The US Department of Defense (DoD) requested a specialized AI version from OpenAI to minimize the refusal rate of military commands. This requirement was found in a document labeled 'P00003', which expanded upon a previous two-year prototype development agreement worth up to $200 million.
OpenAI, Google, xAI, and Anthropic agreed in 2025 to develop militarized prototypes of their advanced AI for logistics, intelligence decision-making, and routine operations. The DoD later sought to deploy these AI tools within classified computer networks, leading to a dispute with Anthropic over the limitations of military use.
OpenAI and the DoD have denied the existence of the 'lowest refusal rate' clause in final contracts. OpenAI spokesperson Nate Evans stated that the company rejected the phrasing and the DoD agreed to remove it. However, a US Department of Justice lawyer initially confirmed the document was a signed and effective contract before later retracting the statement.
