Pentagon's OpenAI contract sought "minimal refusal rates"

The Pentagon asked OpenAI for "minimal refusal rates" in its AI systems, according to contract language obtained by The Intercept. The outlet got the documents through a Freedom of Information Act lawsuit. OpenAI says that phrasing never made it into the signed version.

A refusal is a safety feature in these systems, not friction. The contracts fund prototypes for military decision-making, intelligence analysis, and operational planning, exactly the kind of software where a model declining a request is the point.

The documents cover four companies: OpenAI, Anthropic, Google, and xAI. Each holds a contract with a $200M ceiling. The scope goes past building software. The labs agreed to advise the Pentagon on AI strategy, train military personnel, and help forecast the risks of the technology they're selling to it. A vendor assessing the risk of its own product is grading its own homework, whatever the contract language says.

What the filing doesn't settle is how far the walk-back goes. OpenAI disputes that the minimal-refusal request reached its final contract, but The Intercept's reporting doesn't say whether Anthropic's, Google's, or xAI's agreements carry similar language, or whether any of the four have already delivered a prototype under these terms.