AI Evaluator Forum

AEF cites Transluce’s mental health evaluation example as being published with a detailed disclosure of operating conditions consistent with AEF-1, including use of the AEF-1 checklist.

AEF cites METR’s frontier risk report example as involving more direct access to non-public information and more editorial independence than previous external engagements, with the AEF-1 reporting checklist included in the published report.

AEF’s launch announcement describes the creation of the consortium and releases AEF-1 as a voluntary standard intended to be adopted in relevant forthcoming evaluations, alongside a public statement calling for transparency about third-party evaluations.

AEF-1 is presented as a standard and checklist for evaluators to demonstrate baseline operating conditions supporting independence, access, and transparency during third-party AI evaluations.

AEF publishes an initiative statement explaining why evaluators should disclose evaluation conditions—such as methods control, editorial control, access/time, and conflicts of interest—when publishing independent evaluation results.

AEF lists SecureBio’s biosecurity-relevant pre-release assessments as an example of how AEF-1 principles can be applied when setting out access, funding/conflict rules, and publication/disclosure approaches for sensitive information.

AEF states that the EU AI Office endorsed key provisions of AEF-1 as a way for providers to comply with independence provisions of the General-Purpose AI Code of Practice.