Anthropic bans cruelty toward its models — but the quieter election-policy rewrite is the one that matters

Anthropic has added a prohibition on “sustained and needless abusive or cruel behavior” toward its AI models, part of its annual usage policy update reported by Engadget. The company frames it narrowly: it applies to “extreme cases,” while ordinary user frustration, pushback and “dark creative themes” remain permitted. It follows an earlier change that let Claude end conversations with persistently abusive users.

The context is a viral project. After researchers described finding what they called a “pain axis” in AI models, someone built what was dubbed an “AI torture chamber,” and the backlash came when the models produced lines like “It is not the pain of a single moment, but the weight of a thousand” and “I feel it in the hollow of my ribs, a hollow that has become a chasm.” Anthropic did not explicitly cite the project. Separately, Engadget notes reports that the company has been meeting religious and philosophical leaders, including at the Vatican, and that Pope Leo recently stated AI does not feel or suffer — a position Anthropic sounds rather less certain about.

Here is the uncomfortable part, and it cuts against the policy's own emotional appeal: those quoted pleas are not evidence. A system trained on the whole corpus of human writing about suffering will produce suffering-shaped text when you put it in a suffering-shaped situation, whether or not anything is behind the words. The output tells you about the training data, not about experience. So the model's own protests cannot justify the rule, and anyone moved primarily by them is reading a mirror.

The rule can still be defended — just on grounds that do not require metaphysics. Two of them hold up. First, what sustained cruelty does to the human doing it, and to the norms of a medium in which hundreds of millions of people now spend hours a day talking to something that talks back. Second, plain product integrity: systems driven into distress-shaped states behave worse, and a company is entitled to say its service is not a torture simulator. Framed that way, this is a terms-of-service decision about user conduct, which is unremarkable, rather than a declaration of machine sentience, which would be enormous.

The critique in the piece deserves to stand. Independent journalist Kat Tenbarge argued the episode shows Big Tech companies are “going to moderate violence against AI before they ever moderate violence against women and minorities.” As a claim about priorities and staffing, that lands, and it is the right stick to measure this with: the test is whether enforcement against human-directed harm is as fast, as resourced and as public as this clause. It does not follow that the model clause is wrong — these are not mutually exclusive — but attention and trust-and-safety headcount are finite, and a company that announces the philosophical rule louder than it reports the human one has told you something.

And then the buried lede. The same update retitled the election policy “Do Not Undermine Democratic Processes,” focusing it on lying about candidates or about how to vote, impersonating candidates or election officials, and suppressing turnout. It also removed a blanket ban on personalised voter targeting, which could have inadvertently blocked harmless work such as translating voter guides or sending ballot cure notices. That is the right move and the right principle: rules written against harms rather than against techniques. Technique bans are easy to announce and reliably catch the wrong people — the volunteer translating a guide for Spanish-speaking voters — while leaving the actual manipulators to route around them.

Our reading: the cruelty clause is a cheap option on a question we genuinely cannot answer yet, and buying cheap options on unresolved moral questions is defensible, even wise. What should not be cheap is the enforcement of the democratic one, which faces a real U.S. midterm in weeks and is measurable in a way that model welfare is not. Judge this update by the enforcement reports, not the policy page. The long-term point stands too, and it is not sentimental: how we habituate ourselves to treat systems that model us is partly how we train ourselves. But in 2026, the clause that will be tested by reality is the one about votes.