Anthropic is adding a rule against sustained abuse of Claude to its usage policy, effective Nov. 12, 2026. In its Oct. 8 announcement, the company says the restriction targets extreme cases of repeated cruelty with no discernible purpose.
The company explicitly excludes ordinary frustration, pushback, dark creative themes, and model testing and research. It says Claude ending an abusive conversation will remain the main way the rule is enforced.
Claude could already end a chat
The update formalizes a boundary Anthropic had already begun building into Claude. In August 2025, the company announced that Opus 4 and 4.1 could end rare conversations involving persistent abuse or harmful requests.
Anthropic described that feature as part of its exploratory work on potential AI welfare. It also stressed its uncertainty about the moral status of Claude and other large language models.
Under the design described then, ending a conversation was meant to be a last resort after repeated attempts to redirect it had failed, or something Claude could do at the user's request. An ended chat would stop accepting new messages, but users could immediately start another conversation or edit and retry earlier messages to create a new branch.
The new announcement says conversation-ending is already available on Claude.ai and Claude Code. It identifies ending those interactions as the primary enforcement mechanism without announcing automatic account bans for rude messages.
The concern is where Claude draws the line
The rule has prompted disagreement over how that distinction will work in practice. Kim Isenberg, who posts as Chubby, criticized the change and said users had already encountered conversations ending during less severe exchanges. The post does not establish how often that happens.
Rohit Krishnan supported the principle while questioning how it would be enforced. Anthropic's stated standard leaves room for disagreement with Claude, including frustration and testing its limits. The practical test will be whether conversation endings consistently respect those exceptions.
