The lie detector that cannot see a lie

The polygraph has been used for about a century, in police work, government security checks and, for a long time, in hiring. It records body signals such as heart rate, blood pressure, breathing and sweating while a person answers questions. An examiner then reads the chart and decides whether the person was deceptive. In films and on TV, the jumping needle is the moment the liar is caught.

In 2003 a committee of the US National Research Council published a review of the scientific evidence, called The Polygraph and Lie Detection. It found that, in tests about a specific incident, the polygraph can separate liars from truth-tellers better than chance but far from perfectly, and only for people who have not trained to beat it. For routine security screening of staff, accuracy was likely lower still.

Real cases show the gap. The CIA officer Aldrich Ames began spying for the Soviet Union in 1985 and passed two polygraph tests while doing so. He was arrested only in February 1994. In the United States, a 1988 federal law already barred most private employers from using lie detector tests on job applicants and staff.

The core problem is that there is no body signal unique to lying. A racing heart can mean guilt, but it can also mean fear of being disbelieved, worry about the test, or simply a nervous person. The committee also noted that every signal the machine measures can be changed on purpose. Its own worked example shows the cost: screening 10,000 staff with 10 spies among them, a setting strict enough to flag 8 of the spies would also flag about 1,600 loyal employees. Set it to flag far fewer innocent people, and 8 of the 10 spies pass.

The AI angle: many AI tools make the same promise in a new form. AI-text detectors claim to read authorship from word patterns, and "emotion AI" claims to read feelings or honesty from faces and voices. Like the polygraph, they turn a weak signal into a precise-looking score. A number on a screen feels objective, but if the signal does not reliably track the hidden thing, the score is not evidence.

The idea behind this caseLesson: What is AI, really? →

1. 1When a tool claims to reveal something hidden, such as lies, feelings or who wrote a text, ask what signal it actually measures and how closely that signal tracks the hidden thing.
2. 2Ask for the false-alarm rate, not just the hit rate. A tool that catches most cheaters but flags many honest people can do more harm than good.
3. 3Remember that people can learn to game a signal, so a tool can be beaten by exactly the people it is meant to catch.
4. 4Never let one score decide a hiring, school or legal outcome. Use it, at most, as a reason to look closer with other evidence.
5. 5If a chatbot or a vendor tells you such a tool is "scientifically proven", ask for independent reviews, not the company's own claims.