The BBC reports that Anthropic made the change on Thursday and that it was first reported by The Verge. The company said Claude can end interactions in which people are "cruel", something it did in "rare" cases before the update.
The new clause sits on a list of forbidden conduct that already covered bullying, promoting self-harm and non-consensual intimate imagery. Anthropic stressed the rule is for "extreme cases" of repeated abuse. The exemptions it named are user frustration, pushback, dark creative themes, and model testing and research.
The definition is the weak point. The BBC notes that what Anthropic would count as "abusive or cruel behaviour" towards its models "remains unclear". Screenshots of the policy spread on social media. Some people called it good manners or model welfare. One commenter called it "deeply wrong" and another argued it could undermine abuse directed at humans and animals.
Dr Barry Scannell, a technology partner at Irish law firm William Fry, wrote on LinkedIn: "You can't be cruel to numbers and maths." He said anthropomorphising AI at this level is harmful because it leads people to believe it is something it is not. Mustafa Suleyman, Microsoft AI director, who criticised Anthropic in September for treating AI like it is human, reacted with a melting face emoji.
The same update touches elections. With US midterms ahead, the policy now says Claude cannot be used to "deceive voters or disrupt elections", and it forbids "deceptive campaigns". Anthropic also said its policy already blocked using its tools to develop weapons, which it restated in the update.
For product teams, the practical points are small. The rule changes how Claude may end a chat, not what developers may build. The open question is enforcement: no figures on how often Claude has ended a conversation were in the source.