AI Evaluators Publish Demands for Independent Access to Frontier Labs

More than 100 AI experts and evaluators have signed a public letter calling on frontier AI companies to permit genuinely independent safety reviews. The demand goes beyond inviting outsiders in: the signatories want reviewers with access, protection from retaliation, and a path to report findings to boards and the public.

The letter follows Anthropic CEO Dario Amodei’s recent proposal to give some outside evaluators employee-like access to inspect frontier models and their development processes. It is not simply an endorsement of that idea. The signatories set out conditions they say would make embedded reviews credible across frontier AI companies, including Anthropic and OpenAI.

We believe that all frontier AI companies should embed evaluators to independently assess AI risks.

From the public letter signed by AI experts and evaluators

The letter says evaluation organizations should maintain editorial control, disclose and mitigate conflicts of interest, and avoid ownership, governance, or significant commercial ties to the company they are examining. It also rejects payment or rewards tied to an evaluator’s findings.

The group’s minimum conditions

• Access comparable to highly privileged employees, including relevant systems, data, tools, physical spaces, and direct conversations with staff.
• Prompt, unfiltered communication with boards and other privileged oversight bodies.
• Public release of findings and evidence, subject to a limited process for redacting protected information.
• Protection against retaliation, including for reasonable methods or conclusions that reflect poorly on the company.

The letter’s core concern is practical: a reviewer cannot be fully independent if the company under review can limit what it sees, how it works, or what it can say. The signatories want companies to embed multiple organizations with relevant expertise and allow them to describe disagreements with one another and with company employees.

The letter was organized by the AI Evaluator Forum. Its chair, Conrad Stosz, said the effort is meant to establish common principles so independent oversight can become a meaningful tool for managing AI risk, rather than promote one single safety regime.

Amodei’s proposal appears to offer more access than evaluators have previously received, according to Stosz. OpenAI CEO Sam Altman, Elon Musk, and Microsoft CEO Satya Nadella have publicly supported it, but questions remain over which evaluators would be chosen and how deeply they could inspect closely held technology.

The signatories say embedded evaluations should complement internal safety work and broader outside oversight, not replace either. OpenAI has separately said trusted third-party evaluations are important to the safety ecosystem, while arguing that evaluations must clearly show what they tested and how their setup shaped the result.