The Pentagon seeks $30.3 million to develop Polygraph Next, an AI lie detector

What the Pentagon is requesting—and what the figure means

The US Department of Defense included Polygraph Next—also referred to as Polygraph+ in some reports—in its fiscal year 2027 budget justification for research, development, test and evaluation. The program appears in the budget justification for the Defense Counterintelligence and Security Agency (DCSA). The official document allocates $6.421 million to the project for that fiscal year and provides spending estimates for subsequent years.

The $30.3 million figure is the projected total cost over five years, according to MIT Technology Review’s reporting on the request. It is important to distinguish this multi-year horizon from the amount requested for a single fiscal year: the two figures cover different periods. The official justification confirms that the project exists and identifies the FY2027 allocation; the $30.3 million total appears in the reporting consulted.

What is described is a request within the budget process—not proof that all the funding has been approved or that a contract has been awarded. Nor does it show that a finished product exists or is ready for use. In an R&D budget justification, an allocation supports planned activities; by itself, it does not demonstrate that the intended technical results have been achieved.

What role would AI and sensors play?

The proposal combines two lines of work: developing scoring algorithms based on artificial intelligence and machine learning, and studying contactless physiological detection techniques. Broadly, the first line would aim to analyze data and produce a score; the second would seek to capture a person’s signals without the contact sensors associated with conventional polygraphs. The budget documentation identifies these goals, but the available information does not provide a complete specification of the equipment, its conditions of use or its decision thresholds.

The term “contactless” describes how signals might be captured, not what those signals mean. Physiological responses can vary for many reasons, and a score generated by an algorithm is not a direct observation of a lie. Interpreting a result would require knowing which signals are measured, how they are compared, which population has been studied and what independent criteria are used to determine whether an assessment is correct.

AI could be used to combine variables or apply classification rules systematically. But automating the analysis does not remove uncertainty in the input data, or turn a correlation into proof of deception. And without published details about the model, the data used and how it was evaluated, it is not possible to assess the system’s accuracy, potential biases or consistency across different contexts.

A development goal is not a demonstrated capability

The official justification describes planned activities, including validation studies. That wording matters: it indicates that evaluating the system is part of the work still to be done, not that conclusive independent validation has already taken place. The available sources do not provide evidence of independent tests of Polygraph Next that would support a claim that it identifies lies reliably, or results that would allow its performance to be compared with existing methods.

It is also important not to confuse studies of machine learning applied to traditional polygraphs with tests of this new program. A paper published in Scientific Reports develops a “second-opinion” tool for data from a conventional polygraph. That work is relevant to understanding research into models that could assist assessments, but it does not validate Polygraph Next, its contactless sensors or its use in real-world decisions.

Scientific uncertainty is not resolved by a score that merely looks more precise. Evaluating a detector would require knowing, among other things, how many false alarms it produces, how many cases it misses and how those errors vary with test conditions. An overall error rate would not be enough if it conceals differences between groups or if the system is used with a population different from the one in which it was developed.

Known limits of polygraph-based detection

The US National Research Council’s report on the polygraph and lie detection examined the available evidence, particularly in relation to security screening. Its conclusions point to limitations in that evidence. The report provides relevant context for assessing new proposals, but it is not an evaluation of Polygraph Next, which is a separate, later project.

The underlying problem is that measuring physiological responses is not the same as directly measuring whether a statement is truthful. Interpretation depends on how a test is designed, the circumstances in which it is conducted and the relationship between recorded signals and the conclusion someone seeks to draw. That is why any claim about a new system’s capabilities requires system-specific testing for its intended use.

For now, the sources provided support a description of a research proposal and some of its components. They do not establish what results it will achieve, whether it will pass independent evaluation or whether it will be adopted. They also do not specify the particular processes in which Polygraph Next would be used or what consequences it could have for the people assessed.

What safeguards would be needed before it is used?

If the project moved on to testing with people or operational use, safeguards should be defined before a score could influence a decision. A basic requirement would be to publish sufficiently detailed evaluation protocols so that independent reviewers can examine what is measured, how the result is calculated and what errors occur. Validation should be independent and conducted under conditions representative of the proposed use.

The tool should also be kept separate from the person making the decision. Human review is a meaningful safeguard only if the reviewer can challenge the result, understand its limitations and is not required to treat the score as conclusive. For decisions with serious consequences, it would be necessary to clarify which authority is accountable and what avenues of review or appeal would be available to the person assessed.

The available sources do not indicate that these safeguards are already part of a deployment, or describe a plan to use the system in employment, investigations or other individual processes. Those uses therefore cannot be attributed to the program as if they were confirmed. The immediate issue is narrower: the Pentagon is requesting funding to research a combination of algorithms and sensors; their validity and conditions of use remain unresolved.

To follow the project’s development, it helps to distinguish three milestones: budget approval, technical test results and any later decision to use the system. The budget documentation describes the first as a request and outlines planned work; the sources provided do not settle the other two. That distinction helps prevent an R&D initiative from being presented as a proven capability.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.