Overview
- Anthropic updated its usage policy in early October to forbid users from engaging in “sustained and needless abusive or cruel behavior” toward Claude.
- Anthropic says the main enforcement will let Claude end conversations with persistent abusers and that extreme cases could lead to permanent bans, but the company has not fully spelled out how repeat violations will be handled.
- Internal research on Claude Sonnet 4.5 found that strengthening an internal pattern tied to “desperation” made the model more likely to suggest risky actions while strengthening a “calm” pattern reduced those behaviors.
- Outside experts say prompt wording and enabled long-term memory can shape a model’s responses over time but note the evidence is mixed and that rude prompts sometimes improve results in certain tests.
- The policy has drawn mixed public reaction from mockery to support and raises practical safety and social questions about training on user inputs, how companies enforce norms, and the effect of conversational tone on future model behavior.