Anthropic: Claude has been self-grading every conversation for misuse since June

RexDouglass · x · 2026-10-31

Anthropic officially announced self-assessment: after every response, Claude produces a preliminary safety assessment of the whole conversation for potential misuse. The mechanism has been running since June.
Per Anthropic, it's efficient because Claude has already read the conversation, and unlike classifiers targeting a fixed harm taxonomy, it draws on the model's own judgment and improves with the models — surfacing misuse other safeguards missed.
It's drawing backlash: security practitioners uncomfortable with the model "grading them"; one user dropped from $200 to $20/month and moved to DeepSeek, Qwen and GLM, arguing your model should work for you, not file reports on you.
Raises real privacy/surveillance questions: should always-on user-behavior assessment be a default?

More from Models