Key Info
Anthropic has trained Claude to flag potential misuse through a technique called self-assessment, in which Claude gives a preliminary safety assessment of the conversation immediately after it responds.
Highlights
- Since self-assessment was introduced in June, it has surfaced misuse that other safeguards were not designed to catch.
- The method is efficient because Claude has already read the conversation.
- Unlike classifiers that target a discrete set of harms, it draws on Claude's own judgment and becomes more accurate as models improve.
- Details are published on Anthropic's new Safeguards blog.