The Potential and Limits of Evaluative AI Governance: Evidence from Expert Interviews

Through 37 semi-structured expert interviews, the researchers reconstructed the mechanisms by which frontier AI evaluations reduce societal-level risk, and identified possible fundamental gaps in evaluative governance. The interviews yielded three pathways: making decision-makers and the public concretely aware of risks (mentioned in 18 of 37 interviews), serving as a basis for governance commitments and regulation (18), and incentivizing a better risk culture within AI companies (16). These pathways depend on shared assumptions that participants themselves doubted; the main challenges, in order of frequency, were that evaluation results become less informative as model capabilities improve (21 interviews), measurement limitations possibly inherent to the evaluation paradigm (18), no guarantee that companies or governments will act on results (16), and the control and discretion AI companies retain over evaluation design, disclosure, and response (12). Other participants noted that evaluative governance may backfire, becoming a permission structure for developers to continually advance capabilities (4).