The Pentagon wants $30 million to build an AI-powered lie detector

The Department of Defense is seeking $30.3 million over five years for Polygraph+ (also called Polygraph Next), according to a budget request first reported by Inside Defense and detailed by MIT Technology Review. The program, run by the Defense Counterintelligence and Security Agency if Congress approves it, aims to modernize federal polygraph and credibility assessment with AI and machine-learning scoring algorithms plus “standoff sensing” — physiological readings without clipping sensors to a person.

It arrives amid reported leak hunts: the New York Times said about 50 Joint Staff officers were polygraphed after coverage of U.S. weapons stockpiles in the Iran war. Planned uses include vetting prospective employees and insider-threat detection. Experts in the piece call AI-enabled lie detection “the worst of both worlds.”

We’re watching the federal government try to upgrade a technology that has barely changed since the 1920s — blood pressure, pulse, breathing, sweat — by adding cameras, heat maps, pore tracking, and scoring models. Prior DIU work already pointed at Presage Technologies’ camera-based heart and breathing estimates and Altec Research’s head movement, facial skin temperature, and pore-activity prototype. That is the direction: multi-modal, contactless, algorithmically scored.

The scientific problem does not go away because the interface gets fancier. There is still no reliable Pinocchio’s nose. Humans spot lies barely better than chance without tools; the American Polygraph Association claims high accuracy, but the 2003 National Research Council called the evidence weak, and the Office of Technology Assessment in 1983 found limited support for employee screening. At DOD scale — millions of people — even an “80 to 94 percent” test produces tens of thousands of false accusations. Minority bias, examiner subjectivity, and countermeasures that beat the machine remain on the record.

AI could, in theory, find patterns examiners miss and combine stress, cognitive load, and concealment cues into a harder-to-game score. History is not encouraging: Silent Talker, iBorderCtrl, and AVATAR all chased multi-modal deception and faded. Legal scholars in the story argue you cannot train your way to ground truth when you never know which past polygraphs were actually right. That is invalidity plus opacity — confidence theater with a model card.

Useful AI belongs in security where it measures something real. Here the likelier read from the experts is deterrence and intimidation: people confess because they believe the machine works, and a modernized version under a leak-anxious administration risks amplifying that dynamic. I’m watching whether Congress funds this, which sensors get chosen, and whether “credibility assessment” stays a screening prop or gets sold as validated science it has not earned.

Current polygraphs remain rarely admissible in court. The budget document has not been approved. Presage, Altec, and the DIU did not expand on the Technology Review reporting beyond what the DIU previously disclosed about prototypes.