investigation 10 · statistics & forensics
For a century the polygraph has decided who gets hired, cleared, and believed. It is the perfect statistics case, because the machine itself is actually real — it measures genuine physiology, reproducibly — the search for a better test is genuinely open, and "it detects lies" is an overclaim. And the part that decides careers isn't in the box at all. It's in a formula you can run yourself: when liars are rare, what fraction of alarms are real?
The machine reads arousal. Bayes reads the machine. Only one of them is on your side.
▽ run the screeningTake the honest core first. The polygraph measures real things: skin conductance, heart rate, respiration, blood pressure — all genuine, reproducible channels of autonomic arousal. The traces are not faked and the physiology is not pseudoscience. On lab studies the instrument really does separate deceptive from truthful subjects well above chance: the National Academy of Sciences put the accuracy near an AUC of 0.86 — far above coin-flipping, far below certainty.
Now run the arithmetic the brochure skips. Screening is a base-rate problem: with sensitivity $se$, specificity $sp$, and a fraction $\pi$ of actual deceivers, the chance that a flagged person is actually lying is
Feed it the espionage numbers — 10,000 employees, 10 real spies, and a good operating point on the NAS curve (specificity 75%, sensitivity 80%) — and the machine catches 8 of the 10 spies while flagging 2,498 innocent employees. One real spy per ~313 alarms; a flagged employee is lying 0.32% of the time. The instrument didn't fail. The base rate did — and no threshold on the curve fixes it.
Then drag the prevalence slider to a criminal-suspect pool, where half the subjects are deceptive, and the same machine suddenly looks brilliant: PPV 76%. That is the whole illusion in one slider — field examiners work where liars are common, so the polygraph feels accurate; security screening works where spies are rare, so it mass-produces false accusations. Aldrich Ames, meanwhile, passed twice.
The same claim, read by parallax — from the measured pulse to the security clearance.
Signal. Skin conductance, heart rate, respiration and blood pressure genuinely track autonomic arousal; psychophysiology is a working experimental science, and the polygraph's channels are standard instruments in it. Stress really does move the needle.
Kill test. Startle a subject and predict the trace. It survives, every time, in any lab — the measurement layer is not the problem.
↳ run the screening aboveSignal. The Concealed Information Test flips the logic — instead of "are you lying?", it asks whether you recognize details only the culprit could know, and the statistics are honestly framed as chance against recognition. Japanese police use it routinely; EEG and fMRI memory-detection research is active.
Unresolved. Lab performance is promising; field validity — messy crime scenes, leaked details, stressed innocents — is genuinely undecided.
Kill test. Pre-registered field trials against ground truth, with information leakage controlled. Being run now; still open.
Signal. The kernel is real: deceptive subjects are, on average, more aroused by the questions that matter — above-chance lab accuracy (AUC ≈ 0.86) is a measured fact, not folklore.
Noise. Arousal is not deceit: fear, anger, and surprise write the same trace, countermeasures work, and the honest-but-terrified fail while trained liars pass — Ames in 1986 and 1991, Hanssen never caught by it. And in screening, the base-rate arithmetic above condemns thousands to catch a handful.
Kill test. The NAS ran it (2003): accuracy claims collapse outside the lab, and no scientific basis supports screening use. Run, and failed.
Signal. The ritual works on people — confessions really do happen in the chair. The fear is real; feelings are data about people.
Noise. There is no lie signature in human physiology to detect. The courts said so first — Frye (1923), the case that created the admissibility standard, was about this very machine — and Congress banned most private-employment use (EPPA, 1988). Yet it still gates security clearances: an instrument whose real mechanism is intimidation.
Kill test. Name the physiological signal unique to lying. A century in, no one has — a detector with no defined signal forbids nothing.
Notice the sift: the needle is honest — it's the inference that lies. The measurement layer survives every test; the "detector" fails the moment you write Bayes' rule next to it. The arithmetic of rare liars condemns the honest by the thousand, and no polish on the sensors can buy back a base rate.