The Pentagon wants $30 million for an AI lie detector: what it means

The US Department of Defense is asking Congress for $30.3 million over five years for “Polygraph+” — a next-generation lie detector that uses artificial intelligence. MIT Technology Review reported it on 25 September 2026. Here is a practitioner’s reading: what exactly is planned, why researchers are sceptical, and what is already available to an ordinary client today.

Need an actual test rather than theory? Take the VerifEye online test — from 2,500 UAH (≈ €50), with no office visit, or calculate the cost of a polygraph test.

Update, 6 October 2026. On 5 October, Military Times reported that DCSA said in a written statement that current research under the “Modernizing Polygraph” effort includes no generative AI, no large language models, and no analysis of facial expressions, micro-expressions or voice. As recently as July, Defense Department officials had described such tools as part of Polygraph+. The project itself remains: budget documents price it at $31 million, covering scoring algorithms and non-contact sensors, including thermal imaging. The department says it will keep monitoring face and voice analysis technologies. The scientists quoted in the report say that neither people nor software can reliably detect a lie from a face or a voice.

Update, 8 October 2026. A fresh example of how much a polygraph weighs in the US clearance system. On 7 October, Reuters, citing two people familiar with the matter, reported that Justice Department lawyer Kurt Olsen lost his security clearance in September after failing a polygraph test. It is not known whether other factors also contributed. Without the clearance it became difficult for him to work on an investigation involving a large amount of classified material. Olsen himself said he holds an active clearance from another agency and did not answer questions about the polygraph. Reuters could not establish what he was asked or why he failed. The agency notes that a failed polygraph does not prove guilt and that its results are generally not admissible in US courts.

What the Pentagon plans

According to the budget request, Polygraph+ (also called Polygraph Next) has two directions:

• scoring algorithms based on AI and machine learning — the result is calculated by software, not only by a person;
• contactless “standoff sensing” — physiological readings are taken without sensors attached to the body.

The programme is to be run by the DCSA, the agency that vets candidates and staff for the US federal government. The stated purpose is personnel vetting and insider-threat detection. Note that Congress has not approved the budget yet, and it is not known which technologies will be used.

An earlier selection in 2023 gives a hint: the Defense Innovation Unit chose two companies. One measures heart and breathing rate with an ordinary camera; the other tracks head movement and facial skin temperature.

Why researchers are sceptical

The MIT Technology Review article is critical, and that should be retold honestly:

• back in 2003 the US National Research Council called the evidence for the polygraph in mass screening weak;
• there is no single sign of lying that is the same for everyone. As deception researcher Sophie van der Zee puts it, “There is still no Pinocchio’s nose”;
• AI can only be trained on data where it is known who really told the truth, and real-life tests often have no such ground truth;
• similar projects — iBorderCtrl in the EU, AVATAR in the US — have quietly faded away.

I do not argue with that: no method is 100% accurate. That is why I rely on published figures. The accuracy of the classic polygraph is 85–89% according to the American Polygraph Association meta-analysis (2011); EyeDetect is 86–90% and VerifEye 84–89% according to the developer (Converus).

The key idea: lies are approached from three sides

The most useful part of the article explains what any lie detection tries to capture:

1. physiological stress — breathing, pulse, skin response;
2. cognitive load — lying is harder than telling the truth, and the brain works differently;
3. conscious attempts to conceal — controlling behaviour, countermeasures.

The classic polygraph works with the first one. The more independent channels are combined, the harder a test is to beat — that is the point of the “multi-modal” systems the money is requested for.

What already exists today

The Pentagon’s system does not exist yet. But both of its principles — algorithmic scoring and working without sensors on the body — are already used in the technologies I work with in Lviv:

• EyeDetect — an eye-based lie detector. A camera records pupil reactions and eye movements while you read statements, which is exactly cognitive load. No sensors on the body, the result is calculated by an algorithm, and the test takes up to 45 minutes including the result. It is taken at the station in Lviv;
• VerifEye — an online test through a smartphone camera, 10–25 minutes, with a PDF report. Suitable if you are in another city or country;
• the classic polygraph — physiological reactions, 1.5–3 hours, a written conclusion on the day of the test.

In complex cases I combine the polygraph and EyeDetect: two different channels out of the three that researchers describe. Beating both at once is much harder than beating one.

Will AI replace the polygraph examiner?

No — and the article indirectly confirms it. An algorithm scores well, but it does not word the questions, conduct the pre-test interview or notice that a person is nervous for an entirely different reason. A badly worded question will spoil the result of even the smartest system. AI is a tool for the examiner, not a replacement.

The market is growing

Analysts at Business Research Insights estimate the global polygraph equipment market at about $1 billion in 2026 and up to $1.81 billion by 2035. The main drivers are personnel screening and interest in AI-based systems. It is an estimate from a commercial report, so treat it as a rough guide.

Short answers

Does an AI lie detector already exist?

The Pentagon’s Polygraph+ does not — funding has only been requested. But tests in which an algorithm calculates the result do exist: EyeDetect and VerifEye.

Can a person be tested by camera without knowing it?

No. We conduct any test only with the person’s voluntary written consent.

Which method is more accurate?

The published ranges are close: 85–89% for the polygraph and 86–90% for EyeDetect. The choice depends on the situation, and the most reliable option is to combine two methods.

By polygraph examiner Andriy Maryniak, in practice since 2012. Based on the MIT Technology Review publication of 25 September 2026; as of October 2026.

More on the topic: the future of the polygraph — AI and biometrics
Prices: price list