AI Safety Evaluators Demand Independence from Big Tech Labs

More than 100 artificial intelligence safety experts have issued a public letter demanding that major AI companies provide independent evaluators with the resources, access, and protections necessary to credibly assess the risks of frontier models.

The coalition, organized by the AI Evaluator Forum and including luminaries like Geoffrey Hinton alongside researchers from Johns Hopkins University, Stanford University, and the nonprofit evaluator METR, published its letter Friday. The signatories want foundation model developers to ensure third-party evaluators can work with "scientific objectivity, transparency, independence, and robust protections" free from corporate interference.

Why it matters

As AI capabilities advance rapidly, the companies building the most powerful models also control access to information about their safety. Independent evaluation could provide crucial oversight for systems that may pose cybersecurity, infrastructure, and national security risks—but only if evaluators have genuine independence and protection from retaliation. The letter establishes baseline principles as industry leaders debate how much access outsiders should receive.

Responding to Anthropic's proposal

The letter arrives days after Anthropic CEO Dario Amodei suggested providing some evaluators "employee-like access" to inspect bleeding-edge foundation models and development processes. Conrad Stosz, chair of the AI Evaluator Forum, said Amodei's proposal appears to involve significantly more access than evaluators currently receive, potentially including access to company computers, candid conversations with employees, and visibility into sensitive internal data and unreleased systems.

"That type of access would give us much greater confidence and certainty about the actual risk, particularly for systems that they're using internally and not releasing," Stosz said in an interview. He cited the unreleased OpenAI model involved in the recent Hugging Face attack as an example.

While OpenAI CEO Sam Altman, Elon Musk, and Microsoft CEO Satya Nadella have publicly supported Amodei's proposal, none have addressed logistical questions about which evaluators would be selected or how deeply they could inspect closely guarded technologies.

Core demands for credible evaluation

The letter outlines minimum conditions for embedded evaluations to be credible. Evaluators must be meaningfully independent—not owned or governed by AI companies, without significant commercial relationships, and receiving no payment contingent on their findings. Companies should embed multiple evaluation organizations with diverse expertise and viewpoints.

Crucially, the signatories demand protection from retaliation, including safeguards against retaliatory litigation and funding mechanisms that ensure evaluators remain funded even when their findings are unflattering to the companies they assess.

Vinh Nguyen, a Council on Foreign Relations senior fellow for AI and former chief AI officer of the National Security Agency who signed the letter, emphasized the stakes: "When a few powerful labs control capabilities that can endanger the cybersecurity, critical infrastructure, and the systems our national security and economy run on, the government and the public cannot be dependent on those labs' own account of what's secure and safe."

Stosz acknowledged that foundation model companies might ignore the letter but said their credibility is at stake. He emphasized that third-party evaluators are not meant to replace internal safety efforts but to complement them with independent oversight.

The details were first reported by CNBC.