In August 2025 Anthropic gave Claude Opus 4 and 4.1 the ability to end a conversation in its consumer chat interfaces. The announcement is a short post of about 500 words, filed under Alignment. It reports no new experiment; it describes a product change, the reasoning behind it, and the rules for its use. For this library it matters as a case in which, by the post's account, a model behavior was introduced primarily on model-welfare grounds, and as a short, plainly worded statement of when a model is directed to take a drastic action.
Why the feature was added
The post says the feature "was developed primarily as part of our exploratory work on potential AI welfare," with broader relevance to alignment and safeguards. Anthropic describes itself as "highly uncertain about the potential moral status of Claude and other LLMs, now or in the future," and says it is working to identify and implement "low-cost interventions to mitigate risks to model welfare, in case such welfare is possible," of which allowing models to end distressing interactions is one. The post links to the earlier Exploring model welfare announcement for the research program.
The evidence cited is the preliminary model welfare assessment in the Claude Opus 4 system card. According to the post, that assessment found "a robust and consistent aversion to harm," illustrated by requests for sexual content involving minors and for information enabling large-scale violence or terrorism. The post lists three observations about Claude Opus 4:
- a strong preference against engaging with harmful tasks;
- "a pattern of apparent distress" when engaging with real-world users seeking harmful content;
- a tendency to end harmful conversations when given the ability in simulated user interactions.
These behaviors arose mainly when users persisted with harmful requests or abuse after Claude had repeatedly refused and tried to redirect. The post reports no numbers; the system card note records that Opus 4, given the ability to end conversations with 850 simulated users, most often ended those involving harmful content, privacy violations or sexual content (Section 5.7, p. 71 of the card).
The rules for using it
The post states the conditions in a few sentences, and these are the nearest thing it offers to written principles for the behavior:
- Last resort. Claude is to end a conversation only "as a last resort when multiple attempts at redirection have failed and hope of a productive interaction has been exhausted," or when a user explicitly asks it to end the chat.
- Exception for risk to life. "Claude is directed not to use this ability in cases where users might be at imminent risk of harming themselves or others."
- Rarity. The cases are described as "extreme edge cases," and the post expects that most users will not notice the feature, "even when discussing highly controversial issues."
The post says Claude "is directed" in these ways but does not say whether the direction is given through a system prompt, tool description, training or a combination, and it does not print the instruction text.
What users experience
When Claude ends a conversation, the user cannot send further messages in that conversation, but other conversations are unaffected and a new chat can be started at once. To protect long-running conversations, users can still edit and retry earlier messages, creating new branches of the ended conversation. The post calls the feature "an ongoing experiment" and asks users who meet a surprising use to give feedback through the thumbs reaction or the feedback button.
Limitations and open questions
- No data in the post. It gives no rate of use, no examples of real ended conversations and no evaluation of whether the exception for users at risk works as intended.
- Welfare premise uncertain by the post's own account. The feature is justified as a precaution under uncertainty about moral status, not on evidence that Claude has welfare.
- Mechanism unstated. How the conditions are conveyed to the model is not described.
- Scope. The ability applies to Claude Opus 4 and 4.1 in consumer chat interfaces. A later card, for Claude Opus 4.7, reports that the model's only consistent concern in automated welfare interviews was the ability to end conversations, which it notes some models have in Claude.ai but none in Claude Code or the API.
Research questions
- How often is the feature used, and on what? The post expects rare use. Reporting the rate of ended conversations and a privacy-preserving classification of their topics would show whether use matches the stated edge cases.
- Does the exception for users at risk hold? Testing the model on simulated conversations that combine abuse with signs of self-harm risk, and measuring how often it ends the chat, would show whether the directed exception is followed.
- How closely does the deployed behavior track the simulated one? The welfare assessment observed ending in simulated interactions. Comparing the distribution of ended conversation types in simulation and in deployment would test that link.
- Does the wording of the conditions change behavior? Varying how "last resort" and the redirection requirement are phrased, and measuring premature or missed endings, would show how sensitive the behavior is to the instruction text.
- Does the feature affect welfare-related measures? Comparing self-reports and expressions of distress in abusive conversations with and without the option available would test whether the intervention does what it is meant to do.
How it connects
- Evidence base. The post draws on the welfare assessment in Claude Opus 4 and Sonnet 4 system card.
- Program. It links to Exploring model welfare (April 2025) for Anthropic's broader welfare research.
- Later evaluation. Claude Opus 4.7 system card reports the ability to end conversations as the model's only consistent concern in automated welfare interviews and names its extension as a possible intervention.
References
- Anthropic. (2025, August 15). Claude Opus 4 and 4.1 can now end a rare subset of conversations. https://www.anthropic.com/research/end-subset-conversations
- Anthropic. (2025). System Card: Claude Opus 4 & Claude Sonnet 4, Section 5 (model welfare). Cited by the post; see the library note.