Can AI detect lies from video? What the evidence says

← All articles · By the RedFlag team · September 4, 2026

Short answer: no. No AI system — and no human technique, and no polygraph — can determine from a video whether someone is lying. That is not a hedge; it is the consistent finding of five decades of deception research. What follows is the evidence, and an honest account of what is measurable.

Humans are barely better than a coin flip

The largest meta-analysis of deception judgments ever run — Bond and DePaulo (2006), covering about 25,000 judges across nearly 400 studies — found people distinguish truths from lies at roughly 54% accuracy, where 50% is chance. Training helps very little. Confidence in one's own judgment correlates almost not at all with being right.

There is no Pinocchio's nose

DePaulo and colleagues (2003) reviewed 158 candidate cues to deception across 120 samples. The result: most cues showed no reliable relationship with lying at all, and the ones that did were weak and inconsistent across situations. Gaze aversion — the single most popular folk cue — is not a reliable indicator. If a single behavioral tell existed, this literature would have found it.

The machines have not solved it either

• Polygraph: the U.S. National Research Council's 2003 review concluded that polygraph accuracy claims are "unfounded" for screening and that the theory linking arousal to deception is weak.
• Voice-stress analysis: independent evaluations of commercial VSA software (Harnsberger et al., 2009; a National Institute of Justice-funded field study by Damphousse, 2007) found deception-detection rates at or near chance.
• Vision-based AI: systems claiming to read deception from faces have drawn sustained criticism from researchers; the EU-funded iBorderCtrl border-screening trial became a cautionary tale, with its scientific basis publicly challenged. Micro-expression research (Porter & ten Brinke, 2008) found full one-sided micro-expressions to be rare even in high-stakes emotional lies.

An AI model trained on "lie vs. truth" labels inherits every problem above, plus dataset bias: most training corpora are low-stakes laboratory lies told by students, which behave differently from consequential real-world deception (Mann, Vrij & Bull, 2002).

What is honestly measurable

What survived the research is narrower and more interesting: lying is often harder work than telling the truth. Vrij, Fisher and Blank (2017) showed in a meta-analysis that approaches which increase a speaker's cognitive load widen the observable differences between truth-tellers and liars. Mental strain leaks into measurable channels — pitch variability flattens, blink patterns shift, pausing and speech rate change. None of these signals means "lying". They mean the speaker is working harder at this moment than at other moments.

That reframing is the honest version of this entire product category: not "this person is lying", but "this moment differs from this speaker's own baseline — look closer". The judgment about why stays with the human.

How RedFlag fits in

RedFlag is built on exactly that reframing. It is not a lie detector and does not output truth/lie judgments. It measures a speaker's pitch variability, blink rate relative to their own baseline, and voiced continuity, and flags the moments where the combination deviates from normal speaking ranges. The method, thresholds and test results are public: how RedFlag scores a video.

References

• Bond, C. F., & DePaulo, B. M. (2006). Accuracy of deception judgments. Personality and Social Psychology Review, 10(3), 214–234.
• DePaulo, B. M., et al. (2003). Cues to deception. Psychological Bulletin, 129(1), 74–118.
• National Research Council (2003). The Polygraph and Lie Detection. National Academies Press.
• Harnsberger, J. D., et al. (2009). Stress and deception in speech: evaluating layered voice analysis. Journal of Forensic Sciences, 54(3), 642–650.
• Damphousse, K. R. (2007). Voice stress analysis: only 15 percent of lies about drug use detected in field test. NIJ Journal, 259.
• Porter, S., & ten Brinke, L. (2008). Reading between the lies. Psychological Science, 19(5), 508–514.
• Mann, S., Vrij, A., & Bull, R. (2002). Suspects, lies, and videotape. Law and Human Behavior, 26(3), 365–376.
• Vrij, A., Fisher, R. P., & Blank, H. (2017). A cognitive approach to lie detection: a meta-analysis. Legal and Criminological Psychology, 22(1), 1–21.