The Defense Department is reportedly seeking $30.3 million over five years to develop Polygraph+, also described as Polygraph Next, a proposed modernization of credibility-assessment technology that would explore AI and machine-learning scoring along with potentially contactless collection of physiological signals.
MIT Technology Review, which reported the request on September 25, said the proposed program would be run by the Defense Counterintelligence and Security Agency, or DCSA. The reported uses are vetting prospective employees and detecting insider threats. Congress had not approved the request at the time of the report, and the underlying budget document and a detailed Defense Department explanation of the program were not publicly available in the material reviewed for this article.
What the proposed upgrade would change
If pursued, Polygraph+ would represent a meaningful change in how a familiar screening tool is administered and interpreted. Conventional polygraphs use attached sensors to measure signals including blood pressure, pulse, breathing and perspiration. An examiner compares reactions to different questions and makes an assessment; the test does not directly measure whether a person is telling the truth.
The reported upgrade has two distinct elements. One is algorithmic scoring: software could analyze the collected signals rather than leaving all interpretation to a human examiner. The other is “standoff sensing,” a term for measuring physiological indicators without attaching devices to the person being assessed. MIT Technology Review reported that multimodal systems could seek patterns associated with stress, cognitive load or efforts to conceal information, rather than relying only on the measurements emphasized by conventional instruments.
That description does not establish that the system can reliably detect lies. It instead identifies a research and development direction: collecting more signals and applying machine learning to make sense of them. The central scientific question is whether those signals have a sufficiently dependable relationship to deception in the real-world settings where a screening result could affect employment or security-clearance decisions.
Longstanding questions about deception detection
That question has shadowed polygraphy for decades. MIT Technology Review reported that Congress’s Office of Technology Assessment found very limited evidence supporting polygraph use for employee screening in 1983. It also reported that the National Research Council concluded in 2003 that evidence for polygraph efficacy was “weak at best.” Those findings matter because an AI layer can change the consistency or scale of analysis, but it does not by itself solve the problem of establishing what a physiological pattern means.
Researchers working on newer audio-visual deception-detection systems continue to identify that generalization problem. A 2025 study in Knowledge-Based Systems described multimodal methods as potentially useful on public datasets, while cautioning that their ability to generalize across different scenarios remains insufficiently explored. That limitation is especially relevant to a proposed government vetting system, where the conditions of data collection, the people being assessed and the consequences of errors can differ sharply from a controlled research dataset.
The proposal also sits within a broader government interest in applying AI to sensor-rich screening systems. A 2025 Department of Homeland Security report on non-intrusive screening described AI as a way to make greater use of data from existing sensors and to reconsider what data should be measured. That work concerns screening for physical threats and contraband rather than judging truthfulness, so it is not evidence for Polygraph+ or for automated deception detection. But it illustrates why agencies may see AI as a route to extracting more information from signals that were previously difficult for people to interpret at scale.
Validation, governance and oversight
For Polygraph+, the distinction between gathering a signal and drawing a conclusion from it will be crucial. A camera may be able to estimate some physical features remotely under suitable conditions, but the program’s public case would need to show that its complete process—sensing, data processing, model scoring and human use of an output—works reliably for the intended task. The available reporting does not specify a model, sensor suite, performance target, evaluation protocol, vendor, deployment schedule or rules for human review.
Those gaps are consequential in a screening context. NIST’s current digital-identity guidance says organizations using AI or machine learning in identity systems should document their use, provide information on model training, data, updates and testing, and conduct privacy-risk assessments. The guidance addresses identity systems rather than polygraphs specifically, but it offers a useful benchmark for questions a future Polygraph+ program would need to answer: What information is collected? How is performance tested in conditions resembling actual use? How are errors tracked? And what protections apply to sensitive physiological data?
The Department of Defense has experimented with adjacent ideas before. MIT Technology Review reported that the Defense Innovation Unit selected Presage Technologies and Altec Research in a 2023 deception-detection prototype effort. It described those projects as possible clues to the broader direction of Defense Department research, not as confirmed components of Polygraph+. Treating them as the same program would overstate what is known.
What to watch next
For now, Polygraph+ is a proposed budget request, not a funded or deployed system. Its stated purpose, as reported, is to modernize credibility assessment and improve accuracy and reliability. That aim could justify carefully designed research. But because the proposed technology would be aimed at high-stakes federal screening, the key standard cannot simply be whether AI produces more elaborate scores or enables less obtrusive sensing. It must be whether the resulting decisions are valid, independently evaluated, appropriately limited and transparent enough for people affected by them to understand how consequential judgments are made.
The next reporting milestones are therefore as important as the funding figure: whether Congress authorizes the request; whether the Pentagon releases program documentation; and whether any eventual evaluation plan tests performance across realistic populations and operating conditions. Until then, the reported plan is best understood as an effort to apply AI to an enduring and disputed problem, not as evidence that technology has resolved it.
