Toward Best Practices for AI Evaluation and Governance

A promising way of identifying and mitigating the systemic risks posed by artificial intelligence (AI) development and deployment is general-purpose AI (GPAI) evaluations. Although GPAI evaluations play an increasingly central role in institutional decisionmaking and policymaking, including the European Union (EU) AI Act’s mandate to conduct evaluations on GPAI models presenting systemic risk, no standards exist to promote GPAI evaluations’ quality. To strengthen GPAI evaluations in the EU, the first and only jurisdiction that mandates GPAI evaluations, the authors outline four desiderata for evaluations: internal validity, external validity, reproducibility, and portability. To uphold these desiderata in a dynamic environment of continuously evolving technologies, the authors propose a dedicated EU GPAI Evaluation Standards Task Force to be housed in the bodies established by the EU AI Act. The authors outline the task force’s responsibilities, discuss its potential influence on global AI governance, and address potential sources of failure that policymakers should heed. This paper is primarily intended for EU policymakers and emerging EU AI governance bodies, particularly those involved in implementing the EU AI Act’s provisions on GPAI evaluations. Secondary audiences include AI safety and security institutes, GPAI providers, civil society organizations, and international standards bodies that might interact with EU frameworks.