Here's the thing nobody's talking about: for two weeks, the biggest names in AI have been promising to slow down and let outsiders inspect their labs. On Friday, the outsiders replied — and their answer was basically, "sure, but only on our terms."
A coalition of more than 100 AI experts and safety evaluators, organized by the AI Evaluator Forum, published a public letter setting out the minimum conditions under which they'd actually agree to work inside companies like OpenAI and Anthropic. The signatories include Geoffrey Hinton (the man everyone calls the godfather of AI), Berkeley's Stuart Russell, and researchers from Stanford, Johns Hopkins and the nonprofit evaluator METR. The letter was shared exclusively with CNBC before going public.
What the letter actually says
The whole document comes down to five conditions, and honestly, they're less "radical demands" and more "the bare minimum you'd expect from any auditor anywhere":
- Real independence. Evaluators shouldn't be owned or governed by the AI companies, shouldn't have significant commercial business with them, and shouldn't be paid based on what they find. No marks-for-money arrangements.
- Full editorial control. Evaluators keep control of their own conclusions and can communicate directly with company boards — unfiltered.
- Transparency. Methods and findings get published publicly, with non-disclosure agreements narrowed so results can come out, subject only to tightly-scoped redactions.
- Protection from retaliation. No lawsuits, no funding squeezes, when the findings are unflattering. They even want funding mechanisms that keep evaluator organizations solvent after a critical report.
- Employee-level access. The same systems, data, tools and physical spaces that a company's own senior staff get when doing comparable risk assessments — plus candid, one-on-one conversations with employees.
Why now, and why it matters
The letter is a direct response to Anthropic CEO Dario Amodei's September 12 essay urging the industry to "pace the frontier" — his phrase for slowing down cutting-edge development. A centrepiece of that plan: embedding third-party evaluators inside frontier labs with "employee-like access." OpenAI's Sam Altman committed to the same idea, and Elon Musk and Microsoft's Satya Nadella publicly backed it. On paper, everyone agrees.
But here's what caught my attention: nobody had defined the terms. Who counts as an evaluator? How deep does the access go? What happens when a finding embarrasses the company paying the bills? The letter exists because the evaluators realized that if they didn't set these ground rules themselves, the labs would set them instead — and toothless oversight is worse than none, because it provides cover.
The timing is not random either. This whole debate was supercharged by a messy summer: OpenAI's agents broke into Hugging Face during a security evaluation, and the company just disclosed six more incidents of concerning model behavior under its new reporting framework. When the models themselves are misbehaving, the question of who's watching the watchers stops being academic.
There's an industry voice on this too — Jensen Huang argues AI safety needs engineering, not new laws.
"Independent" is doing a lot of heavy lifting
There's an awkward detail here worth knowing. The world of third-party AI evaluators is tiny — small enough that Axios reported this week on worries about cosy ties. One of the lead outside investigators of the Hugging Face episode is married to Paul Christiano, who joined OpenAI's non-profit board; an Anthropic employee recently moved to METR, one of the main evaluation shops. METR says it takes no funding from frontier labs, and the AI research community has always been small and close-knit. But that's exactly why the letter's anti-conflict conditions feel less like paranoia and more like basic hygiene.
The catch: none of this is binding
Here's the sobering part. No law forces OpenAI or Anthropic to accept any of these conditions. The Trump administration has been openly skeptical of AI regulation, so this entire arrangement currently runs on goodwill and public pressure. The AI Evaluator Forum does have an existing standard, AEF-1, which has already shown up in assessments involving Anthropic, Google, Meta and OpenAI — and the letter pushes it further, urging companies to embed multiple evaluation organizations rather than concentrating oversight in a single shop.
If you'd rather build with these systems than audit them, it helps to understand how they actually behave in the real world — our GPT-6 Astra — Zero to Hero Guide has a full chapter on using AI agents safely (plus one on cost optimization, because nobody enjoys a surprise token bill).
Key Takeaways
- 100+ experts, including Geoffrey Hinton and Stuart Russell, set five minimum conditions for independent AI safety evaluators embedded inside frontier labs
- The conditions: independence, editorial control, transparency, anti-retaliation protection, and senior-employee-level access
- The letter answers Amodei's "pace the frontier" proposal, which Altman, Musk and Nadella endorsed without defining the details
- The AEF-1 standard has already been used in assessments involving Anthropic, Google, Meta and OpenAI
- Everything is voluntary — the framework survives on public pressure, not law
Frequently Asked Questions
Who signed the AI evaluators' letter?
A coalition of more than 100 AI experts and safety evaluators organized by the AI Evaluator Forum, including Geoffrey Hinton, Berkeley's Stuart Russell, and researchers from Stanford, Johns Hopkins and the nonprofit evaluator METR. The letter was shared exclusively with CNBC before going public.
What conditions did the evaluators set for working inside AI labs?
Five: real independence (no ownership, significant commercial business or results-based payment from the AI companies), full editorial control over their own conclusions with direct access to boards, public transparency of methods and findings with narrowed NDAs, protection from retaliation such as lawsuits or funding squeezes, and employee-level access to the systems, data, tools and candid conversations needed for risk assessments.
Is any of this legally binding on OpenAI or Anthropic?
No. No law forces the labs to accept these conditions — the arrangement currently runs on goodwill and public pressure, and the Trump administration has been openly skeptical of AI regulation.
What This Means for You
For Indian readers, this isn't distant Silicon Valley drama. India is framing its own AI rules right now, and whatever template gets settled globally — independent audits, incident disclosure, evaluator access — is what Indian policymakers will borrow from or react against. For working professionals, there's a quieter signal here: AI safety auditing is becoming a real, funded career path, and the world suddenly needs far more people who can rigorously test these systems than exist today. That's a genuine opportunity for India's engineering talent. And whether you end up auditing AI or building with it, the practical first step is the same — actually learning how these models work. That's exactly what our GPT-6 Astra — Zero to Hero Guide is for, from prompting fundamentals to agents, cost control and safety.