Pentagon excludes 4 AI techniques from Modernizing Polygraph R&D

For AI researchers and ML engineers, the Pentagon's partial retreat is a high-stakes case study in deploying emotion and behavior detection where errors are unacceptable. By excluding generative AI, LLMs, facial action coding, and vocal analysis, DCSA is signaling that unvalidated models have no place in security clearance vetting—at least not publicly.

On Wednesday, September 30, 2026, the Pentagon's Defense Counterintelligence and Security Agency (DCSA) issued a statement narrowing the public scope of artificial intelligence work inside the Modernizing Polygraph program. DCSA, which vets security clearance holders and applicants, said that "active research and development" for the effort "does not incorporate generative super intelligence... or large language models" and does not involve "facial action coding [of emotions], micro-expression analysis or vocal analyses." The statement came after Defense News pressed DCSA about earlier declassified presentation slides and prior Department of Defense statements indicating that, as of July, the Pentagon had been preparing AI tools to analyze text, voices, and faces to estimate trustworthiness.

“

On Wednesday, September 30, 2026, the Pentagon's Defense Counterintelligence and Security Agency (DCSA) issued a statement narrowing the public scope of artificial intelligence work inside the Modernizing Polygraph program.

The statement is a partial, carefully worded retreat rather than a clean abandonment. It disavows high-profile generative models, large language models, and several contested behavioral biometrics. But it does not address other AI models that, according to the slides, researchers had been developing to scan spoken and written language for "deceptive speech" patterns. That distinction matters deeply: DCSA appears to be constraining the public scope of its AI polygraph work rather than ending it outright. The Modernizing Polygraph effort may still incorporate machine learning techniques that analyze language for deception, just not ChatGPT-style chatbots or facial and vocal sentiment analysis that have drawn the heaviest scientific and civil-libertarian criticism.

The criticism is substantial. National security lawyers, scientists, and AI auditors warned that AI-aided scanning of speech patterns, vocal stress, or facial expressions to "read minds" lacks scientific backing, will compound errors, and may violate civil liberties in a domain where few such rights exist. Deception detection is not a solved problem. The difference between lying, misremembering, stress, and strategic speech is often undetectable by any known biometric signature. In polygraph contexts, where false positives can destroy careers and security clearance processes are already opaque, a model's confidence score is not a legal verdict, but it may be treated as one.

The Pentagon's language suggests no irreversible decision has been made. DCSA said the Department of Defense will "continue to monitor emerging scientific capabilities to evaluate their potential utility down the road." This signals possible future adoption of AI, but only after scientific validation and further scrutiny. For defense contractors and AI firms, the public window for selling generative AI or facial micro-expression tools into Modernizing Polygraph has narrowed considerably. Yet the ambiguous status of deceptive-speech models means language-processing vendors may still have an opportunity, provided they can demonstrate validity under heightened scientific review.