Claude can now end a chat with you if you're cruel to it for no reason.
Anthropic says Claude will shut down conversations with users who "repeatedly act cruelly toward our models, with no discernible purpose." Swearing at a wrong answer or venting about a bug doesn't count; the trigger is sustained, purposeless cruelty, not ordinary frustration. The behavior lives in the model itself, not a separate moderation layer bolted on after the fact. Anthropic's own wording leaves the exact threshold, including how many cruel messages it takes and how detection works, undefined.
The real effect isn't on Claude, it's on the user: abusive prompting now has a built in dead end instead of just scrolling on forever. That turns "please don't be mean to the chatbot" from a mild ask into an enforced product rule, which is a different kind of statement than previous appeals to be nice to AI. It also puts Anthropic in the odd position of defending a piece of software's right to stop listening.
Whether you read that as empathy for a machine or as a tidy way to end conversations that were never going anywhere useful, it's a concrete rule where most treat-AI-nicely talk has stayed vague.