Over 100 AI Researchers Ask for Outside Watchdogs at Top Labs

More than 100 leading artificial intelligence researchers, including Canada's so-called 'Godfather of AI,' have signed a public letter calling for greater independent oversight of the world's most advanced AI companies, with Canadian experts publicly backing the call.

More than 100 of the world's better-known artificial intelligence researchers have put their names to a public letter arguing that the companies building the most advanced AI systems should not be the only ones checking their work. Among the signatories is the researcher widely described in Canada as the country's "Godfather of AI," and a group of Canadian experts has publicly endorsed the call, according to BNN Bloomberg.

The demand is straightforward in principle and difficult in practice: independent oversight, meaning people outside a company's payroll with the standing and the access to examine what its most capable models can do before and after they reach the public.

What "independent oversight" actually means in an AI lab

Today, the dominant model of AI safety assurance is internal. Frontier developers run their own evaluations — sometimes called red-teaming, where staff deliberately try to make a model produce dangerous output — write their own reports on the results, and decide for themselves whether a system is fit to release. Some publish summaries. Some invite outside testers under non-disclosure agreements that limit what those testers can say afterwards.

That arrangement has an obvious weakness. The party with the most commercial pressure to ship is the party grading the homework. The letter's premise is that other high-consequence industries resolved this long ago by separating the builder from the inspector, and that AI has not.

In practice, an independent watchdog function would need several things that do not currently exist as a package:

• Guaranteed access to models before public release, at a level of depth beyond a consumer interface.
• Legal protection for researchers and employees who report problems, so that disclosure is not a career-ending act.
• Publication rights, so findings are not buried under confidentiality terms.
• Technical capacity on the watchdog side — evaluators who can keep pace with the labs on compute and expertise.

Each of those is contested. Pre-release access touches trade secrets. Publication rights collide with the argument that describing a model's failure modes hands a manual to bad actors. And capacity costs money that regulators, in most jurisdictions, have not been given.

Why the Canadian endorsement carries weight

Canada is not a hyperscaler nation. It does not host the largest training runs and it is not home to the companies whose models sit at the frontier. What it does have is an outsized share of the academic lineage behind modern deep learning, concentrated in Toronto, Montreal and Edmonton, and a long institutional habit of publishing research into the commons rather than locking it up.

That gives Canadian signatories a particular kind of leverage. They are close enough to the field to be credible and far enough from the balance sheets to be hard to dismiss as competitors talking their own book. When the country's most cited AI figure signs a letter of this sort, it lands differently than a statement from a rival developer.

It also puts a policy question in front of Ottawa. If the letter's logic is accepted, someone has to fund and staff the inspecting body, and a mid-sized economy with deep research talent but limited compute is a plausible candidate to host part of that function in cooperation with allies rather than to build it alone.

The letter joins a crowded and unresolved argument

This is not the first collective statement from AI researchers, and the pattern of previous ones is instructive. Open letters have reliably generated attention and rarely generated binding rules. The gap between a signature and a statute has been where these efforts have stalled.

The argument has also hardened into recognisable camps. One holds that voluntary commitments plus internal safety teams are sufficient and that formal oversight would freeze a fast-moving field in place. Another holds that the absence of an outside check is precisely what makes voluntary commitments unenforceable — a promise you grade yourself is not a constraint. The letter sits firmly in the second camp.

What makes this round different from earlier ones is context. Frontier AI is no longer an emerging research area with modest commercial stakes; it is a central driver of equity market performance. The Nasdaq 100 tracking fund (NASDAQ: QQQ) closed at $721.45 on Friday, 18 September 2026, up 0.63% on the day, while the S&P 500 tracker (NYSEARCA: SPY) finished at $761.69, down 0.12%, and the Dow tracker (NYSEARCA: DIA) closed at $515.88, off 0.48%. The tech-weighted index outperforming the broad market on a soft day is the shape the tape has held through the AI build-out — which is exactly why oversight proposals now have a cost attached that they did not carry a few years ago.

What would signal the letter is going somewhere

Open letters are cheap. The tells that this one has traction are concrete and observable.

• Access agreements with teeth. A frontier lab granting a named external body pre-deployment model access on published terms, rather than case-by-case NDAs, would be a genuine break with current practice.
• Whistleblower protection written into law. Statutory cover for AI staff who raise safety concerns is a narrow, achievable measure that does not require regulating model architecture.
• Budget lines. Oversight without evaluation compute is a letterhead. Funding commitments are the difference between an institute and a mailing address.
• Cross-border coordination. Frontier developers operate globally; a watchdog confined to one jurisdiction is a watchdog a lab can route around.

Absent those, the letter functions mainly as a marker — a documented statement by a large slice of the field that the current arrangement is not adequate, available to be cited when something goes wrong.

The question the industry has not answered

The developers' strongest reply is that they already employ many of the world's best safety researchers and that an external body would be slower, less informed and more easily captured by political fashion. There is real force in that. Frontier evaluation is genuinely hard, and a poorly resourced inspector can create the appearance of scrutiny without the substance.

But the reply does not address the structural point the signatories are making. Competence is not the issue; independence is. A safety team that reports to the executives whose bonuses depend on shipping is constrained regardless of how good it is. Until that reporting line changes, or an outside line is added alongside it, the letter's central complaint stands unanswered.

Key facts

• Signatories: More than 100 leading AI experts signed the public letter
• Notable name: Canada's so-called 'Godfather of AI' is among the signatories
• Nasdaq 100 (QQQ): $721.45, +0.63%, close of Fri 18 Sep 2026 20:00 GMT
• S&P 500 (SPY): $761.69, -0.12%, close of Fri 18 Sep 2026 20:00 GMT

Frequently asked questions

What does the letter ask for?

The letter, signed by more than 100 leading artificial intelligence experts, calls for more independent oversight of the world's most advanced AI companies. In practical terms that means checks carried out by parties outside the developers themselves, rather than relying solely on the internal safety teams that currently evaluate frontier models before release.

Who signed it?

More than 100 leading AI researchers signed, including the figure widely referred to in Canada as the country's 'Godfather of AI.' A group of Canadian experts has publicly backed the call. The full signatory list and the specific companies named as targets were not detailed in the initial reporting.

Why does Canada feature so prominently?

Canada hosts a disproportionate share of the academic research lineage behind modern deep learning, centred on Toronto, Montreal and Edmonton. Canadian researchers are close enough to the field to be credible on technical questions but are not employed by the largest frontier developers, which makes their endorsement harder to dismiss as commercially motivated.

How do AI companies currently assess safety?

Frontier developers largely run their own evaluations, including red-teaming exercises in which staff attempt to make a model produce harmful output. They write their own reports and decide internally whether a system is ready to release. Some invite outside testers, typically under non-disclosure agreements that restrict what those testers may publish.

What would independent oversight require in practice?

At minimum, guaranteed pre-release access to models at technical depth, legal protection for researchers and employees who report problems, the right to publish findings rather than bury them under confidentiality terms, and enough compute and expertise on the inspector's side to keep pace with the labs. None of those currently exists as a complete package.

Have similar letters changed anything before?

Collective statements from AI researchers have reliably drawn attention but have rarely produced binding rules. The gap between signatures and statute is where earlier efforts stalled. Concrete signals of traction would include published pre-deployment access agreements, statutory whistleblower protection for AI staff, and funded evaluation bodies with cross-border reach.

Sources

• Canadian experts back call for independent watchdogs at the world’s top AI companies — BNN Bloomberg