Anthropic PBC v. U.S. Department of War
Issue
In *Anthropic PBC v. U.S. Department of War*, a coalition of five civil liberties and technology-sector organizations argues that the federal government violated the First Amendment on two independent grounds: by conditioning Anthropic's government contracting relationship on modifications to the safety architecture of its AI model Claude, and by designating Anthropic as a supply chain risk under 10 U.S.C. § 3252 in retaliation for the company's publicly stated values and AI policy positions. The case raises a question courts have not yet squarely resolved — whether the design and training choices embedded in a large language model constitute protected expressive conduct, such that a government mandate to alter them triggers compelled-speech doctrine. The added wrinkle is that senior officials' own public statements explicitly cited Anthropic's "ideology" as the basis for the designation, making the retaliatory motive unusually direct.
What Happened
This is an amicus curiae opening brief filed June 17, 2026, as Document 202 in Case No. 3:26-cv-01996-RFL, in the Northern District of California before Judge Rita F. Lin, by the Foundation for Individual Rights and Expression, the Electronic Frontier Foundation, the Cato Institute, Chamber of Progress, and the First Amendment Lawyers Association, through counsel at Perkins Coie LLP. The brief supports Anthropic's pending motion for summary judgment and advances three core arguments. First, amici contend that because Anthropic's engineers encode values into Claude through training, compelling changes to that architecture is no different from dictating newspaper editorial policy, invoking *Moody v. NetChoice* and *NetChoice v. Bonta* to extend platform-curation precedents to generative AI. Second, they argue that retaliatory motive is established by the government's own words — Secretary Hegseth's announcement cited Anthropic's "ideology," an Under Secretary's memorandum faulted the company for publicly criticizing the agency, and the President publicly labeled the company "RADICAL" and "WOKE" — which they contend satisfies the but-for causation standard under *Hartman v. Moore*. Third, they argue that severing Anthropic's contracting relationships and triggering blacklisting operates as a functional speech ban under *Backpage.com v. Dart*, warranting injunctive relief without proof of a formal content restriction. The brief also urges the Court to account for the structural chilling effect on the broader AI sector if agencies may use supply chain designation authority to penalize companies for their public policy disagreements with the government.
Why It Matters
The federal government has increasingly used procurement leverage and supply chain designation authority as tools of industrial and ideological policy, and this brief argues that when a company's publicly stated values — rather than a genuine security threat — drive such a designation, the First Amendment is triggered and the government loses. If the court accepts the amici's framing, it would be the first ruling to squarely hold that AI model safety architecture is constitutionally protected expression, creating a significant structural limit on the government's ability to use contract conditions to compel AI companies to build less cautious or differently aligned systems. For non-specialists, the core question is whether the Constitution forbids the government from telling an AI company "change how your AI thinks or lose all your federal contracts." The most closely watched doctrinal uncertainty is whether *Moody* and *NetChoice* — both of which addressed platforms *curating* pre-existing third-party content — extend to a model that *generates* outputs through training, a distinction no circuit court has yet resolved. The outcome could also determine whether retaliation doctrine places meaningful limits on the executive branch's use of national security procurement tools as instruments of pressure against private technology companies.
Related Filings
Other proceedings in the same litigation tracked by this monitor.
How accurate was this summary?