Pentagon seeks $30.3 million for AI-powered lie detector program

The US Department of Defense (DoD) has requested $30.3 million over five years to develop an AI-powered lie detector program called Polygraph+ or Polygraph Next, according to a DoD budget request. The program aims to modernize federal polygraph and credibility assessment technologies using AI and machine learning, as well as a technique called 'standoff sensing,' which allows physiological readings without physical contact.

The project will be managed by the Defense Counterintelligence and Security Agency (DCSA), which conducts background checks for federal employees. Its stated purpose is to vet prospective employees and detect insider threats, though specific technologies have not yet been disclosed. The budget request has not yet been approved by Congress.

The move follows increased use of polygraph tests within the Pentagon under Defense Secretary Pete Hegseth, including tests administered to around 50 Joint Staff officers after reports of leaks regarding US weapons stockpiles in the war with Iran.

Current polygraph technology relies on measuring blood pressure, pulse, breathing, and sweat to assess physiological responses to baseline and target questions. However, its reliability has been repeatedly challenged, with the US National Research Council in 2003 calling evidence for its efficacy 'weak at best.' The American Polygraph Association claims 80–94% accuracy, but experts note that even high accuracy rates could result in significant errors when applied at scale to the DoD’s 2.8 million employees.

AI integration aims to improve polygraphs by identifying patterns in physiological data that examiners cannot, potentially combining multiple measurements into a deception score. However, critics argue that AI cannot reliably link detected patterns to lying due to the lack of ground truth in polygraph data. Experts also highlight issues such as subjective interpretations, higher false positives among minority groups, and the ability of interviewees to learn countermeasures to manipulate results.

Previous attempts at AI-enabled lie detection, such as the EU’s iBorderCtrl and the US AVATAR project, have not yielded reliable results outside lab settings. Projects like Silent Talker, which generated deception scores from video footage, have been discontinued.