← Back to AI Frontiers
Anthropic trains Claude to self-assess for misuse, launches Safeguards blog
Claude now gives a preliminary safety assessment right after it responds; since June, self-assessment has surfaced misuse other safeguards weren't designed to catch.
Anthropic has trained Claude to flag potential misuse in conversations, a technique it calls self-assessment: right after it responds, Claude gives a preliminary safety assessment of the conversation.
- Since use began in June, it has surfaced misuse that other safeguards weren't designed to catch.
- Because Claude has already read the conversation, it is very efficient; unlike classifiers targeting a fixed set of harms, it draws on Claude's own judgment and gets more accurate as models improve.
- Per the blog: on Claude Fable 5, at a 1% false-positive rate it flags 78% of potentially harmful conversations and 94% of high-priority ones, adding an estimated <2% inference compute.
Anthropic also launched a new Safeguards blog.