral task, some of the participants committed isn’t borne out by the objective truth. In a 2010 memory MFG IFG–insula the mock crime and others did not, but all experiment supported by the Research Network and con- IPL were instructed to indicate that they did not. The authors were able to correctly detect ducted by neuroscientist Jesse Rissman and colleagues, deception, using fMRI, in 100% of the par- ticipants in the ‘mock crime present’ condi- the brain activity observed when subjects recognized a tion. However, they also mistakenly detected deception in 67% of the participants in the face was comparable to that observed when subjects ‘mock crime absent’ condition. In the lan- believed they had seen a face before but hadn’t. guage of diagnostic testing, the sensitivity of m/SFG this test was high but the specificity was low. Medial Determining the real-world accuracy of a detection test also depends on a critical third factor, which is the probability of the PROBLEMS OF INFERENCE event occurring within the population — the base rate. Indeed, the risks associated It is impossible to infer a specific mental process solely on with a lie-detection test with low specificity the basis of brain activity in a particular region, or even in a will depend on the base rate of lying in the population assessed28,29. Imagine that the test particular set of brain regions. A single brain region is often described in REF. 22 was given to 101 people, 100 of them truthful and 1 deceptive. On involved in a number of mental processes, and a mental FigureFIGURE 1 | Results 1. Results of the ALE of the analysis ALE ofanalysis the functional of the MRIfunctional ‘deception’ MRI literature. “deception” Overlay literature of map the basis of the false-positive rates of Kozel of activation likelihood estimation (ALE) values (orange) on the lateral (top) and medial (bottom) and colleagues22, the test would identify 68 inflatedrevealing PALS regionssurface76 , consistentlyrevealing regions implicated that are consistently in deception implicated acrossNature in deceptionstudies. Reviews The |across Neuroscience meta- studies. process often involves multiple areas of the brain. participants as ‘lying’ — the 1 participant Theanalysis detection was thresholded performed was over set at 321 P < 0.05, foci and from the 28 false-discovery independent rate statistical was corrected contrasts as per the who lied about the mock crime and 67 who methodbetween described lie and in REF. truth13. IFG,conditions inferior frontal reported gyrus; in IPL, 23 inferior different parietal studies. lobule; As MFG, noted middle by frontal others, gyrus; m/SFG, medial and superior frontal gyrus. did not. In other words, given a positive In 2014, neuroscientist Martha J. Farah and colleagues no region was active in all, or nearly all, studies. result, the probability of the test accurately indicating someone as lying is 1 in 68, or less published a meta-analysis of the fMRI-based lie detection (to our knowledge) report data relevant whether the neural correlates of deception are than 1.5%, and the likelihood of incorrectly to detecting deception at the individual- sought using univariate or multivariate analy- indicating deception when it is not present is literature to date. Like the meta-analysis performed by event level. Specifically, Langleben et al.16 ses and regardless of whether the correlates over 98%. As this example illustrates, even in subjects were17 actually lying about the stimulus at the time. neuroscientist Shawn Christ and colleagues in 2009, the and Davatzikos et al. focused on whether are discovered by simple regression analysis or an ideal circumstance in which a laboratory Wasinstructed the lie and brain truth eventsactivation could be a resultmachine-learning of deception, algorithms. Furthermore, then, or lie-detection test is developed and used in study reveals both variability in the particular brain regions discriminated in the same dataset, using even high accuracy rates may decline precipi- identical situations, and is sensitive to decep- attention,either logistic regression that 16is, or non-linearthe salience tously of when the subjects stimulus? use countermeasures This study in tion within an individual 100% of the time, activated across experiments and some notable consisten- machine-learning analyses17. Although these an attempt to conceal their ‘deception’ (REF. 2). its accuracy in a larger population may still callsevent-level into analyses question showed accuracy many rates prior publishedThe laboratory studies reports assessing that the accu used- be unacceptably low if the specificity of the cy. The regions that consistently showed deception-related of 78% and 88%, respectively, the impact of racy of fMRI-based lie detection on the indi- test is low and the base rate of lying is low. athese similar early findings paradigm, is limited because as theof the brainvidual-subject activity level in assess those the sensitivity studies and The real-world validity of fMRI-based lie activity were the ventrolateral and dorsolateral prefrontal above-noted response-frequency confound specificity within an individual by differenti- detection will also depend on the generaliz- mayin the experimental design. not reflect neural responsesating trials to on whichdeception. the individual is decep- ability of the findings obtained in laboratory cortex, inferior parietal lobe, anterior insula, and medial Other studies focused on examining tive or truthful. However, determining the studies (which typically involve undergradu- superior frontal cortex. Predictably, those regions are whether fMRI BOLD activity differs when accuracy of a test in a general population also ate students — that is, healthy, educated Neuroscientistan individual is lying compared F. Andrew to telling the Kozelrequires and an assessment colleagues of the test’s analyzed sensitivity young adult subjects) to the individuals activated during other cognitive processes, as well, in truth (pooling data across events). Various and specificity across individuals within that whose veracity is to be assessed by these datastatistical from approaches three have been independent implemented, population: “mock that theft” is, what is experiments the likelihood of methods. Consider the differences between particular, those processes that form part of what we call including single-subject univariate analy- detecting deception when it is present in a criminal offenders, a group that is likely to inses 2,which7,19,20, univariate subjects analyses combined were with instructed member of to the lookpopulation at (sensitivity),two objects, and be subjected to lie-detection methods, and “executive control,” e.g., planning, working memory (the the counting of above-threshold voxels in what is the likelihood of correctly indicating the undergraduate students on which these select one, take8,22 it,25 from a drawer, hide it in a locker con- system that provides for temporary storage and manip- targeted regions of interest and machine- when deception is absent (specificity)? To methods have so far been tested. A relatively taininglearning classification the subject’s2. The reported personal accura- date, belongings, fMRI-based lie-detection and thentests examin deny- high proportion of criminals meet the crite- ulation of information), inhibitory control (the ability to cies in these individual-subject-level analyses ing accuracy within individuals have gener- ria for psychopathy, a condition that is asso- havinghave ranged takenfrom 69% eitherto 100%, suggesting object. Accuracyally not assessed rates the sensitivity for those or specificity mock ciated with frequent acts of deception and suppress actions and resist interference from irrelevant that these statistical approaches have prom- of the test across individuals. One exception with alterations in both structural MRI and theftise. However, experiments here again, the noted range concerns from is a71 study to by Kozel90 percent.and colleagues 22But, which when fMRI studies27. A study of fMRI-based lie stimuli), and attention. Even the instructed lie is cogni- about attention and memory confounds tested participants who had been success- detection in criminal offenders with a diag- subjects have more7,12,26 and richer memories of one object tively complex: a subject must remember a set of circum- undermine data interpretation . Indeed, so fully classified using the researchers’ method nosis related to psychopathy — specifically, thanlong as studiesanother, use tasks howin which much the effects of what’sas ‘lying’ on being a prior mock detected crime task (25 is antisocial personality disorder — found that stances, attend to stimuli that vary in their significance of deception cannot be separated from the out of 36 participants). These pre-selected a large proportion of these participants did deceptioneffects of attention and and memory, how the much prob- memory?participants were A subsequentthen examined on a sec2012- not show typical prefrontal BOLD response or salience, decide to lie, suppress the truth, and choose lem of confounds will remain, regardless of ondary mock crime task. On this secondary patterns during instructed deception29. study by Mathias Gamer and colleagues suggests that among relevant and plausible details. memory126 | FEBRUARY may 2014 | VOLUME be a 15critical confound in many prior studies. www.nature.com/reviews/neuro © 2014 Macmillan Publishers Limited. All rights reserved CONFOUNDS: Variables that can prejudice results aren’t limited to those DO WE KNOW WHAT WE’RE MEASURING? inadvertently introduced in research studies. Blood flow Even if instructed lies are lies, and there is some com- itself is influenced by a variety of factors independen mon physiological ground shared by all lies, experimental of neural activity, including age, vascular capacity, and confounds in most of the studies to date make it impos- medication. The fMRI results offered in the Semrau case sible for researchers to know whether the neural activity included a confound likely to be unavoidable in civil or measured is associated with lying or with something else. criminal applications of the technology: the length of time between the fMRI and the event in question. Relatively A 2008 experiment by neuroscientist Jonathan Hakun and little research has been done on how such variables as colleagues, for example, included the following finding: subject fatigue, anxiety, fear, the presence of a perceived Brain activation was observed whenever the target or threat, or practice affect fMRI results. “lie” stimulus was presented, independent of whether the

The MacArthur Foundation Research Network on Law and Neuroscience / fMRI AND LIE DETECTION / 02.23.16 2 The uninstructed lie

In contrast to nearly all other studies to date, It was a cover story that both encouraged percentage of wins in the opportunity trials. one fMRI data set shows brain activity during participant honesty (to test the hypothesis Subsequent fMRI data analysis revealed that genuine dishonesty—that is, dishonesty adequately required them to tell the truth) increased activity in the prefrontal cortex— related to a freely exercised choice to lie. It and gave them the opportunity to lie (in anterior cingulate and dorsolateral and ven- was the result of an ingenious experiment some trials, they believed they would be trolateral prefrontal cortices—was associated published in 2009 by neuroscientists Joshua self-reporting their success at prediction). In with the decision to lie in the dishonest Greene and Joseph Paxton. reality, the study was an attempt to deter- group. Interestingly, an even greater increase mine what makes people behave honestly in prefrontal cortex activity in this group was The pair asked participants to predict the when they are confronted with an opportuni- observed in connection with the decision to outcome of random computerized coin flips ty for dishonest gain. refrain from lying. In other words, when while undergoing fMRI. The experiment was individuals who had shown themselves presented as an inquiry into paranormal Throughout a series of “opportunity” (the willing to lie passed up the opportunity and ability to predict the future; the supposed “opportunity” being to lie) and “no instead reported a loss, prefrontal cortical hypothesis was that predictive ability opportunity” trials, participants made their activity was even higher than when they improved when predictions were not made predictions, believing them to be either lied. (Whether this increase is due to public in advance and were associated private or public, depending on the trial. considering deception, resisting temptation, with financial gain or loss. Researchers then classified the participants or something else is currently unknown.) In as honest, dishonest, or ambiguous based the honest group, no significant effects were on the probability of their self-reported observed when choosing not to lie.

COUNTERMEASURES: magnitude of hippocampal activity—a region long known to be important for memory—and distributed neural DETECTING A LIAR AND A CHEAT patterns could be manipulated by retrieval strategies. Another serious obstacle to using fMRI for lie detection in the real world is that very little research exists on POTENTIAL PROBLEMS OF VALIDITY countermeasures, actions taken to make test results Many scientists argue that the conclusions drawn from misleading or unusable. Moving during scanning or not fMRI ”lie detection” experiments conducted to date are following instructions can ruin a test, but they’re also likely only valid within the context of the experimental data. to be spotted. More worrisome is whether unnoticeable physical or mental strategies could nonetheless effectively Group data might not be able to tell us what we interfere with patterns of neural activity or signal strength. need to know about an individual. The holy grail of lie detection is to distinguish truth from lie reliably at One study that looked at countermeasures with respect the level of the individual subject and at the level of the to fMRI lie detection, conducted by Giorgio Ganis and individual question. But most of the studies conducted colleagues in 2011, featured prominently in the Gary Smith on deception to date focus on truth vs. lie differences murder trial. Study participants were instructed to use a averaged over multiple subjects and trials. series of covert actions, such as imperceptibly moving a left index finger or left toe, just before pressing the A sufficient amount of group-averaged data can indicate response button each time they saw irrelevant dates in a that a certain pattern of neural activity is frequently series. In trials without the countermeasure, researchers associated with a particular experimental condition. were able to detect deception with 100 percent accuracy. However, they cannot tell us whether the pattern of When the countermeasure was employed, detection activation is not also common to other experimental accuracy fell to just 33 percent. conditions (or mental processes). Nor, for the moment, can they shed much light on whether fMRI can reli- A 2015 study by Melina Uncapher and colleagues, this one ably detect lies at the level of the individual subject of memory, showed that participants could successfully or question. In his testimony during the Semrau trial, conceal or feign memory for faces. Interestingly, and some Cephos Corporation CEO Steven Laken, who conduct- might say discouragingly, the study showed that both the ed the fMRI lie detection tests submitted as evidence,

The MacArthur Foundation Research Network on Law and Neuroscience / fMRI AND LIE DETECTION / 02.23.16 3 confirmed that they did not indicate whether Dr. Semrau PRINCIPLED OBJECTIONS responded truthfully as to any specific question and that At present, many of the issues that concern the scientific it was “certainly possible” that Dr. Semrau was lying on community with respect to the use of fMRI for lie detec- some of the particularly significant questions. tion are likely to be problematic for the legal community, at least in most contexts. In fact, much of the existing Experimental conditions often poorly approximate research on deception has no bearing on the question that the real world. To date, fMRI studies have focused matters most to judges, lawyers, defendants, and juries, on detecting lies about an event that just occurred. i.e., “Can fMRI-based lie detection methods provide a The event often has no personal relevance and no legally relevant answer to a specific question?” consequences. Real-world fMRI lie detection focuses on events or facts that are likely to have occurred Most scientists—including many who have reported months or even years before, are deeply relevant to detecting lies in the laboratory with a high degree of the subject, and have serious consequences. Little is accuracy—agree that more and different research will known about whether real-world and experimental need to be conducted before fMRI-based lie detection conditions yield similar results. is ready for its day in court.

The sensitivity and specificity of fMRI lie detection have not been established. No diagnostic tool is TO LEARN MORE perfectly accurate. Antiviral software sometimes Brain Imaging for Legal Thinkers: A Guide for the Perplexed detects threats that aren’t there; mammograms miss tumors. The probative value of fMRI-based evidence Through a Scanner Darkly: Functional Neuroimaging as Evidence of a Criminal Defendant’s Past Mental States depends on knowing how many lies the tool misses and how often it identifies the truth as a lie; few Detecting Individual Memories through the Neural Decoding of Memory States and Past Experience research studies to date have reported such data. Brain Scans as Evidence: Truths, Proofs, Lies, and Lessons

Findings may not be generalizable to other Law and Neuroscience in the United States populations. fMRI studies typically are conducted on To download these and other publications exploring the intersection of undergraduates and other healthy younger adults. law and neuroscience, please visit the MacArthur Research Even if we know that there is neural activity in particular Network on Law and Neuroscience, at www.lawneuro.org. regions under the condition of lying when subjects are This brief is produced by the MacArthur Research Network on Law and younger and healthy—a matter of debate, as already Neuroscience. Supported by the John D. and Catherine T. MacArthur discussed—do we know anything at all about what to Foundation, the network addresses a focused set of closely-related problems at the intersection of neuroscience and criminal justice: expect from a woman of 70, or someone with a 1) investigating law-relevant mental states of, and decision-making mental illness? processes in, defendants, witnesses, jurors, and judges; 2) investigating in adolescents the relationship between brain development and cognitive capacities; and 3) assessing how best to draw inferences about individuals from group-based neuroscientific data.

REFERENCES

Christ, S. E., Van Essen, D. C., Watson, J. M., Brubaker, Ganis, G., Rosenfeld, J. P., Meixner, J., Kievit, R. A., Kozel, F. A., Johnson, K. A., Mu, Q., Grenesko, E. L., L. E., & McDermott, K. B. (2009) The contributions of pre- Schendan, H. E. (2011) Lying in the scanner: Covert Laken, S. J., & George, M.S. (2005) Detecting deception frontal cortex and executive control to deception: Evidence countermeasures disrupt deception detection by functional using functional magnetic resonance imaging. Biological from activation likelihood estimate meta-analyses. Cerebral magnetic resonance imaging. Neuroimage, 55, 312–19. Psychiatry, 58:8, 605–613. Cortex, 19, 1557–66. Greely, H. T., and Illes, J. (2007) Neuroscience-Based Lie Rissman, J., Greely, H. T., & Wagner, A. D. (2010) Detecting Farah, M. J., Hutchinson, J. B., Phelps, E. A., & Wagner, Detection: The Urgent Need for Regulation. American individual memories through the neural decoding of A. D. (2014) Functional MRI-based lie detection: Scientific Journal of Law and Medicine, 33, 377–431. memory states and past experience. Proceedings of the and societal challenges. Nature Reviews Neuroscience, 15, National Academy of Sciences, 107:21, 9849–9854. 123-31. Greene, J. D., and Paxton, J. M. (2009) Patterns of neural activity associated with honest and dishonest moral Uncapher, M. R., Boyd-Meredith, J. T., Chow, T. E., Gamer, M., Klimecki, O., Bauermann, T., Stoeter, P., & decisions. Proceedings of the National Academy of Rissman, J., & Wagner, A. D. (2015) Goal-directed Vossel, G. (2012) fMRI-activation patterns in the detection Sciences, 106:30, 12506–12511. modulation of neural memory patterns: Implications for of concealed information rely on memory-related effects. fMRI-based memory detection. Journal of Neuroscience, Social Cognitive and Affective Neuroscience, 7(5), 506–15. Hakun, J. G., Seelig, D., Ruparel, K., Loughead, J. W., 35(22), 8531-45. Busch, E., Gur, R. C., & Langleben, D. D. (2008) fMRI investigation of the cognitive structure of the concealed information test. Neurocase, 14, 59–67.

The MacArthur Foundation Research Network on Law and Neuroscience / fMRI AND LIE DETECTION / 02.23.16 4