As recently as July, the Pentagon was preparing artificial intelligence tools to analyze text, voices and faces while estimating the trustworthiness of personnel. Declassified slides and earlier War Department statements tied those technologies to an effort to modernize credibility assessments and polygraph interviews.

Now, the War Department is backing away from at least part of that picture. The Defense Counterintelligence and Security Agency told Defense News that active research for its “Modernizing Polygraph” effort does not include artificial intelligence that generates new content.

DCSA also said the current effort excludes large language models such as chatbots trained on massive data sets. Facial action coding, micro expression analysis and vocal analysis are not part of the active research either, according to the agency.

The agency left one significant question hanging. It did not address other AI models that earlier slides said researchers were developing to scan written and spoken language for patterns associated with deceptive speech.

The new statement contrasts with previous War Department descriptions, 2027 budget documents and DCSA slides obtained through an open records request. In July, an anonymous Pentagon official described AI driven sentiment and deception analysis as a tool for producing standardized, objective analysis in real time.

“AI-based analysis is a force multiplier for our human investigators, not a replacement,” the official said. The official also said the models were primarily trained with information from controlled laboratory studies involving volunteers.

DCSA presentations covering 2021 through 2025 showed the Air Force, DCSA and academic laboratories working on sentiment analysis and deceptive speech models since 2019. That research appeared under a project called Credibility Assessment Modernization, also described by the Pentagon official as Polygraph+.

Current budget materials price the Polygraph+ and Polygraph Next project at $31 million. Listed costs include AI scoring algorithms, decision aids, thermal imaging equipment designed to identify blood flow changes associated with stress, and other noncontact sensors.

One May 2025 slide depicted a computer generated interviewer questioning a security clearance applicant. Cameras and microphones supplied audio, video and speech to text data to large language models, while algorithms estimated the applicant’s emotional state and a human investigator selected questions.

DCSA said the university that produced that chart is not involved in the current research stage. The agency also said the Air Force is not participating in the active research, despite earlier slides describing Air Force work on deception related text and deep learning models.

Another DCSA slide set a goal of having natural language processing interpret an interviewee’s feelings with 75% accuracy. Critics questioned that target because current research suggests AI analysis of speech, voices or faces identifies lies at odds of roughly 50:50.

Michigan State University professor David Markowitz said reliable diagnostic cues for deception have not been established for either people or artificial intelligence. His experiment using Google Gemini found the model performed no better than chance and was more likely than humans to falsely accuse trivia players of lying.

National security attorney Mark Zaid warned that an inaccurate machine assessment could derail a military officer’s career. Personnel who fail a polygraph after previously passing can face delays affecting promotions, assignments and overseas work while waiting a year or longer for another examination.

The legal safeguards are also thin. The Supreme Court declared in 1998 that no consensus existed on polygraph reliability and barred polygraph results and polygraphers’ opinions from military courts, while civil courts routinely dismiss challenges involving revoked security clearances.

Researchers also raised concerns about disability bias and automation bias. University of Wisconsin professor Jo Ann Oravec said facial analysis can misread involuntary movements or vocal outbursts, while auditors warned that investigators may accept automated recommendations despite conflicting evidence.

Pentagon officials previously said automated findings would remain auxiliary information for trained polygraph examiners. Air Force developmental leader Farakh Zaman also said researchers used diverse training data, documented the origins and modification of data, and employed independent white hat hackers to seek false correlations.

DCSA has not explained the full scope of Modernizing Polygraph or why voice screening and facial expression analysis are absent from the current phase. After years of slides, budget language and enthusiastic talk about algorithmic assistance, the Pentagon’s latest answer leaves its AI loyalty testing ambitions buried in bureaucratic fog.