Documents related to the $200 million contract between the U.S. Department of Defense and OpenAI revealed that the department requested AI models that would refuse as few commands as possible for military use. The document coded P00003 stipulates that “OpenAI Mission Models” should have minimum refusal rates in national security scenarios and be tested accordingly.
OpenAI spokesperson Nate Evans said the language appeared in the Pentagon’s earlier draft, was removed after an objection, and was not included in the agreement in force. The Pentagon also said that the current contract contains no such requirement. However, although the FOIA request called for the signed contract, the document was not marked as a draft; a Justice Department attorney first confirmed that it was signed and in force, then withdrew that confirmation.
OpenAI, Google, Anthropic and xAI began developing military AI prototypes with the Pentagon in 2025. OpenAI’s February agreement prohibits mass domestic surveillance, autonomous weapons and high-risk decisions without human approval. It was stated that Pentagon systems would operate through the cloud and that safety mechanisms would be preserved.
Why it matters
The approach to reducing the refusal rate in the document highlights the question of how to draw the line between operational flexibility and preserving safety mechanisms in military AI. This issue concerns not only OpenAI, but also other companies developing similar prototypes with the Pentagon, as well as the public monitoring the conditions under which these systems are used. The lack of clarity over whether a statement reportedly included in the agreement was a draft or a binding text makes document management and accountability problems in public-sector AI procurement visible. The central question that remains open is who decides which commands will be refused in systems developed for military use and for which high-risk decisions human approval will remain mandatory—and on the basis of which text.
Background
OpenAI is not a new name in the FikirPilot archive: in the past 90 days, we have published 36 reports mentioning it; the latest is dated September 13, 2026.