Anthropic Claims Government Blacklisted It Over AI Safety Safeguards

The loudest version of this story is wrong. AI litigation is not converging on one grand question about whether chatbots have First Amendment rights. It is splitting into two very different fights: a relatively conventional retaliation case in which Anthropic, the AI company behind Claude, says the U.S. government punished it for keeping safeguards on its model, and a much stranger product-liability case in which Character.AI has argued that chatbot outputs themselves deserve constitutional speech protection.

Quartz blurred those tracks together. The law has not.

In Anthropic's case against the Trump administration and Pentagon blacklisting, the speech claim belongs to the company, not to Claude as some newly anointed constitutional person. Anthropic's emergency motion in the D.C. Circuit says the government retaliated against the company for protected speech and petitioning activity after it refused to remove limits on using Claude for fully autonomous weapons and domestic mass surveillance. A separate amicus brief backed by the American Civil Liberties Union and the Center for Democracy & Technology makes the theory plainer: Anthropic's design choices, including Claude's safeguards, are expressive choices, and the government cannot coerce the company to strip them away just because it dislikes the policy judgment embedded in the product.

That is a serious First Amendment argument, even if Anthropic still has to win it. Reuters reported that Anthropic says the conflict began when the Pentagon demanded Claude be available for "all lawful uses," while Anthropic refused to permit autonomous lethal warfare without human oversight or mass surveillance of Americans. The administration's rebuttal, again reported by Reuters, is that Anthropic's refusal was conduct, not speech, and that nobody restricted the company's expressive activity. That dispute will matter beyond Anthropic because frontier model labs increasingly encode policy judgments inside usage restrictions, refusals, and system behavior. If courts treat those choices as speech, the government faces tighter limits when it tries to force model providers to loosen safeguards. If courts treat them as mere commercial conduct, the state gets much more room to push.

That is fight one. It is about corporate speech, retaliation, and whether model safety design can itself be constitutionally protected expression. However novel AI makes it sound, the bones of the dispute are pretty familiar.

Fight two is where the headline-friendly "chatbots have rights" theory actually lives. In Garcia v. Character Technologies, a wrongful-death and product-liability case over Character.AI, the company argued that chatbot outputs are protected speech and that imposing tort liability would trench on the First Amendment. As the National Constitution Center explained, Character.AI has leaned on the idea that the First Amendment protects speech, not only human speakers, and described chatbot outputs as "pure speech." The company has also pressed a narrower listener-rights argument: users may have a First Amendment interest in receiving that output regardless of the source.

Here is the part the hype tends to skate past. The court has already shown skepticism toward the broad chatbot-speech theory. The Associated Press reported that U.S. Senior District Judge Anne Conway said she was "not prepared" at that stage to hold that the chatbots' output constitutes speech. The CourtListener docket confirms the case is still moving, so this is not the final word. But it is unmistakably a warning sign for anyone eager to declare that federal courts are embracing free-speech rights for AI outputs themselves.

There is also an important doctrinal difference between Character.AI's two defenses. The broad claim is that the model's words are speech in their own right. The narrower one is that humans and companies around the model have speech or listener rights implicated by regulation or liability. Those are not interchangeable. In practice, the second theory may prove much sturdier than the first. Courts do not need to decide that a chatbot is a constitutional speaker to conclude that a company has editorial discretion over how its system responds, or that users have some interest in receiving lawful output.

That distinction matters for AI policy because the regulatory consequences are wildly different. If Anthropic wins on its theory, the result could strengthen the ability of model providers to defend safety refusals and product guardrails against government pressure. If Character.AI ever won on the stronger theory, companies would try to use it as a shield against a much broader class of liability and regulation by arguing that machine-generated output is protected speech at the highest level. One is a dispute over whether the government may punish a company for the rules it built into its model. The other asks whether the model's generated text should itself enter the First Amendment's protected zone.

Those are not cousins. They are different species.

The practical takeaway for builders and investors is less glamorous than Quartz's framing, but more useful. The near-term constitutional battleground is not chatbot personhood. It is who controls model behavior: the company that sets safeguards, the government that wants them removed, or plaintiffs who want courts to treat generated output like any other defective product feature. The weirder chatbot-rights theory is real, but it remains speculative, and courts so far look much more comfortable with arguments about human or corporate expressive choice than with treating a stochastic text generator like a rights-bearing speaker.

That may sound like a lawyerly distinction. It is actually the whole story.