How to Tell If Someone Is Lying: What the Research Says
Published August 14, 2026 Β· 6 min read
Ask people how to spot a liar and you'll get the same answers everywhere on earth: they won't look you in the eye, they fidget, they touch their face, their story is too smooth, they hesitate.
A cross-cultural survey of beliefs about deception found gaze aversion to be the most widely held cue in dozens of countries. It is also one of the most thoroughly discredited. Meta-analyses examining the relationship between behavioural cues and deception have repeatedly found the links to be weak, inconsistent, and far too small to use on an individual.
This matters beyond trivia. It's the reason a room full of confident people watching a suspect will convict the nervous one, and it's the reason interrogation practice in several countries has been rebuilt over the last thirty years.
Here's what the evidence actually supports.
The baseline: 54%
The landmark finding comes from a meta-analysis by Charles Bond and Bella DePaulo, pooling hundreds of studies and tens of thousands of judges. Average accuracy at distinguishing truths from lies: around 54%. Chance is 50%.
More recent meta-analytic work has landed in the same neighbourhood β low fifties. Three findings from this literature are worth internalising:
Professional experience barely helps. Police officers, customs agents, judges and interrogators score close to laypeople. What experience reliably increases is confidence, which is why professional lie-detection judgements are dangerous rather than merely useless.
Individual differences are small. The idea of a natural human lie detector β the person who "just knows" β has not held up. Studies looking for consistently superior detectors have found the variation between people to be negligible.
There's a truth bias. People judge more statements as truthful than deceptive, which means accuracy on truths is higher than on lies. If you flip that and become suspicious of everyone, you don't get better; you just move your errors around.
The one variable that genuinely matters turns out to be the sender, not the receiver. Some people are transparent liars and some are opaque ones, and that difference swamps any difference between observers.
Why body language doesn't work
Three reasons, and they compound.
The cues are weak. Comprehensive meta-analysis of behavioural correlates of deception found that most popular indicators β gaze aversion, fidgeting, posture shifts, self-touching β have essentially no reliable relationship with lying. A handful of cues show statistically detectable but tiny relationships, far too small to classify an individual.
Nervousness isn't deception. This is the fatal confound. Being suspected, interviewed under pressure, or asked to account for your movements produces anxiety in innocent people β often more anxiety, because they have no rehearsed story and can't predict where the questions are going. Reading nervousness as guilt systematically targets the anxious, the neurodivergent, the young, and anyone unfamiliar with the situation.
Cues are cultural and individual. Eye-contact norms vary enormously between cultures. Baseline fidgeting varies between people. Without a long, unpressured baseline for the specific person in the specific situation, "unusual behaviour" is undefined.
The unavoidable conclusion of thirty years of work: watching someone's body tells you how they feel, and how they feel tells you almost nothing about whether they're telling the truth.
What does work: listen to the content
The consistent finding across the literature is that observers do better when they attend to what is said rather than how it's delivered. Statement content carries information; demeanour mostly doesn't.
Verifiability
Truthful accounts tend to contain more details that can be checked β named people, specific places, receipts, times that intersect with records. Deceptive accounts tend to be rich in unverifiable detail (internal states, private thoughts, activities with no witnesses) and thin in checkable ones.
This is the basis of the verifiability approach: rather than assessing whether an account sounds true, you count the checkable elements and then check them. It also has a useful property β telling someone in advance that you'll be verifying details changes the incentives, and liars who then add verifiable detail add falsifiable detail.
Consistency with evidence, not with itself
Internal consistency is a poor signal. Rehearsed lies are highly consistent; genuine memories are patchy, self-correcting and often contradictory in small ways. A perfectly consistent account across several retellings is weak evidence of truth and mild evidence of rehearsal.
Consistency with external evidence is a different matter entirely. That's the thing worth testing.
Strategic use of evidence
One of the better-supported interviewing techniques: withhold what you know, let the person commit to an account, then introduce the evidence. Liars unaware of what you hold tend to produce statements that contradict it; truth-tellers tend not to.
The mistake β and it's the near-universal amateur mistake β is opening with the contradiction. Do that and you've taught the subject exactly what you have, and they will repair every other part of the story around it.
Imposing cognitive load
Lying is more effortful than truth-telling: you have to construct a plausible account, keep it consistent, suppress the truth, and monitor whether you're being believed. Load the interview and deception becomes harder to sustain. Aldert Vrij and colleagues developed a family of techniques around this β asking for accounts in reverse chronological order, requiring unanticipated detail, asking someone to draw the scene, or asking questions in a jumbled sequence.
Meta-analytic estimates put the cognitive approach at roughly 60% accuracy β a real improvement over 54%, and an honest indication of the ceiling. Note the important caveat found in this literature: these techniques help observers mainly when they're told what to attend to. Simply running a harder interview doesn't automatically make the listener better.
Unanticipated questions
Liars prepare answers to the questions they expect. Ask something adjacent and unrehearsed β not "what did you do at the restaurant?" but "describe the layout, where the door was, who was seated nearest you" β and a fabricated account has to generate content on the spot while staying consistent with everything already said.
This is the single most practical technique on this page, and it's why detective fiction's best interrogation scenes are built around the sideways question rather than the direct accusation.
What about polygraphs?
The polygraph measures arousal β heart rate, respiration, skin conductance β not truthfulness. There's no physiological signature unique to lying.
A US National Academy of Sciences review concluded that polygraph testing performs better than chance in specific-incident settings but well short of what its use implies, and is unsuitable for screening, where the low base rate of guilt means false positives overwhelm true ones. Most US courts exclude polygraph results. Newer approaches β voice stress analysis, thermal imaging, fMRI-based methods β have not established validated accuracy for real-world individual use.
The technology varies. The underlying problem doesn't: arousal is produced by being suspected, not only by lying.
Why this changed how police interview people
The Reid technique, developed in the United States from the 1940s, was built on exactly the assumptions above: that trained observers can read deception from behaviour, and that a guilty person confronted with confident accusation will confess. Its nine-step structure is accusatory by design β assert guilt, cut off denials, offer face-saving justifications.
The criticism that accumulated against it is that this structure can produce false confessions, particularly from juveniles, people with intellectual disabilities, and anyone susceptible to suggestion or exhausted by a long interview. DNA exonerations have included cases where the convicted person had confessed. In 2017 one of the largest firms training the method publicly announced it would stop teaching it, citing the risk of false confessions.
The UK moved in the other direction after a series of miscarriages of justice, developing the PEACE model in the early 1990s β Preparation and planning, Engage and explain, Account, Closure, Evaluate. It is deliberately non-accusatory: the aim is to obtain a full account, then test it against evidence, rather than to secure a confession. It incorporates the cognitive interview, developed by Fisher and Geiselman, which improves the volume and accuracy of what witnesses recall by using open questions, context reinstatement and uninterrupted free recall.
The direction of travel across the research is consistent: information-gathering interviews outperform accusatory ones, both on accuracy and on the risk of false admissions.
A practical protocol
If you actually want to test an account β in an investigation, at work, or in a mystery game:
- Prepare first. Know what evidence you hold and, crucially, don't reveal it.
- Free recall, uninterrupted. "Tell me about that evening." Do not steer, correct or interject. Let it run long.
- Expand, out of order. Return to specific moments non-chronologically. Ask for the reverse sequence. Ask for sensory and spatial detail.
- Ask the unanticipated question. The one they haven't prepared for.
- Count the verifiable claims β then go and verify them.
- Only then, introduce a contradiction. One at a time, and stop talking afterwards.
- Judge the account against evidence, never against demeanour.
Notice that six of the seven steps involve no assessment of the person at all. That's the point. Deception detection isn't a perceptual skill; it's an evidence-management procedure.
What this means for mystery players and writers
Two things, and they run in opposite directions.
For players: stop reading suspects and start reading statements. In an AI mystery, this is a genuine advantage, because there is no body language to be misled by β the entire signal is content, which is the only channel that carries information anyway. Write down what each suspect claims, in their words, and test claims against each other and against physical evidence. Our 40 questions to ask suspects is organised around free recall, expansion and challenge for exactly this reason.
For writers: the honest version of this research is better material than the folklore. A detective who catches a liar by noticing a wobble in their voice is doing something no human can reliably do. A detective who catches one by asking a question the liar never prepared for, and watching the account produce a detail that contradicts a receipt, is doing something real β and the reader can follow it, which is the whole fair-play contract. See how to write a murder mystery.
Frequently asked questions
What are the real signs someone is lying?
There are no reliable individual behavioural signs. Meta-analyses find the popular cues β gaze aversion, fidgeting, face-touching β have essentially no usable relationship with deception. The usable signals are in content: whether an account contains checkable details, whether it survives unanticipated questions, and whether it agrees with independent evidence.
Do liars avoid eye contact?
No. It's the most widely believed cue worldwide and one of the least supported. Some liars maintain more eye contact than usual, precisely because they know the stereotype and are managing it.
How accurate are people at detecting lies?
About 54% on average, across hundreds of studies β barely better than a coin flip. Professionals score similarly to everyone else but are considerably more confident, which makes their errors more consequential.
Can you get better at detecting lies?
Not much by observing harder. You can improve substantially by changing the interview: withholding evidence until the account is committed, imposing cognitive load, asking unanticipated questions, and verifying checkable claims. Meta-analytic work puts cognitive approaches around 60% β a real gain, and an honest picture of the ceiling.
Do polygraphs work?
They measure arousal, not truthfulness. Reviews find performance above chance in specific-incident testing but far below what their reputation implies, and unsuitable for screening. Most US courts don't admit the results.
How does this apply to a mystery game?
Directly β an AI suspect has no tells, so the only available channel is what they say. That's the same channel the research says is the useful one. Track claims rather than impressions, ask the question they can't have prepared for, and check statements against the physical evidence. Start with a case and try running the seven-step protocol above; The Crimson Rite and The Fourth Suspect both reward it heavily.
