Pentagon Sought OpenAI Military AI With Minimal Refusal Rate, Documents Show

Documents obtained through a Freedom of Information Act lawsuit show that the U.S. Department of Defense once sought an OpenAI artificial intelligence version customized for the military with a “minimal refusal rate” for instructions. According to ChainCatcher, the language appeared in an updated contract, P00003, for a 2025 prototype agreement lasting two years and worth up to $200 million, covering logistics, intelligence decision-making, and combat uses. OpenAI, Google, xAI, and Anthropic all agreed in 2025 to develop such military prototypes. OpenAI spokesperson Nate Evans said the company never agreed to a contract term involving a “minimal refusal rate,” and that the documents were early Defense Department drafts, not the final executed contract. Defense Department spokesperson Jacob Bliss also said the phrase does not appear in any current contract. The Intercept said none of the documents it obtained were marked as drafts. A Justice Department lawyer representing the Pentagon initially confirmed the file was a signed, executed version, then asked that the earlier confirmation be ignored and said further review was needed. The latest contract, signed on February 27, allows OpenAI services to be deployed on classified military networks, with similar delivery terms fully redacted. OpenAI said it has drawn red lines against autonomous killing and surveillance targeting Americans, but the contract text allows any use the government deems lawful.