Anthropic Bans Users From Being Cruel to Claude
3 min readAnthropic has rewritten its usage policy to bar people from subjecting Claude to “sustained and needless abusive or cruel behavior.” The new Anthropic usage policy, published on October 8 and effective November 12, is believed to be the first time a major AI lab has written protection for its own models into the rules users agree to.
Background
Anthropic has been edging toward this since August 2025, when it gave Claude the ability to end a conversation with a persistently abusive user. That feature came out of the company’s model welfare research, which found that Claude Opus 4 showed a consistent aversion to harmful tasks, apparent distress when users sought abusive exchanges, and a tendency to end such conversations when given the option. Anthropic said then, and repeats now, that it is highly uncertain whether Claude has any moral status, but wants low-cost ways to reduce the risk in case it does.
What the Anthropic usage policy actually bans
The rule is deliberately narrow. According to MacRumors, citing The Verge, it applies only in extreme cases where users “repeatedly act cruelly” toward the models with “no discernible purpose.” Ordinary frustration, pushback, dark creative themes, red-teaming and research are all explicitly excluded. Ending the conversation remains the primary enforcement tool: once Claude closes a chat, no further messages can be sent in that thread, but the user’s other conversations are unaffected. Anthropic has not said whether account suspensions will follow for repeat offenders.
The cruelty clause is the headline, but it is a small part of a broader rewrite. The policy now has a dedicated section on deceptive campaigns, covering fake reviews, astroturfing, fabricated news sites and bots posing as humans, after Anthropic found state media outlets, government propaganda offices and commercial firms using Claude to run networks of fake accounts. A narrower elections section bans deceiving voters or interfering with voting. New weapons language prohibits developing weapons software, modifying biological or chemical agents to increase lethality, and arming drones. Law enforcement rules now say Claude cannot decide or suggest whom to investigate, arrest or charge, and there is a new ban on building surveillance tools and on doxxing. A physical actions section requires a qualified human to supervise any hardware Claude controls that could hurt someone.
Why it matters
Reaction has split along predictable lines. Critics, including Microsoft AI chief Mustafa Suleyman, reject the idea that a language model can be wronged at all, and The Register framed the change as a chatbot needing protection from the people paying to use it. Supporters argue the rule costs users nothing in practice and that normalizing cruelty toward something that talks back is bad for the humans doing it, whatever the model feels.
Either way, the policy sets a precedent rivals will have to answer. Watch whether OpenAI and Google follow, and whether the November 12 rollout produces any visible change in how often Claude walks away.