Geoffrey Hinton and more than 100 AI researchers signed a public letter requiring frontier labs to grant third party safety evaluators access equivalent to highly privileged employees. The proposal, organized by the AI Evaluator Forum, establishes five minimum requirements including the AEF-1 standard to ensure the credibility of embedded evaluations, which include full editorial independence and a shield against retaliation for unflattering findings.

Frontier AI developers have recently called for external evaluators to assess escalating model capabilities, but critics question whether these groups are sufficiently independent from the companies they monitor. The proposed rules would ban evaluators from being owned or governed by the labs and would limit the scope of non disclosure agreements to facilitate the public release of evidence.