Will you let your models go to war?
OpenAI once banned military use outright. It quietly dropped the ban in early 2024, and in February 2026 signed a classified deal to put its models inside the Pentagon's most secret networks. By Sam Altman's own words, the deal was "definitely rushed" and "the optics don't look good."¹
Anthropic didn't refuse the military either. It was ready to loosen its terms and sell — but it held two lines: its models must not be used for the mass surveillance of Americans, and must not fire weapons with no human in the loop. For those two lines, the Secretary of War branded Anthropic a "supply-chain risk" — the first US company ever handed a label reserved for foreign adversaries — and the administration moved to phase the company out across government.²
Read it again: A company drew the line at surveilling its own citizens and at autonomous killing, and was declared a national-security threat for it.
One federal judge, blocking the designation, called it "the Orwellian notion that an American company may be branded a potential adversary." An appeals court let it stand anyway.
I've written about three laws of AI delegation. They discipline the person who delegates: keep the consequence, use the simplest tool, keep a human in the irreversible loop. But all three assume the same question — who bears the consequence?
There's a law beneath them. A zeroth one. It asks something different: are there uses no one may authorize, no matter who consents?
An AI must not be used in a way that violates the dignity, safety, or self-determination of human beings — not even for their supposed good, not even for the convenience of those who operate it.
Mass surveillance. Autonomous weapons. Targeting engines. This is exactly where philosophy meets reality, be it as a term in a contract, or a line someone refuses to cross.
Let me be clear: this doesn't make Anthropic a saint. They were willing to arm the state — just not past those two lines — and they meter their tokens like everyone else. But they drew a line at all, and we just watched what drawing it costs. The other three laws govern how you delegate. The zeroth decides what must never be delegated at all — and right now it's worth watching who's still willing to pay for saying so.
Notes & Sources
- OpenAI dropped its blanket ban on military use in January 2024; classified agreement with the U.S. Department of Defense ("Department of War"), February 2026. Altman quotes ("definitely rushed"; "the optics don't look good") via TechCrunch, 1 Mar 2026; OpenAI statement.
- Pentagon designated Anthropic a "supply-chain risk" (Feb–Mar 2026) over two refused exceptions — mass surveillance of Americans and fully autonomous weapons; first U.S. company so labeled. Axios, 16 Feb 2026. The legal fight is split and ongoing: a federal judge in San Francisco (Rita F. Lin) granted Anthropic a preliminary injunction against the designation — calling it "the Orwellian notion that an American company may be branded a potential adversary" — while a D.C. appeals court declined to stay it (CNBC · SiliconANGLE, 8 Apr 2026).