Charles Bond and Bella DePaulo assembled every study they could find in which people had judged whether someone was lying, under conditions where the truth was known. Around two hundred and ninety-two independent samples, tens of thousands of judgments, spanning decades.
The average accuracy was 54 percent. Chance was 50.
That number is the single most important fact in this part of the guide, and it should be held against every confident claim about reading people that you encounter anywhere else.
Several features of the result matter as much as the headline.
The variance between observers was tiny. If some people were genuinely gifted at this, the distribution of individual accuracy would be wide, with a right tail of talented detectors. It is not. Bond and DePaulo found individual differences in judging ability to be very small — smaller than the differences between senders in how easy they were to read. There is far more variation in how transparent liars are than in how perceptive observers are.
Professional experience did not help. Police officers, judges, customs officers, psychiatrists and interrogators score around the same as students. What professional experience does reliably increase is confidence, and confidence is uncorrelated with accuracy in this domain — one of the most consistent findings in the literature.
There is a truth bias: people accept messages as true more often than not, so accuracy on truthful messages runs above chance and accuracy on lies runs below it. Groups trained to be suspicious flip the asymmetry without improving overall accuracy; they simply move the errors.
Why is the ceiling so low? Two reasons, and they compound.
The valid cues are weak. DePaulo's 2003 meta-analysis of 158 behavioral cues found most had effect sizes near zero and the strongest were small. There is no Pinocchio's nose, because there is no behavior that deception uniquely produces — deception produces arousal and cognitive load, which many other things also produce.
And observers use the wrong cues anyway. People watch for gaze aversion, fidgeting, and nervousness, none of which discriminate, while ignoring the weak signals that do. The next chapter is about how a stereotype that fails every test came to be believed almost everywhere on earth.
The consequences are not academic. Credibility findings in courtrooms rest on the assumption that a fact-finder can assess truthfulness from demeanor; asylum determinations rest on it; hiring and security clearance interviews rest on it. In each case a system is treating a 54 percent instrument as though it were reliable, and building consequential decisions on the output.
What follows from this is not despair, and it is not that deception is undetectable. It is that detection through passive observation of behavior is a dead end, and that everything which does work operates differently: by increasing cognitive load so that the liar's account degrades faster than the truth-teller's, by withholding evidence so that the account can be checked against it, by asking questions the liar cannot have rehearsed, and by counting details that can be independently verified.
Those methods are in Part 4, and they beat 54 percent by a real margin. All of them share a property: they work on the account, not on the person.
See Defense: evidence-based questioning over intuition