Anthropic: Claude Trained to Self-Assess Potential Abuse in Conversations

October 30: Anthropic said in a post on X that it has trained Claude to identify potential abuse in conversations, calling the technique "self-assessment": after completing a reply, Claude immediately gives a preliminary safety assessment of the conversation. Anthropic said that since it began using the technique in June, it has detected abuse that other safety guardrails failed to cover; because Claude has already read the conversation content, the method is highly efficient, and unlike classifiers targeting specific harm types, it relies on Claude's own judgment, so its accuracy improves as model capabilities advance.