Anthropic has updated the usage policy for Claude so that its tools may end an interaction when a user is "cruel". The company says it had done this in rare cases already; it now sits formally on a list of forbidden conduct alongside bullying others, promoting self-harm and creating non-consensual intimate imagery.
The company was careful about scope. The rule applies to "sustained and needless" abuse in "extreme cases", and explicitly not to "common versions of user frustration, pushback, dark creative themes, or model testing and research". What it has not published is a definition of cruelty, which is where the argument starts.
The objection
Screenshots of the policy spread quickly, with some readers welcoming it as good manners or "model welfare" and others calling it deeply wrong.
The sharpest criticism came from Dr Barry Scannell, a technology lawyer at the Irish firm William Fry, writing on LinkedIn: "You can't be cruel to numbers and maths. This level of anthropomorphisation of AI is harmful. It leads people to believe that it's something it's not." Microsoft AI director Mustafa Suleyman, who criticised Anthropic in September for treating AI as human, responded to the update with a melting-face emoji.
The objection is substantive. If a company with every commercial reason to be believed starts implying its product can be wronged, it is teaching users something false about what the product is — and that belief is the root of a long list of problems, from people trusting a chatbot's confident invention to people forming attachments to one.
The defence nobody is making
There is a straightforward case for the policy that does not require Claude to have interests at all, and it is odd that it has not featured more in the argument.
A company may decline to operate a service for abusive users simply because it does not want to, in the way a shop can refuse a customer who screams at the staff — even when the staff member is a vending machine. The reason is not the machine's feelings. It is what the business is willing to host, what it costs to moderate, and what kind of use it wants to normalise.
Framed that way the rule is unremarkable. Framed as protecting the model, it invites exactly the confusion Scannell describes. Anthropic chose the second framing, and the backlash is the price of it.
Why this matters more here than it looks
Anthropomorphisation is not a philosophical indulgence; it is a practical risk, and it is larger where AI tools arrive without much context around them.
A user who half-believes a chatbot is a mind is more likely to accept its answer without checking, to tell it things they would not type into a website, and to take its reassurance as reassurance. That is the same failure a Scottish school decided to teach children about directly — the lesson that it has human-like qualities but is "just a machine", which is why a school in Scotland now employs someone to teach it.
A policy that leans the other way, however well intentioned, works against that lesson. And the stakes of the belief are visible in China's move against AI companion apps and in the testing of ChatGPT's teen safeguards.
The rest of the update
The cruelty clause came out of an annual policy review and was not the only change, though it took all the attention. Anthropic also barred using Claude for "deceptive campaigns", and clarified — with US midterm elections approaching — that it cannot be used to deceive voters or disrupt elections.
That second clause is the consequential one. It is enforceable, it addresses a documented harm, and almost nobody wrote about it.
