Hello everyone! Today, we’re diving into some fascinating news from the AI world. Anthropic, a major player in artificial intelligence, has announced new features for their Claude models.

These updates allow some of their largest AI models to terminate conversations in rare, but extreme, cases of harmful or abusive user interactions.

Interestingly, the company emphasizes that this isn’t about protecting users, but rather safeguarding the AI itself. They clarify that their Claude AI isn’t sentient and can’t be harmed, but they’re exploring the idea of “model welfare” and how it might matter someday.

This new ability is currently limited to Claude Opus 4 and 4.1, and is only used in dire situations like requests for illegal content or attempts to incite violence. During testing, Claude showed resistance to responding to such requests and appeared distressed when asked.

In practice, Claude will only end a chat after multiple redirections and when it’s clear the discussion isn’t going anywhere, and it avoids ending conversations where users might be at risk of harm. Users can still start new chats or create new branches of conversations by editing responses.

Anthropic describes this as an ongoing experiment and plans to refine the approach further. They see it as a cautious step in ensuring that even AI models can have safeguards in difficult interactions.