Sam Altman has set an ambitious standard with his Pentagon deal, claiming contractual protections against mass surveillance and autonomous weapons use in a government AI agreement. If that standard holds, it could establish a new baseline for how AI is deployed in national security contexts. If it erodes, the deal will become a cautionary tale of a different kind from Anthropic’s — one about the gap between stated principles and practiced ones.
Anthropic had tried to set a similar standard and was expelled from the government market for the attempt. The company’s refusal to remove two ethical conditions — no autonomous weapons, no mass surveillance — from its Pentagon negotiations was treated by the Trump administration not as responsible governance but as political defiance. The resulting ban was swift, comprehensive, and deliberately public.
The public nature of Anthropic’s punishment was clearly intended to shape behavior across the industry. By framing ethical conditions as ideological obstruction and enforcing that framing with immediate commercial consequences, the administration sent a message that AI companies entering the government market should expect to check their ethics policies at the door or face serious repercussions.
OpenAI’s Sam Altman responded by announcing a deal that he said does not require that check. His claims — that the contract prohibits mass surveillance and autonomous weapons use, that these are OpenAI’s own red lines, and that the Pentagon should offer these terms universally — set a high and specific standard against which the deal can be measured over time by workers, competitors, and the public.
The industry is watching. Hundreds of AI workers who publicly backed Anthropic remain employed by the companies signing these deals, and their expectations of principled behavior are clearly high. Anthropic, stripped of its government contracts, has nonetheless articulated a position that commands widespread respect: its restrictions are narrow, lawful, and have never prevented a legitimate mission. Whether OpenAI can say the same in a year’s time is the question that matters most.