Anthropic Adds Limits on Abusive Treatment of Claude
Anthropic has revised its usage rules to prohibit sustained, pointless cruelty toward Claude, while saying ordinary frustration, dark creative work, and testing remain allowed. The main enforcement method is still for Claude to end a conversation with a persistently abusive user. The policy update also consolidates rules covering deceptive campaigns, elections, weapons, law enforcement, and physical actions.
Anthropic's revised policy targets only extreme cases: repeated cruelty with no clear purpose. It explicitly leaves room for everyday annoyance, disagreement, somber artistic material, and model evaluation or study. When Claude terminates a conversation, that thread closes to further messages, though other chats remain available.
The option to end conversations began in August 2025 during AI welfare research. Anthropic said it was uncertain about Claude's moral status but sought low-cost risk reduction. It observed Opus 4 showing harm aversion, distress in abusive exchanges, and a tendency to stop harmful chats. The update also consolidates rules on deception, elections, weapons, law enforcement, physical actions, and intimate imagery, taking effect November 12.
The change may mainly affect heavy users and researchers who test models through hostile prompts, as well as platform moderators setting norms for AI interaction. It could signal that AI companies are treating model-directed cruelty as a safety and welfare issue, potentially influencing how other developers design refusal and conversation-ending tools. At the same time, enforcement remains limited to extreme cases, so broader effects on free expression, creative work, and ordinary user frustration may be modest.