The Pentagon Is Spending $30 Million on an AI Lie Detector. Scientists Say It Won't Work.

Key points

• The US Department of Defense has requested $30.3 million over five years for a program called "Polygraph+", which would use artificial intelligence and machine learning to detect lies without physically attaching sensors to a subject.
• The traditional polygraph, which measures pulse and sweat, has been scientifically discredited since at least 2003, when the US National Research Council called the evidence for its accuracy "weak at best".
• Scientists say the core problem has never been solved: there is no single physical signal that reliably means a person is lying, across all people, in all situations.
• Around 50 Joint Staff officers were polygraphed after news reports about US weapons stockpiles during the war with Iran, according to the New York Times.
• Critics say adding AI creates a false sense of scientific authority on top of an already flawed foundation whose real effect is psychological pressure, not accurate detection.

There's a reason Pinocchio's nose is a fairy tale. A growing nose is a clean, unambiguous signal of a lie. In real life, no such signal exists. That hasn't stopped the Pentagon from trying to find one.

Five years, $30.3 million, and a program name that sounds like a software update: that is what the Department of Defense is requesting from Congress for "Polygraph+", also called "Polygraph Next". Not yet approved, the program would be run by the Defense Counterintelligence and Security Agency, the body responsible for federal employee vetting. Its goal is to use artificial intelligence (AI) and machine learning, the process by which software finds patterns in large datasets, to build smarter lie-detection tools. Central to the plan is "standoff sensing": reading physiological signals such as heart rate and skin temperature from across a room, no wires or contact required.

As first reported by MIT Technology Review, two companies have already built prototypes under a related Pentagon programme. Presage Technologies claims it can gauge heart rate and breathing from a standard camera. Altec Research, a medical sensor firm branching into remote detection, has demonstrated a system monitoring head movement, facial warmth and pore behaviour.

The science, though, hasn't kept pace with the ambition.

Does any lie detector actually work?

No. Not reliably. The traditional polygraph has been questioned for decades, and the verdict from official bodies is damning. A 2003 report by the US National Research Council found the evidence for polygraph accuracy to be "weak at best". A 1983 assessment by Congress's Office of Technology Assessment reached a similar conclusion. Without any device, humans spot a lie just over half the time, little better than a coin flip.

The American Polygraph Association claims accuracy rates of 80 to 94 percent. The NRC's own analysis showed that even at those rates, applied across the DoD's 2.8 million employees, a screening programme could falsely flag tens of thousands of innocent people. Polygraph results are rarely admitted as evidence in court for exactly this reason.

The deeper problem is structural. Sophie van der Zee, a deception researcher and associate professor at Erasmus University Rotterdam, puts it plainly: "There is still no Pinocchio's nose." Anxiety, guilt and fear all produce the same physiological fingerprint. People from minority groups are statistically more likely to be judged deceptive, and anyone who understands how the machine works can fool it. The most common trick is artificially spiking your body's response to easy baseline questions by, say, pressing a hidden pin into your foot, which muddies the results enough to beat the test.

Does AI fix any of this?

Not according to researchers who've spent careers on the question. Kyri Kotsoglou, a professor at Northumbria Law School who studies polygraph use in justice systems, calls combining AI with polygraph "the worst of both worlds". AI layers statistical uncertainty on top of a tool that was already scientifically invalid. And because researchers have no reliable "ground truth", no definitive record of who was actually lying during past tests, any machine-learning system trained on that data is learning from noise.

AI could theoretically spot patterns human examiners miss, and combining cognitive load measures with behavioural signals is theoretically stronger than relying on a single reading. Van der Zee has argued that more signals improve your odds. But every previous attempt to build this kind of system, from the EU-funded iBorderCtrl border-control pilot to the US AVATAR project combining eye-tracking with voice analysis, has folded after failing outside a laboratory.

We covered the Pentagon's appetite for AI tools in our 31 August report on the rollout of custom ChatGPT and Grok chatbots to 3 million Defence Department staff. That story was about tools where a measurable signal existed. Deception detection has no equivalent: the signal researchers need simply doesn't exist in clean enough form.

Pentagon timing matters here too. Under Defence Secretary Pete Hegseth, polygraph testing has risen sharply as the military hunts for press leakers. Van der Zee's research suggests the machine's biggest practical effect is already psychological: people confess before the test even starts, because they believe it works. "But that only works," she notes, "if people think a polygraph works."

Spend $30 million making it look more scientific, and you have a more intimidating interrogation prop. Whether you end up with a better lie detector is another matter entirely.

Common questions

What is "standoff sensing" and why does it matter?

Standoff sensing means reading someone's physiological data, heart rate or skin temperature, from a distance using cameras or remote sensors, without wires or physical contact. It matters because it removes one obvious way a subject might notice, and resist, the measurement.

Could this program affect ordinary people who aren't in the military?

The DCSA runs security background checks for civilian federal employees and contractors, not just military staff. If Congress approves the budget and the technology is deployed, anyone going through a federal security clearance process could encounter it.

Why do polygraphs still exist if the science doesn't support them?

Researchers point to two reasons: institutional momentum, the federal government has used them for decades and built procedures around them, and psychological effect, the belief that the test works can pressure people into confessions even when the technology itself can't reliably detect a lie.