A federal judge has ruled that the Trump administration illegally punished Anthropic for publicly refusing to allow the military to use its Claude LLM for tasks including mass surveillance of Americans and lethal autonomous warfare.

Judge Rita Lin granted Anthropic summary judgment on its First Amendment, due process, and Administrative Procedure Act claims in a 59-page order. The government conceded that Anthropic has no backdoor access to deployed models and that its systems are no riskier than any other AI model. Anthropic lost on one count: its claim that Trump's directive to federal agencies exceeded presidential power.

The ruling is one of the clearest judicial statements to date that a company's public stance on how its AI may be used is protected speech, and that agencies cannot use procurement leverage to punish a vendor for holding usage limits. The decision comes as AI developers weigh how much control to retain over defense applications of their models, a tension that has grown as the Pentagon pushes for fewer usage restrictions in contracts.

How the dispute began

The conflict started when the Pentagon told Anthropic to strip all usage restrictions and accept a clause permitting "all lawful uses." Anthropic dropped most of the restrictions but held two red lines, refusing to lift its AI safeguards.

On February 27, Trump ordered every federal agency to stop using the company's technology, calling it a "RADICAL LEFT, WOKE COMPANY" on Truth Social. Defense Secretary Hegseth followed, accusing Anthropic in a tweet of "a master class in arrogance and betrayal" and barring any military contractor from doing business with it.

What the record showed

The government has since abandoned its central claim. Its risk assessment rested on the assumption that Anthropic had backdoor access to deployed models, and it now concedes that Anthropic has no such access and that Claude is no riskier than any other "black box" system.

That left a single factor: trust, forfeited by what a Pentagon memo described as Anthropic's "increasingly hostile manner through the press."

Lin found the timeline gave that away. Days before the blacklisting, Hegseth had floated invoking the Defense Production Act, which would have made Anthropic essential to national security. The day after the designation, an under secretary emailed the company to say a contract was "very close." The government is still discussing collaboration on Mythos, Anthropic's newest model, across sensitive contexts, Lin added.

"The empty invocation of national security is not a blank check to punish and retaliate against government critics," Lin wrote, finding the actions retaliatory under the First Amendment, imposed without the pre-deprivation process the Fifth Amendment requires, outside the supply chain statute, and arbitrary and capricious.

A partial win for the government

Anthropic did not prevail on every claim. Lin rejected its argument that Trump's directive exceeded presidential authority, and entered judgment for agencies that took no action.

She vacated the designation and Hegseth's boycott order, granted a permanent injunction, and denied the government's request for a seven-day administrative stay, noting it had operated under a prelinary injunction since March without identifying any harm.

Anthropic told the court that the measures, if left standing, would cut its defense-related revenue by 50% to 100% and its 2026 revenue by billions. That revenue exposure illustrates how much federal and defense contracts now matter to AI labs' business, even as courts have shown willingness to check how agencies wield that leverage.

The government could appeal the ruling, which would extend a case that has already drawn attention from other AI firms navigating similar contract negotiations with federal agencies.

Disclosure: This article was generated using Anthropic's Claude LLM.