Pentagon asked OpenAI for military AI with low refusal rate

Documents obtained by The Intercept through a Freedom of Information lawsuit show that the U.S. Department of Defense asked OpenAI to provide a customized AI tool designed to minimize rejections of military orders, aiming for a 'lowest refusal rate.' This request appeared in a draft contract numbered P00003, involving deploying AI services to classified military networks, with a contract value up to $200 million. Both OpenAI and the Pentagon denied that the final contract included this clause, calling the document a draft, but The Intercept noted the file wasn't marked as a draft, and a Pentagon lawyer initially confirmed it as the final version before backtracking.

Experts worry that such a low-refusal AI might lack safety guardrails, potentially assisting in military decisions, even target selection. Previously, OpenAI and Anthropic agreed to develop AI prototypes for the military, but safety restrictions sparked disputes, with Anthropic clashing with the Pentagon. OpenAI stated its final agreement prohibits AI use for autonomous killing and espionage against Americans, but the actual wording may leave room for interpretation.