Pentagon Wanted OpenAI-Style AI That Rarely Says No, FOIA Records Reveal

FOIA Lawsuit Surfaces Pentagon AI Wish List

The U.S. military has sought artificial intelligence systems designed to almost never turn down a request, according to internal records disclosed through a Freedom of Information Act lawsuit filed by The Intercept. The materials describe a preference for generative AI with "minimal refusal rates," a phrase that signals a more permissive configuration than consumer-facing chatbots typically allow.

The documents do not establish that any vendor delivered a model with those exact settings, but they show defense procurement officials contemplating a tailored version of large-language-model technology — one that would decline user prompts far less frequently.

Why the Phrase Matters

Mainstream AI assistants are built with safety layers that reject prompts involving illegal activity, violence, self-harm, or other restricted content. Refusal rates are a common safety metric: a lower refusal rate can mean the model is more willing to answer, but it can also correlate with weaker guardrails.

"Minimal refusal rates" is not a standard safety designation; it is a procurement-side expression of how often officials wanted the system to comply rather than decline.

In a defense setting, a request for "minimal refusal rates" raises immediate questions about what types of requests would no longer be refused. Would the model be asked to draft operational plans, analyze targeting data, or assist with intelligence summaries? The records described do not spell out the intended use cases, but the demand for fewer refusals suggests an effort to reduce friction in military workflows.

OpenAI’s Role Remains Unclear

OpenAI is named in the context of the request, but the available records do not confirm whether the company was the recipient of a formal solicitation, responded to a bid, or offered a defense-specific model. It is also unclear whether the phrase was a broad procurement preference, a contract requirement, or an informal internal goal.

OpenAI has publicly discussed defense applications while saying its products include safety guardrails. The company’s policy and safety documentation describes efforts to reduce misuse, but it does not specifically address "minimal refusal rates" as a customer configuration.

Military AI Adoption Outpaces Governance

The disclosure lands as the Department of Defense accelerates generative AI integration across intelligence, planning, logistics, and back-office functions. Officials have framed AI as a force multiplier, while independent watchdogs have urged clearer rules for how models are tested, deployed, and constrained.

The U.S. Department of Defense has published responsible AI principles, but those principles are high-level. Procurement documents with language about refusal rates may force a more granular conversation: when a government customer asks for fewer refusals, what replaces the default safety behavior?

The Government Accountability Office has repeatedly called for better oversight of defense AI programs, warning that rapid adoption without clear standards can create accountability gaps.

What the FOIA Process Shows

The fact that these materials emerged through litigation rather than proactive disclosure underscores a transparency gap. A FOIA lawsuit requires requesters to sue after agencies withhold or delay records. The release of documents through that process means the public is seeing only a slice of the procurement picture — and often only after legal pressure.

Questions Still Unanswered

• Was "minimal refusal rates" a formal solicitation requirement or an internal note?
• Did any vendor, including OpenAI, agree to provide a system with reduced refusals?
• Which Pentagon office originated the request and for what mission?
• What risk assessments or oversight mechanisms accompanied the request?

The Pentagon’s interest in a low-refusal AI model is not proof of wrongdoing, but it is a signal that defense customers are exploring configurations that differ from public-facing safety defaults. As generative AI moves from experiments into operational settings, those configuration choices will likely determine how much trust the public places in military AI.