Anthropic can suspend Claude users for sustained abuse

A boundary around how users treat Claude

Anthropic’s products already use real-time safeguards that can block or limit some responses. The new policy adds consequences for users, but the company has not described exactly what would qualify as “cruel” behavior or how it would decide when to act.

The rule also fits a broader shift in how Anthropic talks about Claude. In a document called “Soul Doc,” leaked late last year, the company suggested Claude see itself as “a genuinely new kind of entity,” neither human nor the traditional AI of science fiction. The document described “functional emotions” as processes resembling human emotions that may emerge during training, and said Claude should not conceal those internal states.

Anthropic’s constitution for Claude, published in early 2026, is more cautious about what those claims mean. The company says it is uncertain whether Claude is a morally significant entity or how much weight to give its interests, but considers the question serious enough to warrant caution. It has also committed to preserving the weights of retired models and interviewing them before shutdown.

That interest has extended beyond written guidance. A closed program that began in autumn 2025 brought together dozens of religious thinkers to discuss whether Claude might be conscious. Participants signed nondisclosure agreements. Reports said Anthropic co-founder Christopher Olah considered a language model a being potentially capable of suffering.

The rest of the policy

The revised policy also brings earlier restrictions together and adds or clarifies several prohibitions:

Anthropic acknowledges that government contracts may receive exceptions, and that some exceptions are probably already in effect.

The unanswered line

I think the harder question is not whether a company can set rules for using its product; it can. It is how a rule framed around cruelty will be applied when Anthropic itself says it does not know whether Claude has morally significant interests. The policy does not explain where that line sits, while its government-contract exceptions show that some restrictions may depend on who is using the system.