Supporting case
Brookings compares who evaluates AI—and what evaluation cannot establish
What happened
An October 16 Brookings paper by Brooke Tanner, Cameron F. Kerry, Elham Tabassi, Andrea Renda and Andrew W. Wyckoff compares AI-evaluation approaches across countries. It distinguishes developers, government security institutes and outside evaluators. Tests measure selected properties under particular conditions; their results are not equivalent to assurance about every real-world use. Benchmark adaptation and the separation between predeployment tests and later monitoring complicate interpretation.
The brief was prepared before a September 15 international dialogue, whose discussion operated under rules protecting participants’ identities. It is a comparative framework, not the establishment of a new evaluator or binding international standard. Future European evaluation arrangements it discusses must remain future arrangements. Primary article and underlying analysis ↗.
The Race question is whether an evaluator has independent access, can publish meaningful findings and can cause a remedy. Access negotiated with a developer may be valuable without transferring those powers. Public participation and ordinary institutional harms also deserve attention alongside national-security testing.
Dates and implementation stage
Development: 2026-10-16; publication 2026-10-16
