Pentagon seeks $30.3 million for AI polygraph scoring despite decades of doubt

The US Department of Defense has asked Congress for $30.3 million over five years to modernize polygraph testing with artificial intelligence. According to MIT Technology Review, the proposed program — referred to as "Polygraph+" or "Polygraph Next" — would develop machine-learning scoring algorithms and a technique called "standoff sensing," which takes physiological readings from a subject without attaching any device to them.

What the money would buy

Details of the budget document were first reported by Inside Defense. It describes an effort to modernize federal polygraph and credibility-assessment technologies in order to improve their accuracy and reliability. The program would be run by the Defense Counterintelligence and Security Agency (DCSA), the office that handles background checks for the federal government, and the technology would be applied to vetting prospective employees and detecting insider threats.

Nothing is final: Congress has not approved the request, the specific technologies have not been selected, and the DCSA did not respond to a request for more information, MIT Technology Review reports.

An earlier effort offers hints about the direction. In 2023 the Defense Innovation Unit ran an open submission process for deception-detection products and picked two companies to build prototypes: Presage Technologies, which claims to measure heart rate and breathing rate using ordinary cameras, and Altec Research, a medical sensor firm expanding into non-contact sensing. A screenshot of Altec's prototype released by the DIU shows it tracking head movement, facial skin temperature and pore activity. Neither company responded to requests for comment, and the DIU declined to comment.

A technology with a long credibility gap

Conventional polygraph hardware has barely changed since the device was invented in the 1920s. Examiners measure blood pressure, pulse, breathing and sweat, then judge truthfulness by comparing physiological responses to innocuous control questions against responses to substantive ones.

The federal government administers tens of thousands of these tests each year, but the scientific record is poor. In 1983 Congress's Office of Technology Assessment found very limited evidence supporting polygraphs for employee screening, and in 2003 the US National Research Council judged the evidence on efficacy to be "weak at best." The American Polygraph Association claims 80 to 94 percent accuracy, but the NRC noted that even a screening test at that accuracy would produce large numbers of mistakes — and with 2.8 million staff, the Department of Defense could see tens of thousands of people wrongly flagged.

Other problems compound this. Interpretation is subjective, with different examiners reaching wildly different conclusions, and people from minority groups are more likely to be judged deceptive. The test can also be gamed: with training, subjects can artificially inflate their baseline responses, for instance by pressing on a pin hidden in their shoe. "If you know how it works, you can beat it," Sophie van der Zza, an associate professor who studies deception at Erasmus University in Rotterdam, told MIT Technology Review. She added that the machine's main real-world effect is deterrence, and that this only works while people believe the device functions.

Why AI may not fix it

In theory, machine learning could surface patterns in physiological data that human examiners miss, and could support multi-modal systems that combine several measurements into a single deception score that is harder to trick. Van der Zee notes that deception involves three underlying phenomena — physiological stress, cognitive load and deliberate concealment — while current polygraphs address only the first.

Precedents are discouraging. Manchester Metropolitan University's Silent Talker system, later folded into the EU-funded iBorderCtrl pilot, and the US AVATAR project combining eye-tracking, voice analysis and body-movement detection all quietly faded away. Kyri Kotsoglou, a professor at Northumbria Law School who studies polygraphs in the justice system, calls combining AI with the polygraph "the worst of both worlds," because it layers algorithmic uncertainty on top of an invalid instrument.

A deeper issue is ground truth. Marion Oswald, a law professor who has written with Kotsoglou, points out that even vast archives of polygraph records cannot confirm which results were correct, so there is no reliable label for an algorithm to learn from. She expects new lie-detection tools to serve, like the polygraph before them, as psychological props rather than scientific instruments.

The leak-hunt backdrop

The request arrives amid tension inside the Pentagon. Under Defense Secretary Pete Hegseth, the department has leaned heavily on polygraphs to hunt for sources of alleged press leaks; in September the New York Times reported that roughly 50 Joint Staff officers were tested after coverage of depleted US weapons stockpiles in the war with Iran. Oswald reads the program as a response to administration concerns about leaks and "perceived lack of loyalty," with lie detection used to intimidate and extract confessions rather than to obtain valid information.

Why it matters

If funded, an AI-scored lie detector would be deployed at enormous scale inside the largest US employer, in a climate where tests double as instruments of pressure in leak investigations. Critics argue the money would modernize the packaging of a discredited method rather than fix its core flaw: there is still no reliable, universal physiological signal of deception, and no ground-truth dataset from which an algorithm could learn one. For the millions of federal employees and applicants who could pass through such screening, the difference between deterrence theater and wrongful accusation is the practical stakes of the decision now sitting with Congress.