Over 100 AI Safety Experts Say Current Conditions Make Honest Audits Impossible

Key points

• A coalition of more than 100 AI safety evaluators published a public letter on Friday warning they lack the access and legal protections needed to honestly assess frontier AI systems.
• Geoffrey Hinton, widely called the "godfather of deep learning," signed the letter alongside researchers from Johns Hopkins University, Stanford University and the nonprofit evaluator METR.
• Anthropic CEO Dario Amodei proposed "employee-like access" for third-party evaluators over the preceding weekend; Altman, Musk and Nadella publicly backed the idea but none addressed the operational details.
• The letter calls for evaluators to be shielded from retaliation, including legal threats, if their findings embarrass the companies they audit.
• The letter was published by the AI Evaluator Forum and shared exclusively with CNBC.

The biggest AI labs keep saying they welcome independent scrutiny. A group of more than 100 researchers who actually do that scrutiny just said the current conditions make honest work nearly impossible.

The AI Evaluator Forum, a consortium that organised the letter, published it on Friday. The signatories want frontier model companies, the handful of labs building the most powerful AI systems, to give outside auditors something they have never had: access equivalent to a senior internal employee, the right to talk to staff candidly, and legal protection if their conclusions embarrass the lab. We covered Amodei's original proposal the day it drew wider attention, in "Anthropic's CEO Wants Outside Inspectors Inside His Own Company".

What exactly are these evaluators asking for?

Five things, in plain terms. Freedom from ownership or financial ties to the companies they audit. The right to publish findings publicly, with only narrow carve-outs for customer data or genuine security risks. Direct lines to company boards, not filtered communications through PR teams. Standardised methods across the industry. And legal shelter from retaliation if a lab dislikes their verdict.

That last point matters because nothing currently stops a company from suing an evaluator or cutting their funding after a damaging report.

"When a few powerful labs control capabilities that can endanger cybersecurity, critical infrastructure and the systems our national security runs on, the government and the public cannot be dependent on those labs' own account of what's secure and safe," wrote 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.

Why is this happening now?

Amodei's weekend proposal lit the fuse. Altman, Musk and Nadella quickly offered public backing, but none of them addressed who gets chosen, how deep access goes, or what legal framework governs it.

Conrad Stosz, the Forum's chair, told CNBC the proposal sounds like meaningfully more access than evaluators have ever had, including access to unreleased internal systems. He pointed to a specific example: an unreleased OpenAI model reportedly involved in the Hugging Face security incident. An independent evaluator with proper access might have caught that risk earlier.

Stosz acknowledged the labs could ignore the letter entirely. He noted there are very few organisations with both the technical credibility and the scale to do this work at all, which cuts both ways: the labs need those evaluators if their safety pledges are to mean anything.

Should the political backdrop worry you?

Yes. The Trump administration has pushed back hard against AI regulation, leaving voluntary industry commitments as the main check on how these systems are developed. Voluntary commitments without independent verification are, as Stosz put it, a question of credibility.

This is the sharpest public statement yet that credibility is currently on loan, not earned. AI2Day has tracked this oversight debate across six stories tagged "AI oversight" since 17 July 2026, and the gap between lab promises and verifiable practice has widened each time.

The honest takeaway: if you rely on any service built on a frontier AI model, third-party evaluators with real access and real legal protection are the closest thing to a trustworthy safety check that exists. Right now those conditions don't fully exist. Watch whether any lab signs a binding agreement, not just a press release.