The AI standards body Hassabis proposes could coordinate a slowdown among the frontier labs
Google DeepMind launched an essay site the same day, and the proposal arrived on it under a note saying it isn't Google's official view.
Google DeepMind opened a new essay site on September 16, 2026, the DeepMind Institute. Demis Hassabis used his own piece on it to name the regulator he wants the US to go and build.
He calls it a Standards Body. In his essay it's a federally overseen public-private partnership or self-regulatory organization, and the comparison he picks is the Financial Industry Regulatory Authority. Labs would hand models over for testing before they ship.
- Who
- Demis Hassabis, writing on the DeepMind Institute
- Proposal
- a new Standards Body, a federally overseen public-private partnership
- Pre-release review
- up to 30 days before release, voluntary at first
- Funding
- would need to be substantial and likely mostly come from industry
- Directors
- Shane Legg, James Manyika, Demis Hassabis
How the testing would work
The body writes the benchmarks and refreshes them, quarterly to start, dropping the ones that get saturated. Clear the thresholds and your model is Frontier-class, which makes you a Frontier Lab: model cards, internal cybersecurity, vetted staff, paid-for safety research. Startups and academics stay out of it.
The first version runs on trust.
Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release
"Initially" is doing a lot of work there. To be fair, he does say what ends it: once the protocol has been shown to work, passing it becomes the condition for selling a frontier model in the US.
The tests would cover cybersecurity and biological threats, and they'd look for a model slipping its own guardrails or showing signs of deception. The labs would help write the first ones (which is the awkward bit), and then the body would build up to holding tests back from them. That's how you stop a lab training against the exam.
The sentence that lands in this week's argument
Almost all of it is about testing. One clause isn't about testing at all, and it answers what Dario Amodei asked the industry for on September 12, 2026, in We Must Pace the Frontier.
could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary
Four days earlier, replying to Amodei, Hassabis had pointed at a standards body without saying what one would actually do.
Dario's essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment. This is also why we recently put out our proposal for an industry-wide standards body for frontier AI.
Now there's a document, and its reach is what I'd expect rival labs to fight about first. It would cover any Frontier-class model whatever country built it, open or closed.
For 20+ years @ShaneLegg and I've discussed AGI’s potential impact on the economy, science & society. With the DeepMind Institute, we're expanding interdisciplinary research on key questions for the AI era. We hope it spurs the discussions needed to get the next steps right: https://t.co/Bqf10G3Xep
Who runs it
Legg, James Manyika and Hassabis are the directors (Legg is listed twice, as a director and as the managing editor). The site opened with a handful of essays, on reasoning transparency, economic policy and utopianism, and it says up front that its writers won't always agree and will probably change their minds.
A line at the foot of every piece keeps Google at arm's length from all of it.
DMI pieces are intended to serve as conversation starters, reflecting the author's ideas and research, and should not be read as Google's official view
So the plan for how America should police frontier models turns up under a note saying it isn't the official view of the company that builds one. Hassabis puts no date on any of it.