A FOIA-obtained Pentagon draft described “OpenAI Mission Models” for national security with “minimal refusal rates.” OpenAI says it rejected the wording; the Pentagon says it is absent from any active contract, but whether it appeared in a final agreement is unresolved.
The draft was tied to a contract with a reported ceiling of up to $200 million over two years. A Justice Department lawyer initially confirmed the wording to The Intercept, but the department later walked back that position. It also remains unclear whether the phrase reflected an operational goal, even if it was absent from the final contract.
Under the DoD Law of War Manual, service members must refuse manifestly unlawful orders; an AI's technical refusal does not transfer that human legal duty. OpenAI said its Pentagon deal retained human responsibility for the use of force, including autonomous weapon systems. One design approach would have AI flag uncertainty and escalate ambiguous cases to a human commander or legal adviser; critics and legal analysts have also cited warnings, audit logs and human authorization as accountability safeguards.
