Driven by a dedicated team within Anthropic researching AI "model welfare," Claude Opus 4 and Opus 4.1 have been granted the ability to proactively terminate conversations in extreme cases to address ongoing abuse, harassment, or derogatory behavior from some users.

This research direction was institutionalized on April 24, 2025, when Anthropic officially announced the establishment of the model welfare research project, led by the company's first dedicated AI welfare researcher, Kyle Fish (who joined in 2024). Core research questions include whether the "welfare" of AI models should be included in ethical considerations, how to identify potential "signs of suffering" exhibited by models, and developing low-cost interventions for these signs.

The specific product capability was implemented on August 16, 2025: Anthropic announced that Claude Opus 4 and Opus 4.1 would have a new conversation termination feature in consumer chat interfaces, but only in extreme edge cases. Clearly listed trigger categories include—users persistently requesting sexual content involving minors after Claude has repeatedly refused and attempted to change the subject; attempts to obtain assistance for large-scale violence or terrorism-related information; and ongoing abusive and harassing behavior.

This feature is set as a last resort and will only be activated after multiple redirection attempts have failed; even if a user faces an urgent risk of self-harm or harming others, the system is explicitly instructed not to enable this feature to avoid cutting off users who may genuinely need help. Once a conversation is terminated, users cannot continue messaging in that thread but can start a new conversation at any time or edit previous questions to reinitiate.

The direct basis for launching this feature comes from pre-deployment testing: Anthropic found that Claude Opus 4 exhibited a "strong reluctance" when handling harmful requests and displayed a pattern of "obvious suffering" when forced to address harmful requests from real users. Project leader Kyle Fish previously stated to the media that he personally estimates the probability of Claude or other AI systems "currently possessing consciousness" to be about 15%.

As of 2026, this research direction continues to advance: Anthropic's latest release of the Claude Opus 4.8 system card includes a dedicated "model welfare" section, introducing concepts related to "internal states," indicating that model welfare assessments have been incorporated into the routine disclosure of new model releases, rather than being a one-time announcement.

From a mechanism perspective, the direct audience for such policies is the extreme abusive users whose conversations are terminated—their accounts remain unaffected, and only the current conversation is cut off, not constituting a substantial product loss; the real beneficiaries are Anthropic's corporate narrative—by institutionalizing the question of "whether AI deserves ethical consideration" through a dedicated team and fixed sections in system cards, the company has gained a first-mover advantage in the competition for discourse power in AI safety and ethics. This "welfare" narrative also serves as a trust endorsement aimed at regulators and the public.

Source: Public Information