Google DeepMind has launched an institute to widen the debate about AGI, and its first collection of essays goes straight at regulation. Demis Hassabis proposes a US-regulated standards body to assess the most advanced AI models, starting with developers voluntarily submitting them for review no later than 30 days before release. In a separate essay, DeepMind safety researchers Rohin Shah and Anca Dragan argue that the window into a model's step-by-step reasoning is narrowing, and that nothing about this is inevitable.
The institute opens with a disclaimer rather than a position. Its founders note that Google, Google DeepMind and researchers around the world will not always agree with each other, and that views may shift as new data and information arrive in a fast-moving field. The first collection covers four subjects: economic policy for managing the consequences of a possible spread of AGI, keeping model reasoning in a form humans can follow, principles of human wellbeing, and a framework for evaluating frontier AI models.
The transparency essay is the more technically specific of the two. Shah and Dragan are writing about the ability to see and check how a model got to an answer, step by step, and they treat its erosion as a design decision rather than a law of nature. New architectures make the most powerful models harder to observe, so developers and regulators should, in their view, weigh the safety trade-offs directly instead of absorbing them by default. One concrete option they raise is capping what they call opaque sequential depth: the amount of sequential computation a model may perform without producing a reasoning trace a human can read. Another is shifting the burden of proof, requiring developers to demonstrate that a less transparent system can still be controlled as effectively as the one it replaces.
Hassabis's proposal is structured in stages. The body would first develop evaluations jointly with AI companies, then move to independent closed testing — held-out evaluations, designed so labs cannot tune models against benchmarks they already know. If the evaluation system proves effective, passing the tests could become a condition of deploying frontier models in the US, rather than a voluntary courtesy. Hassabis also noted the framework could be tightened progressively if needed, and named a coordinated slowdown of frontier AI development across several companies among the possible measures.
The timing is not accidental. Industry safety discussion has been moving from general statements about risk toward specific proposals on disclosure, external verification and coordinated slowdown if safeguards fail to keep pace with the technology. That shift accelerated this week after company executives backed individual elements of Anthropic CEO Dario Amodei's call for a slower pace of frontier development.
My read is that the disclaimer at the top does more work than it appears to. An institute that states in advance that Google may disagree with its own authors is a venue, not a commitment. Nothing in this collection binds Google DeepMind to submit a model for review 30 days before release, or to cap opaque sequential depth in anything it ships. Publishing a proposal that would constrain you is cheaper than adopting it, and it buys the same reputational credit — with the added benefit that the company shaping the design of the standards body is the one that would be standing in front of it.
The phrase that should draw the most scrutiny is the one presented most casually. A coordinated slowdown of frontier development by several companies is, stated plainly, competitors agreeing to restrain output together. That may be the right answer if safeguards genuinely lag capability, but it is also the arrangement competition regulators exist to stop, and the proposal says nothing about how one would be distinguished from the other. Amodei's call has now been endorsed in pieces by several executives; the mechanism by which they would actually slow down together is still missing from every version of the argument.
There is a sequencing problem too. The stage that matters — held-out evaluations run independently — comes second. The stage that arrives first has the body building its tests alongside the labs it is meant to test, which is the capture risk the held-out design is there to fix. And the essay sets no date for when the voluntary phase ends. It ends when the evaluations prove themselves, a judgement the evaluated help define.