Anthropic releases recommendations for building robust AI safety evaluations
Anthropicが堅牢なAI安全評価のための推奨を発表
Anthropic released recommendations for building reliable AI safety evaluations, emphasizing interdisciplinary teams, safe testing environments and accurate measurement. Anthropic noted that safety evaluations suffer from problems such as manipulated metrics, insufficient construct validity and a lack of reproducible documentation; whether questions are multiple-choice or free-response, and differences in option order, can significantly change results, a predicament similar to the one psychology faced a decade ago and addressed through reforms such as preregistration. Its recommendations include: forming interdisciplinary teams of methodology experts and domain experts; running evaluations in sandboxed environments without internet access and monitoring them in real time; choosing observational or causal designs according to purpose; and ensuring evaluations truly measure the target concept—for sensitive topics such as suicide, for example, empathetic responses and guiding people to seek human help are more appropriate than refusal rates. Anthropic said it has implemented these principles in its own safety evaluations, using multi-stage evaluations covering entire conversations, scoring with rubrics developed by domain experts, and failing the whole evaluation on a single failure.