Definition
AI safety evaluation encompasses methodologies for measuring model alignment, self-knowledge claims, deception, and risky capabilities — where deployment confounds can invalidate conclusions.
Key Points
- Model self-reports (“I’m just an AI”) used in safety debates may reflect chat-templates not intrinsic model properties (2026-09-27-chat-template-llm-self-referential-voice)
- Eval designers must control for instruct formatting, system prompts, and deployment stack
- activation-steering offers experimental control for ablating template effects