Can a Computer Tell When You Are Lying?

Machines have detected 5 key cues to spotting liars. But is that good enough?

Posted Dec 16, 2014

But could a computer do better?

Sure, they don't have the intuition that humans do, but they are also not at risk of the kinds of mistakes that are specific to humans, such as emotional investment in wanting to think that a certain person is lying, or that another sort of person would never lie to them.

Suppose you could give computers a transcript of the lies and truths that you said, wrote, or typed, and then programmed the machines to look for particular kinds of cues. How would the computers do? Of course, in this sort of approach to lie-detection, computers don't get to make judgments based on your lying eyes (or any other parts of you)—they just have to go by your words. Contrary to the conventional wisdom, though, language cues (what we say) offer even more promising clues about deception than do visual nonverbal cues (how we look).

The promise of finding accurate computer-based lie-detection has gripped researchers, and dozens have gone on a quest to see if it works. There are computer programs written to find and count relevant linguistic cues in transcripts, and social scientists have used them to see if computers can find any reliable differences in the transcripts of communications known to be lies compared to the transcripts known to be truths.

The question addressed in these studies is, "Do certain kinds of linguistic cues show up more (or less) frequently when people are known to be lying, compared to when they are known to be telling the truth?"

For a new review article, the authors rounded up 44 relevant studies. There were 38 different cues measured in enough of the studies that they could combine all the results and see which cues emerged as reliable and telling clues to deception.

First, I'll give you the best case for the success of the computers. Then I'll tell you why the computers were actually pretty unimpressive, much like humans are.

The Best Case for Computers: The Cues to Deception that They Found

Some cues separated the truths from the lies better than others did. In order of the strength of the cues (or magnitude of the effect sizes, for those who prefer the statistical jargon), here are the computer-identified cues to deception:

  1. Liars do not use as many different words as truth-tellers do. This is called "content word diversity." The results seem to suggest that liars don't access the same range of vocabulary that truth-tellers do. They fall back on the same words rather than using a variety of words.
  2. Liars' answers have fewer sentences and fewer words. Liars just don't seem to have as much to say as truth-tellers do. It is as if they are holding back, or maybe they are so busy trying to remember what to say or what not to say that they end up not saying much at all.
  3. Liars express anger more than truth-tellers do.
  4. Liars seem to make fewer exceptions than truth-tellers do. The computers figured this out by counting words such as except, but, and without. People who are telling the truth make more of these kinds of distinctions than liars do.
  5. Liars distance themselves from what they are saying. Specifically, compared to truth-tellers, they are less likely to use the first person ("I") and more likely to use the second person ("you") or the third person ("he" or "she" or "they").

The Unenthusiastic Case for Computers as Lie Detectors

Okay, so computers can separate liars from truth-tellers in the five ways I just described. But the authors looked at computers' use of 38 different linguistic cues, and only for half of them did the computers find any differences that were statistically significant—and some of those differences between lies and truths were very small.

Also, those conclusions I just offered above are based on averages across all of the 44 studies that included those cues. But the 44 studies varied in lots of different ways. For example, in some, the liars and truth-tellers were talking about their own personal experiences—often, emotional ones. In others, they were describing people they liked or disliked. In studies in which people described their own experiences, sometimes they described neutral experiences and other times, they were describing very negative experiences. In some studies, the liars and truth-tellers were typing (as in e-mail communications); other times, they were talking, and still other times they were writing by hand. Sometimes, the liars were highly motivated to get away with their lies, and the truth-tellers really cared about not getting mistaken as liars; other times, it just didn't matter much.

In all, the authors showed results for 15 different variations. You can think of them as different contexts for lying, or different kinds of lies, or different ways of lying, or different feelings about lying—what's important is that these variations matter. Liars lie in different ways in different contexts. For example, liars express more negative emotions (they use more words indicative of anger) than truth-tellers do when they are describing their own personal experiences; but when they are just talking about who they like and dislike, then liars and truth-tellers do not differ in their expression of negative emotions.

Here's something really striking: There was no one cue to deception that statistically separated the liars from the truth-tellers across all 15 different contexts and types of lies. The one that came closest was the number of words. For most of the different contexts and types of lies and feelings about lies, the liars had less to say (they said fewer words) than the truth-tellers. Again, though, the results were true for most contexts, but not all.

A perfect cue to deception would be one that occurs every time a person is lying, and never occurs when someone is telling the truth. The classic example is Pinocchio's nose. But the truth is, there is no Pinocchio's nose. It doesn't matter whether humans are looking for the cues or computers are: They just aren't there.

Reference: Hauch, V., Blandon-Gitlin, I., Masip, J., & Sporer, S. L. (in press). Are computers effective lie-detectors? A meta-analysis of linguistic cues to deception. Personality and Social Psychology Review. (Will probably appear in print in 2015.)

Notes: (1) The e-book versions of almost all of my books are on sale between December 16-22, 2014 (or between the 17th and the 23rd in the UK). Prices start at 99 cents on the first day and then increase gradually to the regular list price by the end of the 7th day. (2) Some of my books are also available in other languages, such as Spanish, Portuguese, Chinese, and Korean. Many of them are on sale this week, too. (3) For my other blog posts on deception, click here. (3) For what you may have actually come here for, discussions relevant to single life, you may want to check out "Harvard Business Review cares about women's success – but only if they are married with children," and other posts there.

Photo credit: PhotoExtremist.com/Google free use