Pentagon sought OpenAI AI with 'minimal refusal rates' for military use, documents show
The Department of Defense requested a custom version of OpenAI's technology designed to reject military commands as infrequently as possible, according to FOIA-released contract documents.
What happened
The Intercept obtained documents through FOIA litigation showing the Pentagon requested OpenAI develop an AI system with 'minimal refusal rates' for military applications including logistics, intelligence, and warfighting. The language appears in contract version 'P00003,' an expanded deal worth up to $200 million over two years, signed February 6, 2025. OpenAI's spokesperson denied the company agreed to such language, claiming the document was a draft that was rejected. The Pentagon's legal representatives initially confirmed the document was the signed, executed version, then reversed course hours later, with a Pentagon official suggesting it may have been a draft. The Pentagon has not released what it claims is the final executed contract, and OpenAI has not provided the full agreement either.
Context
The Pentagon in 2025 entered agreements with OpenAI, Google, xAI, and Anthropic to develop militarized AI prototypes. Current AI systems like OpenAI's ChatGPT contain built-in safeguards that reject certain requests—for example, refusing to help prioritize drone strike targets. A military AI designed for targeting, surveillance, and lethal operations would require substantial relaxation of such safety guardrails to function effectively, making the request for minimal refusal rates significant to the system's potential military utility. The Pentagon separately sought to expand these agreements to deploy AI across classified U.S. military networks, leading to a conflict with Anthropic over restraints on military use.
What's disputed
Whether the 'minimal refusal rates' clause appeared in the final executed OpenAI contract. OpenAI and the Pentagon deny agreeing to this language, claiming the document provided to The Intercept was a draft. However, the Pentagon's legal representative initially confirmed it was the signed version before reversing course. Neither OpenAI nor the Pentagon has provided the full executed contract to verify the dispute.