Towards embedded evaluations for AI monitoring

lesswrong.com14 min read

The authors propose a claim-based embedded evaluation framework for monitoring AI systems during internal deployment, aiming to detect rogue agent incidents. The framework introduces six core monitoring detection claims covering coverage, capability, trust, disclosure, past deployments, and future deployments. The authors argue that evaluators must work inside frontier AI developers to independently assess these claims and verify oversight mechanisms. Open article →

More on this topic