Pentagon Seeks $30.3 Million for AI Lie-Detection Program Despite Reliability Concerns

A job applicant or employee could eventually face a credibility assessment built from AI scores and bodily signals collected without a device attached to them. The Pentagon is seeking $30.3 million over five years to develop that approach under a program called Polygraph+. Congress has not approved the request, and the central question remains: can a system identify deception from signals that have never been a dependable proof of lying?

A screen for people, not just answers

The Defense Counterintelligence and Security Agency, which conducts federal background checks, would run Polygraph+. The budget document identifies two intended uses: vetting prospective employees and detecting insider threats. That makes the proposed technology a potential tool for decisions about people’s access and trustworthiness, not simply an experiment in reading physical signals.

The proposal calls for AI and machine-learning algorithms to produce credibility scores. It also names “standoff sensing”: taking physiological readings without attaching equipment to a person. The budget request does not specify which sensors Polygraph+ would use. Nor has Congress approved the spending, so the request is not a funded deployment of an AI lie detector.

The request arrives amid a wider Pentagon push to investigate alleged leaks. Under Defense Secretary Pete Hegseth, the department has increasingly used polygraph tests to seek their sources. MIT Technology Review cites a New York Times report that around 50 Joint Staff officers were tested after coverage of depleted US weapons stockpiles during the war with Iran. That episode illustrates an existing demand for such tools; it does not show that Polygraph+ could identify a leaker.

What the earlier camera work shows

There are clues to what contact-free sensing might look like, though not a confirmed design for Polygraph+. In 2023, the Pentagon’s Defense Innovation Unit selected Presage Technologies and Altec Research to build deception-detection prototypes. Presage says standard cameras can measure heart and breathing rates. An image of Altec’s prototype shows readings for head movement, facial skin temperature and pore activity.

A conventional polygraph measures changes in blood pressure, pulse, breathing and sweat as an examiner compares responses to baseline and target questions. More sensors could give an algorithm a broader picture. Sophie van der Zee, who studies deception at Erasmus University Rotterdam, says combining measures of stress, mental effort and efforts to conceal a lie might improve detection. But a broader set of signals is still not the same as knowing whether an answer is true.

Even if you have all the records in the world from polygraph tests, you don't know whether those polygraph tests are right or not

Marion Oswald, professor of law, quoted by MIT Technology Review

The missing answer behind a credibility score

Oswald’s objection goes to the data an AI system would learn from. If earlier test results cannot reliably establish who lied, finding new patterns in those results cannot, on its own, establish a dependable measure of deception. Van der Zee says there is no universal physical sign of lying. Kyri Kotsoglou, a Northumbria Law School professor, argues that adding AI could compound the uncertainty of a method whose validity is already in doubt.

The concern predates AI. A 2003 National Research Council review described evidence for polygraph effectiveness as “weak at best.” The American Polygraph Association claims accuracy of 80% to 94%, but the council warned that even a screening test at that level could produce many mistakes. Interpretations can also vary between examiners, and research cited by MIT Technology Review indicates that people from minority groups are more likely to be judged deceptive.

The Pentagon’s proposal aims to make credibility assessments more accurate and reliable. For now, the unresolved test is whether Polygraph+ could establish that its scores reflect deception rather than stress or another response. Without that showing, a more elaborate measurement could still leave the person being screened facing an uncertain judgment.