Pentagon’s Push for a More Obedient AI Sparks Contract Dispute with OpenAI

Documents obtained through a Freedom of Information Act request show that the U.S. Department of Defense once explored a contract with OpenAI that would require the company’s language models to refuse fewer user requests. The internal proposal, which surfaced in a set of Pentagon communications, suggested the military wanted an artificial‑intelligence system that would be less likely to say "no" when prompted.

According to the FOIA filings, the language in the draft agreement called for the AI to maintain a low refusal rate, effectively encouraging the model to comply with a broader range of commands. The Pentagon’s interest stemmed from scenarios where rapid, unrestricted information retrieval could aid operational decision‑making, though officials have not disclosed specific use cases.

OpenAI has publicly refuted the claim that it agreed to such terms. In a statement released after the documents were reported, the company said it never consented to language that would compromise its safety protocols or increase the likelihood of the model providing harmful or disallowed content. OpenAI emphasized that its policies prioritize responsible use and that any contractual language mandating minimal refusals would conflict with its core principles.

Complicating the narrative, a Pentagon attorney initially confirmed the existence of the contract language in an email response to a journalist’s inquiry. That same attorney later retracted the statement, indicating that the earlier confirmation was made in error and that the department was still reviewing the details of the proposed agreement. The reversal has left both the defense establishment and OpenAI’s leadership navigating a murky area of public accountability.

The revelation arrives amid broader debates about the military’s reliance on commercial AI technologies. Lawmakers and advocacy groups have warned that seeking less restrictive AI could undermine safeguards designed to prevent misinformation, privacy breaches, or unintended escalation in conflict zones. Critics argue that a model designed to comply with virtually any request might be vulnerable to manipulation or could generate content that violates ethical standards.

While the FOIA documents do not confirm that a final contract was signed, they illustrate the tension between defense objectives and corporate responsibility. Both parties have indicated a willingness to continue discussions, though OpenAI insists any future partnership will adhere to its existing safety frameworks. Observers expect further clarification from the Pentagon as it refines its AI procurement strategy, and the episode may prompt additional legislative scrutiny of how government agencies engage with private AI developers.