Lie Detectors Don’t Detect Lies

Part 3 of 8 in a series on lies and deception at work.

Polygraphs measure nerves, not lies, and the new AI lie detectors repeat the same mistake faster. What a century of machines says about catching liars at work.

Police officers have been passing this story around since at least the late 1960s. Detectives have a suspect who won’t talk. They put a metal colander on his head and run wires from it to the office photocopier. In the copier is a sheet of paper that says “He’s lying”. Every time he gives an answer they don’t like, someone presses Copy. Eventually he confesses.

The story has been set in Radnor, Pennsylvania, at a US Navy base and in Detroit. David Simon told the Detroit version as fact in Homicide, his 1991 book about Baltimore’s murder detectives, and later put the trick on screen in The Wire. Nobody has pinned it to a real case. It’s probably a legend.

It survives because everyone who hears it understands why it would work. The photocopier detected nothing. The suspect’s belief in it did everything.

Hold on to that, because it explains a century of real lie-detection machines better than their makers ever have.

The first two essays in this series looked at why people are poor lie detectors, and at what actually gives lies away, which turned out to be evidence rather than faces. This one is about the machines that promise to do the job for us: the polygraph, voice analysis, brain scanners and, lately, artificial intelligence.

What the machine measures

A polygraph is a set of sensors and a recorder. Rubber tubes around your chest and stomach track your breathing. A cuff on your arm tracks your blood pressure and pulse. Pads on your fingertips measure how much you’re sweating. The name means “many writings”, from the pens that once scratched it all onto a moving roll of paper.

That tells you, fairly reliably, whether someone is aroused: nervous, frightened, angry, embarrassed or just uncomfortable. It can’t tell you why. There’s no pattern of breathing, sweating or heartbeat that belongs to lying and nothing else. Every polygraph examination has to get round that problem, and how it does so is worth knowing.

The most widely used format is the Comparison Question Test. You’re asked relevant questions about the matter in hand, such as whether you took the money from the safe. Mixed in are comparison questions designed to make almost anyone uneasy, such as whether, before this year, you ever took something that didn’t belong to you. The examiner steers you towards answering no, which for most people is a small lie. The theory is that an innocent person reacts more to the comparison questions and a guilty person to the relevant ones.

So the standard lie detector test works by nudging you into a lie, to give it something to measure against. It also depends heavily on the examiner, who writes the questions and interprets the charts.

In 2004 a British Psychological Society working party reviewed the evidence. Depending on which review of real-world studies you went by, between 11% and 47% of innocent suspects were wrongly classified as lying.

A different design, the Concealed Information Test, checks whether you recognise details only the culprit would know. I covered it in the second essay. It’s far better grounded, and it’s the exception.

Where it came from

In America, the idea goes back to William Moulton Marston, who began measuring blood pressure during questioning as a Harvard graduate student in 1915. It was his wife Elizabeth, according to their son, who noticed that her blood pressure seemed to climb when she was angry or excited.

In the early 1920s Marston tested a murder suspect called James Frye, who passed. The trial court wouldn’t hear the result, and in 1923 the appeal court agreed, ruling that the technique hadn’t gained general acceptance in its field. That test of general acceptance governed scientific evidence in American courts for decades.

Meanwhile John Larson, a police officer with a PhD, built a machine in 1921 that recorded continuously, and newspapers gave it the name that stuck: the lie detector.

By the 1980s America was testing about two million people a year. Around 90% of those tests were ordered by private businesses, and three-quarters were for people applying for jobs.

In 1985 President Reagan ordered thousands of government employees to take polygraphs, then scaled the order back when his Secretary of State, George Shultz, threatened to resign rather than take one. In 1988 Reagan signed the Employee Polygraph Protection Act, which bans most private employers from asking staff or job applicants to take a lie detector test. Its definition of a lie detector covers voice stress analysers too. There are exceptions for security firms and drug companies, and for staff reasonably suspected of involvement in a specific theft or loss. Government agencies aren’t covered at all.

The polygraph never caught on in Britain in the same way. Results aren’t admissible as evidence in court here. Britain also never passed an equivalent of the American ban, and some UK employers do use polygraphs in internal investigations. What limits them is employment and data protection law.

Eight spies and 1,600 suspects

The most thorough review of the polygraph came from the US National Research Council, part of the National Academies. The Department of Energy commissioned it after a scare over nuclear weapons secrets in the late 1990s led it to start polygraphing thousands of its own staff. The committee’s verdict, published in 2003, was that the evidence for the polygraph was “scanty and scientifically weak”.

Its best-known passage is a thought experiment. An agency screens 10,000 employees, 10 of whom are spies. Set the test sensitive enough to catch eight of the spies, and about 1,600 people fail, almost all of them innocent. The agency now has to find eight spies among 1,600 suspects. Set it so that only about 40 people fail, and it catches just two spies.

Stephen Fienberg, the statistician who chaired the committee, spelled out what that meant at a congressional hearing in 2003. On the Department of Energy’s own expectations, anyone who failed its screening test had over a 99% chance of being truthful, while a spy taking it was far more likely to pass than fail, even without trying to cheat.

That’s the arithmetic of screening, and it applies to any test looking for something rare. Most people in any workplace aren’t stealing or leaking anything. Test all of them and even a fairly good machine accuses far more honest people than guilty ones. The rarer the thing you’re looking for, the worse it gets.

The spy who passed and the innocent man who failed

Aldrich Ames was a CIA officer who began selling secrets to the Soviet Union in 1985. He betrayed almost every Soviet agent working for the CIA, and at least ten of them were executed. He passed polygraph tests in 1986 and 1991. Asked later how, he said there was no special magic, just confidence and a friendly relationship with the examiner. His Soviet handlers had told him to relax. He was arrested in 1994 and died in prison in January 2026, aged 84.

The machine can fail the other way. In 1978 the owner of a carry-out shop in Perrysburg, Ohio, was shot by a man in a ski mask, and before he died he named a local man, Floyd “Buzz” Fay. Prosecutors offered Fay a deal: pass a polygraph and the charges would be dropped, fail and the result could be used against him. He failed two tests. He was convicted of aggravated murder, sentenced to life, and exonerated in 1980 after about two years in prison.

In prison he studied the polygraph. As the case is usually told, including in the British Psychological Society’s review, he coached 27 fellow inmates, all of whom had admitted their guilt to him, on how to beat the comparison test. After about 20 minutes’ instruction, 23 of them passed.

Between them, Ames and Fay are the whole problem. The machine waved through a spy whose betrayals got people killed, and helped put an innocent man in prison.

Why it still gets used

Given all that, it’s fair to wonder why anyone still bothers. The answer is the photocopier.

In 1971 the psychologists Edward Jones and Harold Sigall wired volunteers to an impressive-looking machine and told them it could measure their true feelings. It couldn’t. It was a prop. But people hooked up to it gave more honest answers about attitudes they’d have played down on a questionnaire. Jones and Sigall called it the bogus pipeline, and later reviews of the research found it reliably reduces the urge to give the socially acceptable answer.

The National Research Council saw the same thing in the real polygraph. It found no evidence that the polygraph draws out confessions any better than a fake machine that people believe works. Its usefulness comes from belief, not accuracy.

That’s more or less how the British state uses it. Since 2014, high-risk sex offenders released on licence in England and Wales can be required to take polygraph tests, and in 2021 this was extended to some terrorist offenders and, as a pilot, to high-risk domestic abuse perpetrators. Nobody can be sent back to prison just for failing. They can be recalled for what they disclose, or for refusing to take the test. Nothing they say can be used as evidence in a criminal trial. In the pilot that led to the national scheme, 76.5% of tested offenders made at least one disclosure, against 51.2% of those who weren’t tested.

Supporters argue the readings themselves add something beyond the belief. Whatever you make of that, the system has been designed around what the polygraph reliably does, which is get people talking. It isn’t really a lie detector. It’s a prompt to confess.

Using a machine to trick people into honesty raises its own questions, and I’ll come back to them at the end of this series.

Voices

The same thing happened with voice analysis, closer to home.

The Department for Work and Pensions trialled voice risk analysis, software that claimed to pick up stress in benefit claimants’ voices over the phone. It came from an Israeli company, Nemesysco, supplied through Capita. Harrow Council reported early savings in 2007, partly because claimants realised their answers weren’t adding up and didn’t want to risk prosecution. That’s what careful questioning does to someone who thinks they’re being checked. It says nothing about whether the software could tell.

In 2010, after trials costing £2.16 million, the DWP said the evidence wasn’t strong enough to recommend it. Two Swedish phoneticians, Anders Eriksson and Francisco Lacerda, had already examined the technology and found no scientific evidence that it worked. When Nemesysco’s lawyers threatened legal action, the journal’s publisher took their paper off its website. In 2014 the Guardian found that 24 councils had used the software or were considering it. A UK disability law site now warns that voice analysis may unfairly flag people who stammer.

On the telly

For many British viewers, the lie detector means the results on The Jeremy Kyle Show.

ITV cancelled the show in 2019 after the death of a guest, Steve Dymond. He died a week after filming, having taken the show’s lie detector test to prove he hadn’t cheated on his fiancée, and failed it. When MPs questioned the show’s producers, one said viewers had always been told the test wasn’t 100% accurate. The committee chairman called it “astonishing” that he couldn’t say how accurate it actually was. In 2024 the coroner found insufficient evidence that the show had contributed to Mr Dymond’s death. He also found insufficient evidence that Mr Dymond had lied during the test.

Whatever you make of the programme, that’s the polygraph in its purest form: a machine presented to millions as a verdict, by people who couldn’t say how often it was wrong.

Brains, and a toe

If the body is too vague, the obvious next move is to go straight to the brain. By the late 2000s, companies were selling lie detection by fMRI brain scan.

In 2011 Giorgio Ganis and colleagues tested it using a recognition design much like the Concealed Information Test. Without countermeasures they identified deception in individual volunteers with 100% accuracy. Then they taught the volunteers to make a tiny, invisible movement of a finger or toe whenever certain irrelevant items came up. Accuracy fell to 33%.

The ancient description of a liar I opened the first essay with has him rubbing his big toe along the ground. Three thousand years later, a wiggled toe can beat a brain scanner.

Faces

Modern cameras are very good at finding faces. They’ll pick one out of a crowd, lock focus on an eye and follow it across the frame.

In 2022 I was commissioned to photograph the fireworks at the opening and closing ceremonies of the Commonwealth Games in Birmingham, on the strength of my pictures of London’s New Year’s Eve display. Sony lent me a couple of cameras for the job. Before the opening ceremony I had plenty of time to photograph the teams waiting outside the stadium, right next to where I’d set up.

The face detection was very good. It just didn’t know which face mattered to me. I wanted one particular person in the crowd, and at first the camera kept locking onto someone else. You can fix that by narrowing the focus area, but the setting is buried deep in the menus. A beginner might never find it, and would simply accept whatever the camera decided.

Finding a face isn’t the same as knowing what’s in it. In 2009 one camera maker found that out in public. Joz Wang and her brother bought their mother a Nikon Coolpix S630, and whenever they photographed each other smiling, the screen asked “Did someone blink?” Nobody had. The warnings stopped only when her brother stared at the lens with his eyes forced wide open. The family is Taiwanese-American. Wang wrote about it online under the title “Racist Camera! No, I did not blink… I’m just Asian!” Nikon didn’t say what went wrong. Whatever the cause, the camera’s idea of an open eye didn’t fit her family.

Software that claims to read emotion does the same thing with higher stakes. In 2018 Lauren Rhue ran 400 official NBA player photos through two commercial emotion-recognition services. Both gave Black players more negative emotions than white players. Face++ rated Black players as angrier even when they were smiling as much, and Microsoft’s service rated them as more contemptuous when their expressions were ambiguous. Microsoft withdrew the emotion-detecting features from its facial recognition software in 2022.

In the first essay I wrote about the Brokaw hazard: misreading someone whose normal behaviour happens to look suspicious. Software doesn’t escape that hazard. It builds it in and applies it to everyone at once.

At work, the best-known case was HireVue, whose video-interview software large employers used to screen job applicants. It wasn’t trying to catch liars. It scored candidates partly on their faces. It stopped using facial analysis in 2020 and announced it in January 2021, saying the technology “wasn’t worth the concern”. That was a year after a privacy group complained to the US Federal Trade Commission. In 2021 the American insurer Lemonade boasted that its AI could spot fraud from non-verbal cues in customers’ claim videos, then deleted the posts after a backlash and said it didn’t use emotion recognition.

The most ambitious attempt was European. iBorderCtrl was a €4.5 million EU-funded research project in which travellers answered questions over a webcam while software looked for signs of lying. A computer-generated border guard asked the questions, and it was trialled with security agencies in Hungary, Latvia and Greece until 2019. A reporter from The Intercept tried it before crossing into Hungary, told the truth throughout, and was scored as lying on four questions out of sixteen.

The Pirate Party MEP Patrick Breyer went to court for the project’s documents. The EU’s courts let him see general material about the reliability and ethics of such systems, but ruled that the details of this one, including its false positives and its risk of discrimination, could stay secret to protect commercial interests. His final appeal failed in 2023. The most important fact about any lie detector is how often it accuses the innocent. For iBorderCtrl, that was a trade secret.

Regulators have partly caught up. In 2022 the UK’s Information Commissioner’s Office warned organisations off emotion-analysis technology, with its deputy commissioner saying such systems “may not work yet, or indeed ever”. Since February 2025 the EU’s AI Act has banned AI that infers the emotions of people at work or in education, except for medical or safety reasons, and the law’s own explanatory text points to the weak science behind these systems. The same law still allows AI “polygraphs and similar tools” at borders and in policing, as high-risk systems with extra safeguards. Britain isn’t bound by the Act.

The algorithm that made people accuse

The newest machines skip the body altogether and read what you write.

In 2024 researchers in Germany and France asked nearly 1,000 people to write two short statements about their plans for the weekend, one true and one invented, with a £2 bonus for making the invented one convincing. They trained an AI model on the results. It sorted true from false about 67% of the time.

Then they asked 2,040 people to judge the statements, with some groups able to consult the AI. Left to themselves, judges managed about 46%, no better than guessing, and fewer than one in five accused anyone of lying. Only about a third chose to ask the AI, but those who did followed it closely. When it said a statement was a lie, 85% of them made the accusation. They did what the beginner does with a new camera, and accepted whatever the machine decided.

The machine didn’t mainly make people more accurate. It made them readier to accuse. In the first essay I argued, following Tim Levine, that our reluctance to call people liars isn’t a flaw. It’s part of what lets people work together. A machine that strips that reluctance away changes how people treat each other, whether or not it’s right. An earlier study found the mirror-image problem: when people could overrule a machine that was right about 69% of the time, their combined accuracy fell back to chance.

The researchers expected organisations to take this up faster than individuals, naming job interviews and insurance claims. Some of it is already being sold, and it recycles old ideas. One 2026 guide to AI in insurance describes scoring callers on whether they pause before recalling facts and whether their phrasing sounds rehearsed. Those are the kinds of cues the second essay found weak at best. Putting them in software doesn’t make them true. It makes them faster.

There’s a quieter problem with all these systems. A machine learns what lying looks like from examples someone has labelled. In the lab, the lies are invented weekend plans written for a £2 bonus. At work, the lies that matter are told under pressure, often by people who’ve had time to prepare. Nobody has a large, reliably labelled collection of those, because if we could reliably label them we wouldn’t need the machine.

Readings and records

I worked in IT for years. When I saw a machine expose a lie at work, it didn’t read anybody. It remembered. In the second essay I described an audit at a former employer that came down to logs showing who had changed which figures, and when. No sensors and no scores, just one record checked against another.

On another occasion I was asked whether there was a way to look at what staff were doing and assess their productivity. There was. Microsoft 365’s tools showed who was working with whom and what share of their time it took. SharePoint showed who had written each file and shared document.

The SharePoint history is a record of who changed what, and when. The collaboration figures are estimates, built from meetings, emails, calls and chats. Neither tells you whether the work was any good. That still takes a person.

Records can be wrong too. Under English law, courts presume a computer was working properly unless someone shows otherwise. The rule dates from 1999, when an older safeguard requiring proof that a computer was working was repealed, and critics argue it made it far easier for the Post Office to prosecute sub-postmasters on the word of a faulty system. That’s the lie an organisation tells about its own machines, and it gets an essay of its own later in this series.

The difference is that a record can be checked. A reading can only be believed.

What this means at work

If anyone offers your organisation a machine that detects dishonesty, whether it reads bodies, voices, faces or text, don’t ask how accurate it is. Ask how many honest people it will accuse for every liar it catches, among people like yours, according to someone who doesn’t sell it. If nobody can answer, you’ve learned what you need to know.

If an employer asked me to take a polygraph, my answer would depend on what they were trying to find out, and whether I thought they were going about it ethically. Part of me would also be curious enough about the process to run my own experiment on it and see what came out.

Suspicion has costs beyond false accusations. The National Research Council warned that screening whole workforces could lower morale and productivity and drive good people away. What suspicion costs gets its own essay later in the series.

Only one was meant to be fiction

The National Research Council’s report on the polygraph is not a document with many jokes in it, but one of its appendices has a section headed “William Moulton Marston, the National Research Council, and Wonder Woman”. Marston created her in 1941, six years before he died, and the report itself calls her lasso “a link to the original and modern myth of the invincibility of the polygraph”.

So Marston left the world two inventions for making people honest. One was a blood-pressure test that grew into the polygraph. The other was a magic lasso in a comic book.

Only one of them was meant to be fiction.

Disclosure: Sony lent me cameras for my Birmingham 2022 Commonwealth Games commission, mentioned above. This essay isn’t sponsored.

Sources and further reading

Alder, K. (2007). The Lie Detectors: The History of an American Obsession. Free Press.

British Psychological Society (2004). A review of the current scientific status and fields of application of polygraphic deception detection. Final report of the BPS Working Party.

Congressional Research Service (2009). Polygraph Use by the Department of Energy: Issues for Congress (RL31988).

Eriksson, A. & Lacerda, F. (2007). Charlatanry in forensic speech science: a problem to be taken seriously. International Journal of Speech, Language and the Law, 14, 169–193.

Federal Bureau of Investigation. Famous cases: Aldrich Ames (fbi.gov).

Frye v. United States, 293 F. 1013 (D.C. Cir. 1923).

Ganis, G. et al. (2011). Lying in the scanner: covert countermeasures disrupt deception detection by functional magnetic resonance imaging. NeuroImage, 55, 312–319.

Gannon, T. A. et al. (2014). An evaluation of mandatory polygraph testing for sexual offenders in the United Kingdom. Sexual Abuse, 26, 178–203.

GB News (2024). Steve Dymond inquest: coroner rules ‘no causal link’ between Jeremy Kyle Show appearance and his death. 10 September.

Homeland Security Today (2026). Aldrich Ames, CIA’s most damaging spy, dies at 84 in federal prison.

Iacono, W. G. & Ben-Shakhar, G. (2019). Current status of forensic lie detection with the comparison question technique. Law and Human Behavior, 43, 86–98.

Information Commissioner’s Office (2022). Immature biometric technologies could be discriminating against people.

Jones, E. E. & Sigall, H. (1971). The bogus pipeline: a new paradigm for measuring affect and attitude. Psychological Bulletin, 76, 349–364.

Kleinberg, B. & Verschuere, B. (2021). How humans impair automated deception detection performance. Acta Psychologica.

Lykken, D. T. (1998). A Tremor in the Blood: Uses and Abuses of the Lie Detector. Plenum.

National Research Council (2003). The Polygraph and Lie Detection. National Academies Press.

Rhue, L. (2018). Racial influence on automated perceptions of emotions. SSRN.

Roese, N. J. & Jamieson, D. W. (1993). Twenty years of bogus pipeline research. Psychological Bulletin, 114, 363–375.

Rose, A. (2010). Are face-detection cameras racist? Time, 22 January.

Simon, D. (1991). Homicide: A Year on the Killing Streets. Houghton Mifflin.

von Schenk, A., Klockmann, V., Bonnefon, J.-F., Rahwan, I. & Köbis, N. (2024). Lie detection algorithms disrupt the social dynamics of accusation behavior. iScience, 27, 110201.

Wikipedia. Aldrich Ames (en.wikipedia.org).

Employee Polygraph Protection Act 1988 (US). Offender Management Act 2007, ss.28–30. Regulation (EU) 2024/1689 (AI Act), Article 5(1)(f) and Annex III.