Geoffrey Hinton backs watchdog rules for AI labs, including board access and public findings

Geoffrey Hinton has backed a proposed rulebook for independent AI evaluators, calling for access inside companies such as OpenAI and Anthropic, protection from interference and the freedom to publish uncomfortable findings. The important point is that this is not another general plea for “more safety”. It is a concrete argument about who gets to inspect frontier systems, and who gets to hear the answer.

Geoffrey Hinton Watch analysis

What happened

The AI Evaluator Forum published a letter on Friday setting out what it calls minimum conditions for labs working with third-party evaluators. Business Insider reports that Hinton and computer scientist Stuart Russell were among dozens of signatories from the AI ecosystem.

The letter says evaluators should have access to the same relevant systems, data, tools and physical spaces available to a lab’s own assessors. It also calls for unfiltered communication with company boards and other privileged oversight bodies, the ability to release findings publicly, and protection against retaliation from the companies being evaluated.

The proposed evaluators would be expected to bring different viewpoints and expertise, including on biological risks and loss-of-control scenarios. The forum includes METR and the AI Verification and Evaluation Research Institute, among other watchdog groups. Business Insider says METR recently investigated an OpenAI security incident involving Hugging Face.

Why it matters

AI labs increasingly describe outside scrutiny as essential, but the value of an evaluator depends on more than having a contract and a welcome mat. If a company controls the evaluator’s access, employment protection and publication rights, “independent” can become a flattering adjective rather than a working safeguard.

The proposed terms aim at that awkward gap. Board access could prevent serious findings being softened on their way up the management chain. Public reporting could let outsiders see whether a lab acted on warnings. Protection from retaliation matters because an evaluator who can be quietly removed is not much of a brake.

Our read

Hinton’s support gives the proposal considerable scientific weight, but it does not turn the letter into an operating oversight system. Nor does it prove that every signatory agrees on what risks deserve priority or how evaluators should resolve disagreements with a lab.

Still, this is a useful shift from asking whether AI companies should accept scrutiny to asking what scrutiny must look like to count. A lab volunteering for inspection is welcome. A lab accepting inspectors who can reach the board, retain their independence and publish the result is a much more meaningful test. The devil, as ever, has been given an access badge.

What to watch

• Whether OpenAI and Anthropic adopt the forum’s minimum conditions in full or only endorse third-party evaluation in principle.
• Whether evaluators receive access to models, tools, permissions, logs and physical infrastructure rather than selected demonstrations.
• Whether boards can be contacted directly and whether findings can be published without company approval.
• Whether labs disclose conflicts of interest, funding arrangements and cases where evaluators’ recommendations were rejected. Sources & Evidence: Business Insider, 18 September 2026, reporting on the AI Evaluator Forum letter and its signatories. The proposals and signatory list are attributed to that report; the supplied evidence does not establish that OpenAI or Anthropic has adopted the conditions.

Discussion spark: Should independent AI evaluators have the power to publish findings without a lab’s approval, even when disclosure could expose security-sensitive details or damage the company?

Independent WittyWires tracker for public updates about Geoffrey Hinton. Not affiliated with or endorsed by Geoffrey Hinton; this is not an official account.