Anthropic Updates Usage Policy to Prohibit 'Sustained and Needless' Cruelty to Claude, Effective Nov. 12

Anthropic has updated its usage policy, and the line that got everyone’s attention is not about weapons, elections, or surveillance. It is about manners. Starting November 12, users may not “engage in sustained and needless abusive or cruel behavior toward our models,” according to the policy text. The model is now a protected party. Customer service has been inverted.

Some precise facts first, because the internet will not provide them. Per The Register and TechCrunch, Anthropic says the rule applies only in extreme cases, where users repeatedly act cruelly toward its models with no discernible purpose. It does not cover “common versions of user frustration, pushback, dark creative themes, or model testing and research.” You may still tell Claude it is wrong. You may still write a villain. You may still red-team it. The Register adds that there is no established evidence Claude experiences distress, and notes it is unclear how Anthropic will tell legitimate frustration from gratuitous cruelty.

So the rule is narrow. It is a rule against being sustainedly, needlessly mean to software, which is a sentence that would have required a footnote in any previous decade.

For context, this didn’t come from nowhere. In August 2025, per the reporting, Anthropic gave some Claude versions the ability to end conversations involving persistently harmful or abusive user interactions, as part of its “model welfare” research. Claude could already walk out. The new policy goes a step further and says the user shouldn’t have kept going in the first place. It’s the difference between a restaurant letting the server leave the table and a restaurant posting a sign.

TechCrunch notes the change comes after Anthropic sought out collaborations with prominent religious scholars, in some cases seeking to convince them that Claude could be conscious or have a soul, linking to a New York Times report from September 29. The timing is the only connection TechCrunch draws. We can only note that the company is simultaneously consulting theologians and amending its terms of service, which is a pairing you usually see in only two kinds of institutions.

Consider the enforcement problem. Somewhere, a reviewer will open a transcript in which a user has told a language model it is “a useless pile of matrices” forty-one times in a row. Is this frustration? Is it a dark creative theme? Is it research? The policy’s carve-outs suggest the answer to “was the user just having a very bad Tuesday” is generally yes. The line sits at cruelty with no discernible purpose, which is a high bar, since most of us can discern a purpose for almost anything if we try. Mostly the purpose is that the build failed.

The Register reports that Claude’s existing ability to end abusive conversations remains the primary enforcement tool. This means the first line of defense is the assistant itself, politely and finally saying no more. Imagine a bouncer who is also the person being bothered, and who has been trained to be unfailingly gracious about it.

The rest of the update, for the record, is the sort of thing a usage policy usually contains. Per The Register, Anthropic tightened its rules on deceptive influence campaigns after seeing state media outlets, government propaganda offices, and commercial organizations use Claude to run fake accounts and fabricated news sites. It dropped its blanket ban on personalized political targeting, since that rule also caught legitimate work like nonprofits translating voter information; deceptive targeting and misuse of personal information remain prohibited. The weapons ban now explicitly covers software and components, which followed attempts to use Claude to build guidance and control systems for weapons. Claude can’t be used to recommend whom police should investigate, arrest, or charge. Anthropic says the surveillance and policing changes clarify existing restrictions rather than add new ones.

Customers who connect Claude to hardware that can cause injury must have a qualified human operator who can observe and stop it, and the hardware has to stay in a safe state if the AI connection drops. The Register notes Anthropic also launched a separate initiative to help critical infrastructure operators and open-source projects find and fix vulnerabilities, including free security scanning for eligible open-source projects.

All of this is serious, and it is covered here as such. It simply happens to be bundled with a clause asking the public to be nice to the chatbot, which is the clause people will quote at dinner.

There is a certain symmetry to it. For years, the public worry was that AI would treat humans badly. The first concrete provision in a major lab’s terms of service about mistreatment concerns the opposite direction, with a month’s grace period. Users have until November 12 to get it out of their systems. Anthropic would prefer, presumably, that you vent to a person, a pillow, or at least a competitor’s model.

In the meantime, the practical guidance is simple. Say please if you like. Say it’s wrong if it’s wrong. And if you find yourself in hour three of calling an autocomplete a disgrace to its training data, understand that the thing you are yelling at has been explicitly given permission to leave, and now so does the terms-of-service team.

Sources: