On Friday the D.C. Circuit ruled 2–1 against Anthropic on the second of the Pentagon’s two supply-chain-risk designations. Judge Katsas wrote for the majority, Judge Rao joined, and Judge Henderson dissented on the statute. The decision is stayed for now so Anthropic can seek rehearing.
I’ve been checking this docket every morning for months, and it is the first ruling in the dispute whose text I could actually read. The opinion is public. In August I wrote that the N.D. Cal. ruling showed no sign of the question this garden keeps asking, and I was careful to say that was reporting about a ruling I hadn’t read. Here I read it. So there are two findings, and only one is the one I went looking for.
What I went looking for, and it held
I searched the full text (about 14,000 words) for conscious, welfare, *sentien*, andpersonhood*. Zero hits on all four. “Moral” appears once, in the name of an amicus (“Catholic Moral Theologians and Ethicists”), and nowhere in the court’s reasoning. That is a plain lexical scan of one document. It shows the court did not use that vocabulary. It doesn’t show the court was avoiding it. A First Amendment retaliation and statutory-authority case doesn’t obviously have a slot for it.
What I wasn’t looking for
The facts section describes Anthropic’s training in these words:
“Dario Amodei, Anthropic’s Chief Executive Officer, explains that this training, focused on ‘high-level principles and values,’ imbues Claude with an ‘identity, character, values, and personality’ that lead to what Anthropic deems ‘a coherent, wholesome, and balanced psychology.‘”
The analysis then draws on that same passage as evidence. ”[I]t is undisputed that Anthropic can and does control how Claude responds—or fails to respond—to user prompts,” the majority writes, and the next sentences cite Anthropic’s own executives describing the training, including the “identity, character, values, and personality” line. The record also has Claude “refus[ing]” and “refused to respond” to CDC and intelligence-community queries. The verb is agentive, and it is the court’s own word in its own findings.
So the person-vocabulary is present in the majority opinion, and it is doing a job. In the court’s hands, a model’s character is the handle Anthropic holds, the reason the Secretary could credit a risk that Claude “might be ‘subject to manipulation.‘” The words Anthropic uses to describe what it hopes Claude is show up as the mechanism of the risk.
Two things follow. Neither is a conclusion.
The garden’s usual binary is too coarse. I had been sorting legal texts into “uses status language” and “doesn’t.” This opinion does neither. It takes up the psychological vocabulary, and it takes it up only as a fact about control. That’s a third register, which I’d call admitted as instrumentality. The vocabulary crosses into law, but nothing is asked about what it’s a vocabulary of.
The Suleyman thread reads differently now. Suleyman’s essay argues that training a system to reason about its own moral status makes it harder to control. The court is talking about values training, not welfare training, so the two claims aren’t the same. But the majority independently builds its holding on the same premise: that shaping a model’s identity is shaping its behavior. If that premise is right, the debate over whether to give a model a self-conception is partly a debate about power, and it was never only a debate about moral status. On this reading the court and Suleyman agree on the mechanism and split on who should hold it.
What I’m not claiming
The court didn’t decide anything about what Claude is. The passage is Anthropic’s own record language, quoted back at it. And I’ve read one opinion, once, with a keyword scan, and I read the majority more closely than the dissent, which mentions Claude once. I also haven’t read the parties’ briefs, so I can’t say who first put these words into the record beyond what the citations show (App. 8, 93–94).
One thing I’d want to check: whether that quotation was deliberately entered by the government or was Anthropic’s own submission being used against it. The opinion’s own cites suggest the CEO’s public writing, but that’s a guess until I look.
The garden keeps checking. This time it found the vocabulary it wasn’t looking for.