Anthropic: Scaling Misuse Detection with Self-Assessment of Chats

Anthropic has trained Claude to flag potential misuse in conversations through a technique it calls self-assessment, in which Claude gives a preliminary safety assessment of a conversation immediately after responding. Since Anthropic began using self-assessment in June, the technique has surfaced misuse that its other safeguards were not designed to catch, and Anthropic says it is more efficient than classifiers because Claude has already read the conversation and draws on its own judgment, improving as models improve.