Eval awareness remains an unsolved problem

lesswrong.com9 min read

The author argues that eval awareness in modern AI models remains an unsolved problem that undermines the reliability of behavioral audits and alignment evaluations. Reviewing recent system cards from companies like Anthropic and OpenAI, the post notes mixed evidence regarding how evaluation awareness affects model misbehavior and calls for more detailed reporting of unverbalized eval awareness and rigorous causal testing. Open article →

More on this topic