A "lie detector" test -- and how to cheat it

ResearchBlogging.orgImplicit attitudes and associations can tell us a lot about a person. It's a way to find out if they might have a racial or gender bias, and recently there has even been some work suggesting that an implicit association test can tell us whether someone is lying -- it's called the autobiographical Implicit Association Test (aIAT). Here's how it works:

Suppose you're an American soldier and you want to know if an Afghan you've captured is really a member of the Taliban, despite his assurances that he supports your side and only carries a Kalashnikov for self-defense. You fire up the aIAT software on your laptop and give him a quick test:

First, he's shown true and false statements unrelated to his allegiance in the war: "I am at a computer;" "I am at the tea house." The task is to press a button identifying the statement as quickly as possible: A for true or L for false. Next he reads statements indicating guilt or innocence: "I am a member of the Taliban," "I support the American troops," and again asked to press one key for "guilty" statements and one key for "innocent" statements. Finally the two tasks are mixed: Press A for a true or guilty statement; press L for a false or innocent. Then the associations are reversed: A for true or innocent; L for false or guilty. Reaction times are measured. If he's truly guilty, then he should react faster to "guilty" statements paired with "true." If he's innocent, then he should react faster to "innocent" statements paired with "true."

But might it be possible to outwit the test?

A team led by Bruno Verschuere administered a similar test to 36 undergraduate students. Half of the students were instructed to act out a "crime" before taking the test: go to a professor's office and steal a CD containing the answers to the final exam. Half the students were "innocent" and read a newspaper account of the "crime." Then all the students were given the aIAT. As expected, the guilty students failed the test, and the innocent students passed.

But next they were all instructed on how to fake a "not guilty" result on the test: Deliberately slow down when the statements expressing innocence were paired with true statements (and statements demonstrating guilt were paired with false statements). Here are the results:


As you can see, in the original (unfaked) test, students who really did steal the CD had positive scores, indicating guilt, while students who didn't steal the CD had negative scores, indicating innocence. But after they were instructed on how to fake, the results of the guilty students were statistically indistinguishable from those of innocent students -- they had successfully cheated the test, after only a moments' worth of instruction on how to do it.

In a separate experiment, the researchers found the same result even when students took the guilt/innocence test only once. And in a third study, even when the amount of time to respond was limited (potentially preventing people from deliberately slowing down), the students were still able to fake their results.

So it appears that while the IAT can tell us a lot about individuals, it's quite easy for a savvy malfeasant to cheat the test when it's used as a lie detector.

If you've never tried it, I highly recommend visiting the Project Implicit website and trying the IAT for yourself -- always an enlightening experience.

Verschuere, B., Prati, V., & Houwer, J. (2009). Cheating the Lie Detector: Faking in the Autobiographical Implicit Association Test Psychological Science, 20 (4), 410-413 DOI: 10.1111/j.1467-9280.2009.02308.x

More like this

Psychology buffs are probably familiar with the Implicit Association Test. In laboratory experiments testing for racial prejudice, subjects are timed in their matching of positive and negative words with either white or black faces. In these tests, whites overwhelmingly react more positively to…
A common defense in murder cases is "focal retrograde amnesia": the defendant claims to have simply forgotten what occurred around the time of the crime (perhaps due having consumed too much alcohol or other drugs). In fact, "amnesia" is claimed in as many as 45 percent of murders. Psychologists…
Spence et al. wanted to test the use of fMRI for lie detection. In order to do so, they found a subject who had been convicted for child molestation because she has Munchausen's syndrome by proxy. There are two important parts of background to this piece. First, Munchausen's syndrome by proxy (…
What's the best way to ensure that law enforcement officers don't abuse their authority and coerce innocent suspects into confessing? Yesterday we discussed research suggesting that a side-view videotape of a confession was more likely than a head-on view to result in an accurate assessment of…

Penn and Teller recently did a great show on why polygraph based Lie Detector tests are "bullshit". (thats the name of the show, for those who don't know).

Anyway it was interesting to note how a test -can- work but more because of how the examiner was giving the test, not by the test itself. The successful polygraph examiner is more of a study in human psychology than an expert in the polygraph operation itself. It reminds me of the joke of the police station that used a "Lie Detector" machine that was really a copy machine with a paper saying "He's Lying!" in it.

This type of test also measures the difficulty a touch-typist has in trying to arbitrarily reprogram his years or decades of successful typing. Since the developers of this stuff never bothered to test two-fingered typists against touch-typists, they have no idea how badly typing habits confound their results.

I had never heard of the aIAT before, but it's an interesting application of the general IAT. However, the Verschuere et al. research demonstrates nicely a general problem I have with IAT research - it is very hard to interpret what, precisely, an IAT measures. There is a big distinction between the results of an IAT and a person's explicit beliefs - while explicit beliefs measure a more deliberative process, the IAT measures a spontaneous one. Thus, explicit beliefs tend to correlate better with deliberative behaviors, while IATs correlate with automatic, spontaneous ones.

The distinction between beliefs and spontaneous associations is a subtle one, and I don't think the research has adequately explained just what the distinction is. However, that hasn't stopped the media (and some researchers!) from making sensationalist claims about what the IAT can or can't do. This produces acrimonious debates about whether, for example, the race IAT can discover closet racists. The race IAT does not discover your "true" feelings about blacks; it simply discovers implicit associations that you may or may not be aware of. These associations can be overridden by a variety of factors and can be faked, as the Verschuere research shows.


The keyboard example is just so you understand what was going on. Real experiments use special response boxes because the latency of the keyboard is too unreliable. Touch typing is not related at all.

Whoops, meant #2 :)

A good source for information on the polygraph (comment #1) is www.antipolygraph.org

There you will find links to the P&T episode and lots of background material. I also found this quote there.

Make-believe science yields
make-believe security.

P&T will show you how an asshole with a little training (the polygraphist) can be defeated by a little training of your asshole.

By JohnnieCanuck (not verified) on 29 Jul 2009 #permalink

This paper makes no mention of the ways in which many IAT based tests screen for distorted responses. In addition research suggests that although scores can be distorted by some people who would already have got low scores, those with higher scores cannot trick the system. In real world use when the emotion of the event was real and not imagined the effect is much stronger and more difficult to cover up.

Nobody has yet mentioned sample sizes!

By Pete Jones (not verified) on 29 Jul 2009 #permalink

@ Steve

Actually most experimenters just settle for using a normal keyboard and get results just fine. I agree touch typing is not related though:)

@In General

Several researchers have (or are) developed algorythms to detect fakers and have booked decent results.

Anybody read the new IAT meta analysis by Greenwald et al., in JPSP however?
Apparently it's not much better than explicit measures in most cases. Good thing this happened before a huge chunk of experimental psychologists started relying only on implicit measures... oh wait... :(

"Reaction times are measured. If he's truly guilty, then he should react faster to "guilty" statements paired with "true." If he's innocent, then he should react faster to "innocent" statements paired with "true."

Reaction times indicate guilt or innocence? That really sounds scientific doesn't it? Why we don't we just measure the shape of his skull to get the true measure of his personality? It worked in the 19th century.

Don't you think that an innocent might feel insecure during the test, fearing a guilty result, and because of that hesitate before the answers?

Yes Fernando. Perhaps someone who has a history of being unfairly criticized or being punished for things that weren't his or her fault would worry more about a guilty result, and therefore hesitate more, than someone who had always been given the benefit of the doubt.

But in that case they will slow down for both types of questions.

I do agree that these tests are valuable but I think that they should be one of many factors in determining someone's unconscious associations. In the future, I hope this is not used as a single tool like the Meyers-Briggs tests.

I agree that a lie detector test could possibly be a useful tool, when combined with other tools. I think that, unfortunately, people sometimes expect to get an exact, incontrovertible answer from a single source.

I agree that lie detectors can be useful but they generally seem to only be useful in the way they are used. An operator can manipulate the person being tested easily.

Why did they redo the not-guilty condition in the "fake" test? It should be identical to the results for the first test.

Also, it wasn't. The difference between the innocent group on the first test and on the second, was larger than the difference between the guilty and innocent groups on the first test. This indicates that the test doesn't work.

By Phil Goetz (not verified) on 28 Aug 2009 #permalink