It also says four further benchmarks did not provide enough information to receive a verdict.
The update belongs to Epoch AI's public review project for external AI benchmarks. Its documentation says reviews are selected for impact, reach, and coverage diversity. Benchmarks requiring specialist knowledge can take longer because the organization hires external subject-matter experts. Epoch AI says it does not review benchmarks it created itself because of the conflict of interest.
The documentation distinguishes Verified, Flawed, and Not enough information outcomes, and publishes a dated table of reviewed benchmarks. For reviews drawing on private material, Epoch AI says it will disclose only generic findings to avoid leakage. The update does not establish that the newly reviewed benchmarks are representative of all AI evaluations, nor does it provide a comparative measure of model performance or identify causes for every verdict in the batch.