Who Should Run Advanced AI Evaluations -- AISIs?

Public bodies should directly evaluate safety-critical AI models: The paper recommends that government AI Safety Institutes (AISIs) conduct "gray-box and white-box" evaluations of advanced AI systems themselves, rather than relying solely on private evaluators, due to the high risks involved and the difficulty of verifying safety claims made by AI labs.

Private evaluators can handle lower-risk assessments under oversight: Governance audits, security audits, and "black-box" model evaluations (where evaluators can only test inputs and outputs) can be more efficiently provided by private companies, similar to evaluation regimes in other industries, as long as public bodies maintain oversight.

Information sensitivity requires public involvement: The paper identifies the sensitive nature of information involved in AI evaluation as a key reason why public bodies need direct access to AI models and facilities—private evaluators alone cannot be trusted with proprietary model details that are critical for safety assessment.

Public AI evaluation capacity must scale with industry risk: AISIs will need substantial resources and capacity that grows alongside the size and risk level of the AI industry, including extensive access to AI models and evaluation facilities, to effectively fulfill their governance role.

Lessons from other industries inform AI governance: The paper draws from nine established evaluation regimes across different industries to show that effective technology governance typically requires a mixed approach combining public safety evaluations with private market oversight, rather than one approach alone.