The Pentagon wants an AI lie detector, despite decades of weak evidence

The Pentagon is seeking $30.3 million over five years for “Polygraph+”, an AI and machine-learning project intended to modernise lie detection for employee vetting and insider-threat monitoring. The important point is not that a machine has solved dishonesty. The evidence supplied by MIT Technology Review suggests the proposed system is trying to automate a problem that researchers have struggled to validate for a century.

Watch Desk analysis

What happened

According to MIT Technology Review, the Department of Defense budget request describes a programme run by the Defense Counterintelligence and Security Agency. It would explore scoring algorithms, machine learning and “standoff sensing”, meaning physiological measurements taken without attaching equipment to the person being tested.

The specific technology has not been settled, and the budget has not yet been approved by Congress. The DCSA did not respond to a request for more information. Possible ingredients include camera-based measurements of breathing, heart rate, facial temperature and movement, based on prototypes previously explored by the Pentagon’s innovation unit.

Why it matters

The attraction is obvious: a system that could process several signals at once, perhaps without the familiar wires and sensors, sounds more modern and more objective. But the core difficulty remains stubbornly analogue. Stress, cognitive load and attempts to hide information are not unique fingerprints of lying.

MIT Technology Review cites the National Research Council’s 2003 conclusion that evidence for polygraph screening was “weak at best”. The supplied report also notes that earlier efforts involving thermal cameras, pupil tracking, voice analysis and brain scans have not produced reliable results outside the laboratory. An AI score does not magically create ground truth where the underlying tests cannot establish who was actually telling the truth.

That matters because the proposed system is aimed at employment screening and insider-threat detection, where a wrong suspicion can affect a person’s career and reputation. At the scale of the federal workforce, even a system that sounds broadly accurate could generate a great many false accusations.

Our read

This is a serious computing and public-sector technology story, but it is not a breakthrough story. It is a large proposed investment in the hope that better sensors and more elaborate models can rescue a measurement problem that has repeatedly resisted rescue.

The eyebrow-raising bit is that “AI-powered” may make an uncertain judgement look more scientific rather than make it more reliable. Any deployment should require transparent validation against independently established outcomes, published error rates and meaningful safeguards for people who are flagged. Until then, Polygraph+ is a research proposal with a very consequential use case, not a lie detector anyone should trust by default.

What to watch

• Whether Congress approves the $30.3 million request.
• Which sensing methods and machine-learning models the DCSA actually selects.
• Whether the programme publishes independent validation and subgroup error rates.
• Whether AI-generated deception scores are used for decisions, investigations or merely as interview prompts.

Discussion spark: Should governments be allowed to use AI-generated deception scores in employee screening before the technology has demonstrated reliable accuracy outside controlled laboratory settings?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.