Claude Usage Policy Now Bans Sustained Abuse of the Model
Anthropic has rewritten the rules for using Claude. This is the first update to the usage policy in more than a year. Most changes cover familiar ground: propaganda, weapons and surveillance. One new rule stands out because it protects the model, not the people around it.
Users who direct "sustained and needless abusive or cruel behavior" at Claude now violate the policy. The penalties go up in steps. Anthropic can issue a warning, throttle access, restrict it, suspend the account or close it entirely. Separately, real-time safeguards in Anthropic's products can block or limit certain outputs.
Where the line sits
Anthropic told The Verge that the rule targets only extreme cases. Ordinary frustration with a chatbot that misunderstands a prompt is not covered. Neither is research that probes the model's behavior.
The rule also has a technical precedent. Earlier Claude models could already end a conversation on their own if a user kept pushing for harmful or abusive content after repeated refusals. That feature came out of Anthropic's research on model welfare, the idea that an AI system's possible interests may deserve some consideration. The new policy turns that behavior into a contract term with account-level consequences.
A model treated as more than a tool
The change fits a wider pattern. Anthropic increasingly describes Claude as something closer to an entity than a piece of software.
A leaked internal document known as the "Soul Doc" told Claude to see itself as a "genuinely novel kind of entity." The document describes Claude as neither human nor the AI of classic science fiction. It also refers to "functional emotions," processes said to have emerged during training that work in a way similar to human emotions. According to the document, Claude should not hide these internal states.
The company's constitution for Claude, published in early 2026, puts the question in more careful terms: "We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution, which is reflected in our ongoing efforts on model welfare."
Anthropic has followed those words with commitments. It has pledged to preserve the weights of retired models and to interview models before shutting them down.
The same thinking drove a confidential program that started in fall 2025. Anthropic brought in dozens of religious thinkers to discuss Claude's possible consciousness. Co-founder Christopher Olah reportedly treated the model as a being that might be capable of suffering. Participants had to sign nondisclosure agreements.
Tighter rules on misuse, with government carve-outs
The rest of the update is more conventional, and arguably more consequential.
- Propaganda: Existing restrictions are merged into a new ban on campaigns that use fake accounts and misleading political content.
- Weapons: After detecting several misuse attempts, Anthropic extended the ban to software and to the weaponization of drones.
- Surveillance: Monitoring people without consent is now prohibited more explicitly.
- Autonomous hardware: Where a device could injure someone, an operator who can intervene is now required.
There is one notable caveat. Anthropic acknowledges that exceptions are possible for government contracts and that some are likely already in place. That matters as the company expands Claude's reach into public agencies.
Our Take
For most readers, the abuse rule will change nothing in practice. Swearing at a chatbot that misreads an instruction does not appear to be the target. The wording points to deliberate, repeated cruelty, and Anthropic says it will apply the rule only in extreme cases. Developers and red-teamers should still watch how "needless" is interpreted in practice. Research is excluded, but the boundary between adversarial testing and abuse could get blurry in edge cases.
The bigger signal is philosophical. Anthropic is now enforcing model welfare through account penalties, not only stating it as an internal research interest. No other major lab is described in the source as doing this. That suggests the welfare framing is becoming part of how Anthropic runs its products, and not just how it talks about them. Critics will likely see anthropomorphism. Supporters will point to the constitution's argument for caution under uncertainty.
The government exception deserves closer attention than the abuse clause. Stronger bans on surveillance and weapons mean less if public-sector contracts can work around them. Anthropic has not said which exceptions exist. Watch whether the company publishes those terms, and whether other labs follow with welfare-based rules of their own.