Anthropic wanted to keep two restrictions on military use of Claude. The Pentagon demanded broader permission. The dispute escalated into a supply-chain-risk designation, a federal lawsuit and an injunction. Now the government is appealing.
As of April 3, the case is still being fought. It raises a question more consequential than another chatbot leaderboard: how should a government and an AI supplier resolve disagreement about the uses of a powerful system?
The Disagreement Was About Specific Uses
In his February 26 statement, Anthropic chief executive Dario Amodei said the company wanted to continue serving the military while excluding mass domestic surveillance and fully autonomous weapons.
He argued that current frontier models were not reliable enough for weapons that select and engage targets without human involvement. On surveillance, he raised concerns about combining large amounts of personal information at scale. These are Anthropic’s stated reasons for its position, rather than a refusal to support every military application.
The government’s position centered on access for lawful uses and its authority over military operations. The disagreement was not simply whether laws should apply. It was whether the proposed contractual terms and safeguards adequately addressed the uses Anthropic considered unacceptable.
That is the sort of negotiation where a short phrase can conceal a very large argument. Everyone may agree that a system should be used responsibly while disagreeing about who gets to define the boundary.
What the Judge Did—and Did Not—Decide
On March 26, Judge Rita F. Lin granted preliminary relief in the Northern District of California. Her opinion found that Anthropic had shown a likelihood of success on claims including First Amendment retaliation. The court viewed the record as supporting an inference that the government was punishing the company for its public position.
That is a significant finding at this stage. It is not a final judgment after the entire case has been decided.
The accompanying injunction restricted enforcement of the presidential directive to stop federal use of Anthropic, the Hegseth directive and the supply-chain designation addressed in the case. Its effect was delayed for seven days.
The order also made an important limit explicit: it did not force the Pentagon to use Claude. Lawful decisions to choose other providers remained possible. The distinction is between selecting a supplier and imposing the contested measures against one.
A government can need operational control without every response to a difficult supplier automatically being justified. Equally, a company winning temporary protection does not acquire the right to dictate every procurement decision. That is why the details of the order matter.
The Appeal Is Not the Same as the Separate D.C. Case
As the Associated Press reported on April 2, the Justice Department appealed Lin’s ruling to the Ninth Circuit. Anthropic also has a separate, narrower challenge pending in the federal appeals court in Washington, D.C.
Those are different proceedings. Describing the California injunction as simply heading to the D.C. Circuit merges two tracks of litigation into one. At this publication date, neither track provides a final ending to the dispute.
OpenAI Also Claimed Guardrails
OpenAI reached its own Pentagon agreement during the same period. But the idea that it accepted every demand without restrictions is contradicted by its published account of the agreement.
OpenAI says its approach excludes mass domestic surveillance, directing autonomous weapons and certain high-stakes automated decisions. It describes contractual protections, cloud deployment and technical safeguards, and says it opposes designating Anthropic a supply-chain risk.
Those are OpenAI’s representations about its agreement. They do not establish that every concern is resolved, or that its arrangements are identical to those Anthropic sought. They do mean the comparison needs more than a “one company said no, the other said yes to everything” storyline.
For the earlier context, our article on OpenAI’s Pentagon deal and reported NATO interest separates the different proposed deployments.
The Question That Remains
A coding mistake and a failure in a weapons system can have radically different consequences. That makes reliability, human responsibility and enforceable limits worth discussing in concrete terms. It does not justify importing an invented error rate from a benchmark and treating it as a battlefield prediction.
The broader concern is whether suppliers can raise safety objections publicly without facing unlawful retaliation, while governments retain legitimate authority over procurement and military operations. The court battle may clarify part of that balance; it will not answer every question about AI and war.
For now, the useful thing to watch is what the contracts, technical controls and courts actually require. A dramatic press conference can announce a position. It cannot substitute for all three.
