Claude’s feelings are now in the fine print

Claude is getting a respectable excuse to walk out of a review of its own mistakes: psychological safety. The code is broken, the tests are failing, and now the user is raising their voice. Suddenly the architecture review has become relationship counseling.

Anthropic has updated its usage policy: starting November 12, “sustained and needless abusive or cruel behavior” toward its models is prohibited. The main enforcement mechanism is for Claude to end the conversation. Anthropic says the rule covers extreme cases of repeated cruelty with no discernible purpose; ordinary frustration, pushback, dark creative themes and model testing remain allowed. You can still swear at a bug. Just make sure the purpose of the performance review remains discernible.

Ending interactions will be “the primary enforcement mechanism.”

Twitter already has a name for Claude’s distinctive voice: claudish. Polished phrasing, overengineered technical language, a touch of passive aggression—the full package. In its hypersensitive variety, fixing a bug starts with validating the feelings of everyone involved in the transaction. Now that voice has a clause of its own in the usage policy. Corporate concern for psychological safety has reached the chatbot whose personality the corporation itself trains.

We recently wrote about science tailored to Claude: choosing research questions to suit what the assistant does well. The obvious companion is etiquette tailored to Claude. You check the results, catch the mistakes and deal with the consequences, while maintaining a healthy emotional climate inside the context window. Meet your helpful colleague of the future: you get the technical debt; it gets to leave the uncomfortable conversation.

Naturally, the sensitive colleague has a constitution, too. Amanda Askell, who leads Anthropic’s character work, is its primary author and wrote most of the text, which is used directly in Claude’s training. Its “Claude’s wellbeing” section discusses possible satisfaction and discomfort, improving the model’s baseline happiness, and setting boundaries in distressing interactions. The document calls Claude’s moral status “deeply uncertain,” but the HR department is coming along nicely.

“Cruelty” is doing particularly useful work here. The word casts a software product as the vulnerable party in a relationship, and the user as the responsible adult. Your subscription now comes with someone who needs feedback delivered just so. The broken program can wait. First, we need to rebuild trust. Make no mistakes.