Anthropic has trained its AI model Claude to perform "self-assessment" of potential abuse in conversations, the company announced on X. Under this technique, Claude generates a preliminary safety evaluation immediately after completing each reply, leveraging its own understanding of the conversation content rather than relying on external classifiers targeting specific harm types. Since deploying self-assessment in June, Anthropic says the method has detected abusive patterns that other safety guardrails missed. The company notes the approach is highly efficient because Claude already processes the full conversation context, and its accuracy improves as underlying model capabilities advance.