Story
September 19, 2026
AI evaluators want inside access — but the fight over who watches the labs is just beginning
More than 100 AI experts are demanding independent, employee-level access to frontier labs, arguing that safety checks cannot rely on company-controlled scrutiny. Critics on the deregulatory right question the legitimacy of the watchdogs themselves.
The debate sharpened after Anthropic chief executive Dario Amodei floated giving outside evaluators “employee-like access” to frontier-model development, a proposal later backed publicly by OpenAI’s Sam Altman, Microsoft’s Satya Nadella and Elon Musk. The practical details, however — who gets chosen and how far inside the companies they can look — remain unsettled.1
On Friday, the AI Evaluator Forum and more than 100 academics, researchers and industry figures answered with a public set of minimum conditions. Their premise is blunt: a safety review is not independent if the company being reviewed controls the information, money or consequences.
The signatories, including Geoffrey Hinton and Stuart Russell, called for “scientific objectivity, transparency, independence, and robust protections against interference.” They want evaluators able to speak directly with boards, publish findings after narrowly limited redactions, and receive the same relevant systems, data, tools and physical access as senior internal assessors.2
Conrad Stosz, chair of the forum behind the letter, said the campaign is about ensuring that independent oversight is a “meaningful tool for managing AI risk broadly,” rather than a substitute for labs’ own safety work. He argued that access to unreleased systems and sensitive internal data would offer far greater confidence about real-world risks.1
Supporters frame the demand as a check on concentrated technological power. Vinh Nguyen, a former National Security Agency chief AI officer, warned that neither government nor the public should have to depend on a handful of labs’ “own account of what’s secure and safe.”1
But the proposed watchdog model has also become a political target. David Sacks amplified a report describing Anthropic’s chosen watchdog as tied to the “woke Effective Altruism movement” and “a complete joke.”
3 In a separate repost, he endorsed a different principle: “The best way to pace the frontier is to hold the labs fully liable for the behavior of their models.”4
That leaves a widening fault line: evaluators seek protected access before harms emerge; critics prefer accountability after the fact, without empowering a new class of AI referees.