Pentagon asks $30.3 million to put AI scoring on the 1920s polygraph

Under Defense Secretary Pete Hegseth, the Pentagon has been turning more often to polygraphs to find the sources of alleged press leaks ^[5]. In September, the New York Times reported that around 50 officers on the Joint Staff had been given polygraph tests after news coverage of depleted US weapons stockpiles in the war with Iran, according to MIT Technology Review ^[18]. The test they took has barely changed since the 1920s. It records blood pressure, pulse, breathing and sweat ^[9]. The examiner compares the reaction to a baseline question such as "Is the sky blue?" with the reaction to a target question such as "Have you ever committed a crime?" ^[9]

Polygraph+ is pitched as modernization, a way to make those tests more accurate and reliable ^[2]. The work it would actually do is screening job applicants and hunting insider threats ^[3]. The federal government already runs tens of thousands of polygraphs a year on employees, and courts rarely admit the results ^[11]. The budget line would upgrade a tool already in routine use, at about $6.06 million a year if the $30.3 million is spread evenly over five years ^[2].

Here's what teams tell themselves about the device. The American Polygraph Association claims it is 80 to 94 percent accurate, while research suggests people with no equipment spot a lie just over half the time ^[12][13]. Taken at face value, the association's own range leaves 6 to 20 wrong results per 100 tests, or 60 to 200 per 1,000 ^[1]. The 2003 National Research Council report made the same point: a screening test at that accuracy could still produce a lot of mistakes ^[14].

Here's what users actually do. Interviewees learn countermeasures, most often by inflating their response to the baseline questions, for example by stepping on a pin hidden in a shoe ^[16]. Different examiners get wildly different results, and people from minority groups are more likely to be judged deceptive ^[17]. An AI scorer could make results consistent from one examiner to the next. Standoff sensing removes the cuff and wires ^[1]. I'd expect the pin to work against a camera about as well as it works against a cuff. The Altec prototype the Defense Innovation Unit released tracks head movement, facial skin temperature and pore activity, all physical reactions ^[7], and the pin trick works by producing a physical reaction on cue ^[16].

Kyri Kotsoglou, a professor at Northumbria Law School who studies polygraph use in the justice system, said: "It's a misguided effort to reduce the complex to something that is tangible." ^[4] DCSA, Presage and Altec did not respond to the Review's requests for comment, and the DIU declined to comment ^[8].

For an operator weighing a scoring upgrade on any screening tool, two axes sort the decision. One is whether the input signal has been validated for the decision it will feed. The other is how many people with nothing to hide will pass through it. With a validated signal and a small pool, a better scorer is a plain improvement. With a validated signal and a large pool, it is worth buying if a human review sits between a flag and any action. An unvalidated signal and a small pool make a research project whose results should decide nothing. An unvalidated signal and a large pool mean a faster, more consistent model delivers the old error rate to more people. Polygraph+, scoped for applicant vetting and insider-threat work across the department, lands in that last box, on the signal the 1983 and 2003 reviews doubted ^[3][10].