❯ Anthropic uses post-response Claude self-assessment to screen for misuse efficiently

A check after the answerAnthropic drew renewed attention this week to self-assessment, a safety technique described in its October 16 research post. After answering, Claude produces a brief assessment invisible to the user, screening the conversation for potential misuse and feeding offline analysis and further review. The method covers a large share of production traffic on models released from Fable 5 onward.

Reusing existing contextA separate reviewing model would normally need to reread the conversation. Self-assessment reuses the responding Claude model and its KV cache, the intermediate computation already available for that context. It adds a short instruction and roughly ten output tokens on average. Anthropic estimates less than 2% additional inference compute across its production API traffic distribution, without charging those tokens to users.

An initial screening layerIn Fable 5 tests, compared with a stronger prompted classifier, self-assessment flagged 78% of potentially harmful conversations and 94% of high-priority conversations at a 1% false-positive rate. Flagged conversations feed offline analysis and further review to identify aggregate risks, cases missed by other detectors and new misuse patterns. Other safeguards continue to handle real-time blocking.

▮ SIGNALReusing computation from the original task can make sophisticated safety screening affordable across a much larger volume of traffic.