FOIA Draft Said Pentagon Sought OpenAI Models With 'Minimal Refusal Rates' — Both Sides Say Final Contract Did Not

Documents released via a FOIA lawsuit suggested the Pentagon once sought tailored OpenAI models for defense work that would register "minimal refusal rates" for certain military requests. OpenAI and Department of Defense officials later disputed that the phrase appears in the final, executed agreement.

Records obtained by The Intercept described contract language for testing, evaluation and improvement of specialized AI models intended for national security tasks. The draft language referred to mission-focused systems designed to decline fewer queries during particular operations — a formulation critics said could weaken safety guardrails.

What the Draft Said and How Parties Responded

The released draft included wording about models with "minimal refusal rates" for some defense applications. OpenAI said that phrasing came from an earlier Department of Defense draft that the company rejected.

"OpenAI has never agreed to contract language requiring 'minimal refusal rates.' This language does not appear in our executed contract," said Nate Evans, an OpenAI spokesperson, adding that the company provided feedback on the Department’s draft.

A Department of Defense representative likewise said the phrase does not appear in any active contract. The situation became more confusing after a Pentagon lawyer initially confirmed the document and later withdrew that confirmation. Trevor Tiedeman, a Pentagon research official, said there may have been confusion between records released under FOIA and the versions ultimately executed and that he would review how the earlier draft passed internal review before public release.

How the Draft Was Released

The Intercept obtained the paperwork after filing a Freedom of Information Act request seeking completed contract records tied to military AI agreements. Officials later argued the released files corresponded to an earlier draft rather than a final, signed contract. OpenAI has not published the complete contract publicly, leaving some details unavailable for independent inspection.

Google logo on a black background next to text reading 'Click to follow TechRadar'

Context and Concerns

Experts warn the exchange highlights ongoing tensions about how much control private tech firms retain when their AI is used by governments. Commercial AI products typically include safeguards to prevent assistance with harmful activities — for example, OpenAI’s public chatbot declines requests that would plan attacks or select military targets.

"'Minimal refusal' is likely referring to little or no safeguards on the model being used," said Heidy Khlaaf, chief scientist at the AI Now Institute. She said private firms face difficult choices when their technologies are integrated into military operations and emphasized the need for transparency and oversight.

What Else We Know

The Pentagon has had similar disagreements with other AI vendors over restrictions for defense uses. Despite the dispute over the draft, OpenAI later expanded its agreement to allow use of its services across classified government networks; however, many portions of the publicly released agreement remain redacted.

OpenAI has said the contract includes explicit prohibitions on autonomous weapons use and on certain surveillance activities involving Americans, but the company and government agencies have not fully explained how those restrictions would be applied across different military scenarios.

Why It Matters: The episode underscores the need for clearer public disclosure about the limits, safeguards and oversight mechanisms that govern commercial AI when deployed for national defense — especially when parts of agreements remain confidential or redacted.

Credit: Getty Images/VCG — Reporting based on documents published by The Intercept.