Hinton and 100 Researchers Say AI Labs' Oversight Pledges Are Hollow Without Five Structural Fixes

Since Anthropic CEO Dario Amodei's September 12 essay calling for AI companies to grant external evaluators "employee-like access" to their systems, and the wave of CEO endorsements that followed, no lab has named a specific evaluator, defined what they'd actually see, or set a start date. That silence is the backdrop for a letter published Friday by more than 100 AI researchers and security professionals, who say the oversight everyone just spent a week praising doesn't structurally exist.

The letter was organized by the AI Evaluator Forum and shared with CNBC ahead of its public release, according to Tech Times. Signatories include Nobel laureate Geoffrey Hinton, Princeton computer scientist Arvind Narayanan, researchers from Johns Hopkins and Stanford, and members of METR, the nonprofit that has become the primary independent evaluator working with frontier AI labs.

Amodei's proposal landed fast. Sam Altman endorsed it within hours. Elon Musk posted that "Dario is right." Satya Nadella and Demis Hassabis both signaled support. What none of them did, per Tech Times' reporting, was commit to a name, a scope, or a timeline.

The Argument Is About Structure, Not Motive

The letter is careful to frame its complaint as structural, not an accusation of bad faith against any specific lab. Its authors argue that self-assessment is unreliable no matter how sincere the company running it is, because the conditions that would let an outsider actually verify a lab's claims haven't been built anywhere in the industry.

The coalition lists five minimum requirements before embedded evaluation arrangements can be trusted. The one spelled out in detail: "substantive independence," meaning evaluation agencies cannot be owned or financially controlled by the same companies whose models they're assessing, according to the AI Evaluator Forum's letter. The letter also raises how evaluators are funded as a separate point of concern, tying compensation structure to the risk that an evaluator dependent on a lab's goodwill has an incentive not to find problems.

If METR or any other evaluator is paid by the same company it's supposed to be checking, and that company can walk away from the arrangement whenever it wants, the evaluator's incentive to flag serious problems is compromised before a single audit happens. This isn't a hypothetical from activists with no stake in the field. It's coming from the researchers who'd have to sign their names to these audits.

The Labs' Side of This

Building genuine third-party access to frontier model weights, training data, and internal deployment decisions in a matter of weeks is not realistic. Anthropic, OpenAI, and Google DeepMind are commercial entities protecting trade secrets and, in some cases, national-security-sensitive capabilities. A public commitment made on September 12 turning into a fully staffed, contractually independent evaluation regime by September 19 was never a plausible timeline, and no source here claims any CEO promised that speed. METR itself already has an existing working relationship with several labs, which is more oversight infrastructure than existed two years ago.

The letter doesn't dispute that some progress has occurred. Its point is narrower: praising "third-party scrutiny" in a press essay is not the same as building it, and until specific evaluators, specific access scopes, and specific funding firewalls are named, the public commitments are promises, not policy.

No regulator has mandated any of this, and no legislation currently requires frontier labs to submit to outside audits at all. That makes the entire framework dependent on companies keeping pledges they set for themselves, with no enforcement mechanism if they don't. Whether that's a feature or a flaw depends on whether you trust markets and reputational pressure to do the job, or think an industry moving this fast on something this consequential needs a backstop with teeth.

The open question now is whether any lab responds with specifics. Amodei, Altman, Nadella, and Hassabis have all endorsed the concept of external evaluation in public. None of them, as of this writing, has named an evaluator, an access scope, or a start date in response to Friday's letter.