Prospects of functional magnetic resonance imaging as lie detector

Frontiers in Human Neuroscience 7:594, doi 10.3389/fnhum.2013.00594; open access (CC BY 3.0), PMC3781577. Review for a mixed science and law readership. Read for research direction R2, 4 October 2026. Provenance: papers/amadeus/rusconi2013_fmri_lie_detector.provenance.json.

What was read

• Read in full: the PMC XML converted to text: abstract, introduction, fMRI basics, "the lying brain", scientific hurdles, legal and ethical hurdles, concluding points, footnotes 1 to 12, statements and the reference list. No supplement exists.
• Choice and status: chosen as the open-access review of fMRI lie detection after two better candidates proved unavailable. Rustad and Brennen (2026, Applied Cognitive Psychology) is open access, but Wiley served a Cloudflare bot check, which the rules forbid circumventing. Farah et al. (2014, Nature Reviews Neuroscience) is paywalled, and its Penn-hosted copy no longer resolves. This review is from 2013. It predates the multivariate (MVPA) lie-detection and countermeasure studies after 2013. Rissman and Murphy 2024 covers those for memory detection.

What it argues, with the numbers it gives (verified as the paper's statements)

• People are poor lie detectors. Observers judge a stranger's truth or lie correctly about 54% of the time. Lawyers, police, judges and psychiatrists do no better (Bond and DePaulo 2006).
• What fMRI "lie" signals are. Lie-minus-truth contrasts show a parieto-frontal network, and the reverse contrast shows nothing. The usual reading is that lying takes extra effort: suppressing the truth and building an alternative. Most results are group averages, while application needs individual-level evidence.
• The scientific hurdles.
  • Reverse inference. Brain states map to mental states many-to-many, so prefrontal activity does not mean a lie (Poldrack 2006).
  • Construct validity. fMRI "lie detection" really detects suppression of a competing response, without knowing what was suppressed. Defendants suppress anger and the urge to blurt; anxiety and fear also raise the signal.
  • Individual differences and untested groups. Juveniles, older people, people with addictions, delusions or confabulation, and other cultures have not been studied. What counts as a lie is itself a social convention.
  • Countermeasures. Trained participants defeat fMRI tests, for example with covert mental arithmetic during control questions (Ganis et al. 2011). Whether a well-rehearsed story, retold with little effort, escapes detection is open.
  • Questioner expectancy shapes the subject's brain activity.
  • Analysis choices. Each group uses its own statistical method, and a cited meta-review reports non-independence errors in more than half of studies.
  • Data loss. About 20% of subjects in fMRI lie studies from 2001 to 2006 were rejected for head motion or insufficient data. Implants and claustrophobia exclude others.
  • Reported accuracies. 78–90%, from few individual-level studies with compliant volunteers. Two companies (No Lie MRI, Cephos) claimed about 90%, against about 70% for the polygraph.
  • External validity. Lab lies are instructed, low-stakes and of one type. Real lies vary: denial, alteration, or a truth told so as to sound like a lie. Validation would need objective ground truth, and court verdicts cannot supply it.
• Legal and ethical hurdles.
  • Rights engaged: unlawful search, the right to silence and against self-incrimination (whether brain activity exists "independent of the will"), privacy (ECHR art. 8), freedom of thought (art. 9), dignity and mental integrity (EU Charter arts 1 and 3), and data protection, including a right to rectify a contested interpretation of one's own brain data.
  • France (Civil Code art. 16-14, 2011) limits brain imaging to medicine, research and expert judicial enquiry, with express, written, revocable-at-any-time consent.
  • Other risks: covert or compelled scanning, "fishing trips", jurors over-weighting brain images, and fairness when only one party is scanned.
• Conclusion. fMRI lie detection meets neither scientific nor legal standards. Accuracy figures near 0.90 come from compliant participants and would "drop dramatically" with non-compliance and countermeasures. fMRI is "unlikely to constitute a viable lie detector for criminal courts".

Limits

• Narrative review, 2013, with no quantitative synthesis. Several claims rest on secondary legal sources, and some figures (the 20% loss, the "more than half" faulty studies) are relayed second-hand.
• Courtroom framing. It does not consider a cooperative subject who wants to be read accurately.

What it means for Amadeus (inference)

• Neural lie detection, by construction, cannot catch self-deception. It detects the effort of suppressing a truth the person holds. Someone who believes their own account suppresses nothing. So the signal P6 would want for self-deception does not exist in this paradigm, and practised lies may also escape it. This agrees with Rissman and Murphy 2024 for memory: the brain shows what the person believes.
• Neither people nor models should be trusted to spot lies from speech. If human judges are at 54%, an interviewing model has no basis for treating a person's style as a lie signal. Lies to others must be caught structurally:
  • against records (what they did);
  • by consistency across sessions and contexts;
  • by asking in ways that make a lie costly to maintain.
• Rules for a month of recording.
  • Consent revocable at any time, as in the French rule, and the participant's right to see and contest interpretations of their own neural data. P6's "believed by the person" field should be contestable by the person.
  • Expect around 20% data loss in scanner sessions from motion.
  • Treat other people's speech captured in the recording as their personal data too.