Antropic says some Claude models can now end “harmful or abusive conversations”

Antropica has announced new ability that will allow some of their most new patterns to the final conversations in what the company describes as “cases of persistent interactions of persistent users.” Strischantly, Anttropica says that’s done this for not protect the human user, but I prefer the ai model itself.

To be clear, the company is not claiming that their patterns are claude are sensitive or can be damaged by their conversations with users. In their words, antiircle remains “highly uncertain from the Potenti’s moral of claude and other llms, now or in the future.”

However, his point of a point to A recent program created to study what you call “welfare model“And tell her anthropic is essential to take a justifying approach” to identify interventions to the risk mitigation to modify the welfare, in case such wellness is possible. “

This last change is currently limited to Claude opus 4 and 4.1. And again, it’s just happening in the extreme edge cases, “questions” from users of sexual content that will be complying with large scale violence. “

While those types of questions could create legal or advertising problems on the anthropic itself (recent reviews recently around as Chatgpt can reinforce potentially or contribute to user disappointment “), society says that in pre-implementation tests, Claude Opus 4 has shown a “strong preference and a” apparent distress model “when did that.

As for these finance capabilities: The company says: “In all cases, Claude is only exhausted, or when a user explicitly clade to complete a chat.”

Antropic says Claude too

When Claude ending a conversation, antenopic tells that the users still able to have new conversation on the same, and creating a branches of ingriding conversation by educating their answers.

“Treating this feature as an educational experiment and will continue to refresh our approach”, the company says.