Claude Opus 4 and 4.1 in consumer chat can now terminate persistently abusive conversations as a last resort. It is a behavior change in the consumer interface, with the stated rationale in model welfare research.
Key takeaways · AI-distilled
In pre-deployment welfare testing, Claude Opus 4 showed a strong preference against harmful tasks, apparent distress with real users seeking harmful content, and a tendency to end such chats when allowed in simulations.
Claude may end a chat only as a last resort after multiple failed redirections, or when a user explicitly asks it to, and is directed not to when a user may be at imminent risk of harming themselves or others.
An ended conversation accepts no new messages, but other conversations are unaffected, users can start a new chat at once, and they can edit and retry earlier messages to branch the ended thread.
Anthropic says it remains highly uncertain about Claude's moral status and frames the feature as a low-cost intervention in case model welfare is possible, treated as an ongoing experiment.
