The Confidence Problem
Ask any police officer, trial lawyer, or hiring manager and they will tell you the same thing: they can tell when someone is lying. The eyes dart, the story shifts, something feels off. This confidence runs so deep that courts in every U.S. state allow judges and jurors to weigh "demeanor evidence" (a witness's body language, eye contact, and apparent nervousness) when deciding who to believe.
It turns out that confidence is unfounded. The largest study of human lie detection ever conducted reached a stark conclusion. The ability barely exists.
What 206 Experiments Revealed
In 2006, psychologists Charles Bond (Texas Christian University) and Bella DePaulo (UC Santa Barbara) published a meta-analysis synthesizing every available experiment on lie detection. They gathered data from 206 documents and 24,483 human judges, each attempting to distinguish lies from truths without special aids.
The average accuracy was 54%. A coin flip would give you 50%.
But the 54% average conceals a troubling asymmetry: people correctly identified 61% of truthful statements as truthful, a bias toward assuming honesty, but they caught only 47% of lies, making humans actually worse than a coin flip at the specific task of detecting deception.
Experience Changes Nothing
The most uncomfortable finding comes from a companion meta-analysis by Aamodt and Custer, published that same year, which compared lie detection accuracy across professions to test whether years of experience with deception sharpen the skill.
The results were blunt: judges achieved 59% accuracy, police officers 55.3%, customs officers 55.3%, federal officers 54.5%, and detectives scored just 50.8%, literally indistinguishable from flipping a coin. Untrained college students averaged 54.2%, no professional group showed a statistically significant advantage, and a detective with twenty years of interrogation experience reads deception no better than a sophomore in a psychology lab.
The Cues Do Not Exist
The reason is not that professionals rely on the wrong signals; the signals themselves are almost nonexistent. In 2003, DePaulo and five colleagues published a landmark review of 158 potential behavioral cues to deception across 1,338 separate estimates drawn from 120 independent samples. Their conclusion was unequivocal: most behavioral cues show "no discernible links, or only weak links" to whether someone is actually lying. Gaze aversion, fidgeting, nervous hand movements, vocal pitch changes: the entire catalog of folk wisdom about liars' bodies is not reliably produced by people who are being deceptive, because anxious truth-tellers display them while confident liars suppress them entirely.
Hartwig and Bond confirmed this in a 2011 meta-analysis in Psychological Bulletin that applied Brunswik's lens model to deception judgments, and their finding was counterintuitive in its own right: people do not even rely on the wrong cues, exactly. The intuitions humans hold about which behaviors correlate with deception are weakly correct, but the behavioral differences between liars and truth-tellers are so faint that even accurate intuitions produce almost no detection advantage; the signal is not misread so much as it barely exists.
The popular claim that facial microexpressions reveal deception has fared worst of all. Jordan and colleagues tested Paul Ekman's Micro Expression Training Tool in 2019 and found that trained participants scored below chance when judging real-world high-stakes lies, meaning the training actually made them worse.
A Calculation the Legal System Should Run
Here is where the 54% average stops being an academic curiosity and becomes a structural problem: consider a courtroom where a judge must decide whether to credit a witness's testimony. Using Bond and DePaulo's asymmetric detection rates (47% lie detection, 61% truth detection), Bayesian analysis reveals the positive predictive value of a "lying" judgment.
If the base rate of deception is 50%, then when a judge concludes "this person is lying," the person is actually lying only about 55% of the time.
Now consider a more realistic scenario: most witnesses testify under oath and most testimony contains substantial truth, so if the base rate of deception is closer to 30%, the positive predictive value drops to 34%. For every three people a judge or juror tags as deceptive based on demeanor, two are telling the truth; the math bears this out (true positives = 0.30 ร 0.47 = 0.141; false positives = 0.70 ร 0.39 = 0.273; PPV = 0.141 รท 0.414 = 34%), numbers at which no diagnostic test in medicine would be considered reliable.
The Strongest Counterargument
Timothy Levine's Truth-Default Theory offers the most sophisticated challenge to the bleak 54% figure, arguing that the number is partly an artifact of experimental design: subjects receive equal numbers of lies and truths and classify each one, a situation unlike real conversation where most communication is honest. Our tendency to assume truthfulness, which drags accuracy down in these experiments, is actually adaptive because it is usually correct, and more critically, when people can ask questions, probe inconsistencies, and access verifiable evidence rather than passively watching behavior, their accuracy improves considerably, with Levine's own experiments using diagnostic interview questions pushing detection rates above 70%.
This is a genuine limitation. But the passive-observation paradigm that Bond and DePaulo measured is precisely the paradigm courtrooms use: jurors watch witnesses, assess whether they "seem" truthful, cannot cross-examine or independently verify, and are doing exactly what the 54% figure measures.
What We Didn't Prove
The meta-analysis examines primarily low-stakes laboratory deception, where subjects lie about trivial matters. High-stakes deception in real criminal investigations may produce different cues, though the available evidence is not encouraging: Mann and colleagues' 2004 study of police suspects in actual interrogations found accuracy rates of 65% for lies and 64% for truths among experienced officers, better than the laboratory average but still far from reliable for individual judgments. The studies draw heavily from Western, educated populations, cross-cultural variation in deception cues remains understudied, and while computational methods using linguistic analysis have shown modestly higher accuracy than unaided human judgment, none approach the reliability that legal fact-finding demands.
The Bottom Line
A meta-analysis of 206 experiments and 24,483 judges found that humans detect lies at 54% accuracy, just four percentage points above a coin flip. Trained professionals โ judges, police officers, customs agents, federal investigators โ perform no better than untrained college students, and the behavioral cues that folk wisdom associates with deception are so faint as to be functionally nonexistent. Every courtroom that admits demeanor evidence and every interrogation room that trusts an officer's gut is relying on a tool that barely outperforms random chance.
What You Can Do
Stop trusting your gut about whether someone is lying, because the research is clear that the feeling of certainty and actual accuracy are uncorrelated. If you serve on a jury, weigh testimony based on consistency with physical evidence and with other testimony, not on how the witness "seemed." If you manage people, structure interviews around verifiable claims rather than reading body language. If you are a parent, know that the belief you can always tell when your child is lying is almost certainly wrong, and building trust through open conversation will serve you better than interrogation. And if you work in policy or law enforcement, advocate for evidence-based interview protocols like the United Kingdom's PEACE model, which focuses on gathering information through open-ended questions rather than reading nonverbal behavior to catch deception.