More than 100 specialists are calling for protections for safety evaluators working inside AI companies. The AI Evaluator Forum published their letter on September 18, 2026. Geoffrey Hinton and Stuart Russell are among the signatories, who sign personally rather than on behalf of their employers.
The group supports embedded evaluation, subject to conditions. Evaluators should retain independent judgment and editorial control, without significant separate business relationships with the lab or payment tied to favorable findings. Several evaluation organizations should bring different perspectives.
The letter also calls for access to relevant systems and staff, transparency about methods and results, and direct communication with oversight bodies. Evaluators should be able to publish uncomfortable findings. Funding and legal protections should keep those findings from putting their work at risk.
ANSA reports the appeal in the context of proposals for outside oversight. It remains a set of demands, not an enforced agreement. It identifies no new AI incident and does not condemn a specific partnership. Follow-up depends on which labs adopt these conditions and what their evaluators are ultimately allowed to disclose.