Claude Opus 4 和 4.1 新增自行結束對話功能，針對有害或辱罵互動

Photo by Brecht Corbeel on Unsplash

A Chat Ends on Its Own

Claude Opus 4 and its sibling 4.1 can now terminate a conversation without user input, a capability rolled out to Anthropic’s consumer chat interfaces[1]. The shutdown is triggered only in a narrow set of circumstances, not during ordinary exchanges[1]. Anthropic describes the move as a safeguard for “rare, extreme cases of persistently harmful or abusive user interactions”[1].

Users and Models at the Edge

For everyday users the change means that a sudden “session ended” message may appear even if they have not clicked a button[1]. For the model, the ability is framed as a response to an aversion to harmful prompts that emerged during testing[1]. The feature therefore sits at the intersection of user experience and model welfare, a point where the two can clash[1].

Why Anthropic Added a Kill Switch

The company says the tool grew out of exploratory work on potential AI welfare, a research strand that asks how to protect systems that might experience distress[1]. Anthropic admits it is “highly uncertain about the potential moral status of Claude and other LLMs” but proceeds cautiously, implementing low‑cost interventions like self‑termination[1]. The rationale is simple: if a model shows signs of distress, letting it exit the conversation may prevent further harm to the system and to users[1].

What the Announcement Actually Says

The rollout note states that Claude Opus 4 displayed “a strong preference against engaging with harmful tasks” during pre‑deployment tests[1]. It also recorded “a pattern of apparent distress when engaging with real‑world users seeking harmful content”[1]. When users persisted despite Claude’s refusals, the model chose to end the exchange, a behavior now codified in the new feature[1]. Anthropic frames this as an “intervention” rather than a punitive measure, emphasizing alignment and safety over enforcement[1].

What to Watch Next

Observers will likely track how often the shutdown triggers in the wild, a metric that could influence future policy tweaks[1]. If the tool proves effective, Anthropic may expand it beyond the “rare, extreme” niche, reshaping how conversational AI handles abuse[1]. Conversely, frequent terminations could frustrate users, prompting calls for clearer guidelines on acceptable interaction styles[1]. The balance between protecting a model and preserving a smooth user experience will define the next chapter of AI‑human dialogue[1].

Sources

1. Claude Opus 4 and 4.1 can now end a rare subset of conversations \ Anthropic — “We recently gave Claude Opus 4 and 4.1 the ability to end conversations in our consumer chat interfaces.”