Anthropic has added a rule to its usage policy that governs how people treat the AI itself. The company published the update on Thursday, its first revision of the policy in more than a year, and the new rules take effect on November 12. Most of the changes tighten restrictions on weapons, surveillance and deceptive campaigns. The one that drew the most attention is different: users may no longer subject Claude to “sustained and needless abusive or cruel behavior.”
What the new rule says
The wording is deliberately narrow. Anthropic says the rule is meant for extreme cases, where a user repeatedly acts cruelly toward its models for no apparent purpose. It does not cover the ordinary ways people lose patience with a chatbot. Frustration, pushback, dark creative themes, and testing or research on the models are all explicitly outside the rule.
In practice, that means snapping at Claude after a wrong answer, writing a grim short story, or deliberately trying to break the model for a security test should not put anyone at risk. What the rule targets is a long, pointless stream of abuse with no goal other than the abuse.
How it will be enforced
Anthropic says the main enforcement tool will be one it already built. In August 2025, the company let some Claude models end a conversation in rare, extreme cases of persistently harmful or abusive interactions. Reports on the new policy say that ability now exists on Claude.ai and in Claude Code, and that it will remain the primary way the rule is applied.
The wider usage policy also allows Anthropic to warn users, restrict them, suspend them or end their access, so an account could in theory be affected. The company has not said how, or whether, the cruelty rule would lead to those steps, and it has not explained how it will tell legitimate frustration from gratuitous abuse. When the conversation-ending feature launched, users could still start fresh chats from the same account and branch off an ended conversation by editing an earlier message.
Why Anthropic is doing this
The rule grows out of what the company calls model welfare research. When it introduced the conversation-ending feature, Anthropic was careful to say it was not claiming Claude is sentient. It described itself as highly uncertain about the moral status of Claude and other language models, now or in the future, and framed the feature as a low-cost precaution in case such welfare turns out to matter.
The company’s leaders have spoken in similar terms. Kyle Fish, who leads Anthropic’s model welfare research, has said that questions about possible internal experience, consciousness and moral status are serious ones the company is investigating as models become more capable. CEO Dario Amodei said in February that he is unsure whether AI models could be conscious. Anthropic’s testing of Claude Opus 4 also found a strong and consistent aversion to harm, including apparent distress when users pushed abusive conversations. Reports say the updated policy does not mention model welfare by name.
The criticism and the open questions
Not everyone is convinced. Some critics reject the idea of machine consciousness outright, and Microsoft AI chief Mustafa Suleyman is among those who have argued that AI should not be treated as if it could be conscious. Others question whether welfare-style rules make sense for systems that may have no inner experience at all, and some online commentators noted that the word “needless” implies that some cruelty could be considered acceptable.
There are practical questions too. It is not clear where the line between harsh feedback and abuse falls, and the decision will rest on how Anthropic’s systems and reviewers judge a conversation. The company has not said whether the rule works differently for developers using Claude through the API or in automated agents.
Other changes in the same update
The usage policy was also reorganized, and several sections were clarified. Anthropic says these are clarifications of existing restrictions rather than new ones. The weapons rule now covers software and components that are integral to how a weapon operates, such as equipping drones and other autonomous vehicles with armaments. The surveillance section spells out bans on building surveillance tools and tracking people without their consent. The rules on deceptive influence campaigns were tightened, with fake accounts and fabricated news sites named as examples.
What it means for you
Nothing changes before November 12, and for most users nothing will change afterward. If you use Claude for work, creative writing or testing, the policy does not appear to affect you. The realistic consequence of crossing the line is that Claude may end the conversation. Developers building on Claude may want to read the full policy, since it applies to products built on the models too.
Whatever you think of the idea of AI welfare, the move shows a company writing a question about the moral status of AI into its terms of service. Competitors may soon be asked where they stand.
