Fifty-Four Percent: Bond and DePaulo on Human Lie Detection

This is the same meta-analysis as chapter 51, and it appears again here because Part 4 exists to answer it.

Bond and DePaulo pooled decades of studies in which people judged whether someone was lying under conditions where the truth was known. Accuracy averaged 54 percent against a 50 percent baseline. Professionals scored no better than students. The variance between individual judges was negligible — there is no population of gifted detectors. Confidence was uncorrelated with accuracy.

DePaulo's earlier meta-analysis of 158 individual behavioral cues explains the ceiling. There is no Pinocchio response. The cues that show any relationship to deception have small effect sizes, they are inconsistent across contexts, and the cues people actually rely on — gaze aversion, fidgeting, nervousness — have none at all.

Timothy Levine's Truth-Default Theory supplies the last piece and reframes the finding as adaptive rather than pathetic. Human communication operates on a presumption of honesty, which people abandon only when triggered by something specific. This is not gullibility; it is correct. The overwhelming majority of statements people hear are true, verification is expensive, and a suspicious default would make ordinary life unworkable. The truth bias caps detection accuracy on lies while raising it on truths, and the net is 54 percent.

Levine's further point is that in real life we detect lies well — not from behavior in the moment, but afterwards, through evidence, third-party information, confessions, and inconsistency that emerges over time.

Which is the argument of this entire part.

Every technique in Part 4 that works shares one property: none of them assess the person. Strategic Use of Evidence produces a discrepancy between a statement and a document. The verifiability approach produces a list of checkable claims. The cognitive load approach manipulates the task and observes degradation. Unanticipated questions compare a prepared answer to an unprepared one. The Concealed Information Test measures recognition of a specific fact. The cognitive interview and PEACE improve the accuracy of the account rather than judging the speaker.

Every technique that fails shares the opposite property: it looks at the person and tries to read a state. Behavior analysis interviews, SCAN, voice stress analysis, gaze aversion, demeanor evidence, and the intuition of the experienced investigator all attempt to detect a lie by observing a liar, and they all land within a few points of chance.

That distinction is the single most useful thing in this part of the guide, and it generalizes well outside interrogation.

When you find yourself trying to work out whether someone is lying to you, the question you are asking has a known answer rate barely better than a coin. Replace it. Ask instead: what in this account can be checked, what would I expect to find if it were true, and what have they told me that is inconsistent with something I already know but have not mentioned.

Those questions have answers. Is he lying does not.