Building strong safety evaluations: Anthropic's recommendations for AI safety evals

Anthropic published recommendations for building AI safety evaluations, drawing on reforms from psychology's replication crisis and its own experience building evals for system cards. The post argues safety evals are easy to game and often fail to measure what they claim, urging teams to combine methods and domain experts, run evals in sandboxed environments, and match observational or causal designs to the question asked.