The man who built one of the world's most powerful AI systems would like a regulator to oversee AI systems, please. He'd also like that regulator to be funded by the industry it regulates. Demis Hassabis, CEO of Google DeepMind and Nobel laureate, published a personal manifesto Tuesday calling for a new U.S. AI watchdog with the power to screen frontier models and, if things get bad enough, coordinate an industry-wide slowdown.
What Hassabis Is Actually Proposing
According to Axios, which spoke with Hassabis exclusively ahead of the manifesto's release, the DeepMind boss is calling for a "systematic" approach to AI regulation. The proposed watchdog would have real teeth: the power to screen the world's most advanced AI models before they go public, and the authority to pump the brakes across the entire industry if the risk calculus tips the wrong way.
The document, titled "A Framework for Frontier AI and the Dawning of a New Age," is framed as a personal statement rather than a Google corporate position. That distinction may matter legally. It almost certainly does not matter practically, given that Hassabis runs one of the two or three most consequential AI labs on the planet.
This is not a fringe voice issuing a crank warning. Hassabis won the Nobel Prize in Chemistry in 2024 for his work on protein structure prediction. He is about as close to the scientific establishment's idea of a legitimate AI authority as anyone alive. When he says the moment has come for more oversight, people listen.
The Obvious Problem Sitting in the Middle of This
Here is where it gets uncomfortable. The proposed watchdog would be funded by the AI industry itself. Staffed by technical experts, yes, but funded by the companies building the systems being regulated.
Let's think about what that looks like in practice. Google, OpenAI, Anthropic, and Meta write the checks. Those checks pay the salaries of the people deciding whether Google's, OpenAI's, Anthropic's, and Meta's products are safe enough to release. The referees are on the teams' payrolls. This arrangement has a name in every other industry, and the name is not "independent oversight."
To be fair, Hassabis is not the first person to propose industry-funded regulatory bodies, and not all of them have been disasters. The Financial Industry Regulatory Authority operates on a similar model in the securities world. But FINRA also spent decades being criticized for being too cozy with the banks it supposedly policed. The AI stakes, by Hassabis's own account, are considerably higher than a bad quarter for a brokerage.
Why He's Saying This Now
The timing here is not random. The AI development race has reached a velocity that even its participants seem to find alarming. Every major lab is shipping frontier models faster than governments can draft legislation about them. The EU AI Act exists, but enforcement is slow and its reach outside Europe is limited. The U.S. under the current administration has moved in the opposite direction, rolling back what modest AI oversight frameworks the Biden years produced.
In that vacuum, the people building the most powerful systems are increasingly the only ones with a coherent opinion about how to govern them. That is genuinely concerning regardless of how sincere their intentions are. Hassabis may be entirely good-faith here. But good faith from one CEO doesn't bind the next one, or the one after that.
There's also a competitive angle that's worth sitting with. A rigorous screening process for frontier AI models, run by technical experts who understand what makes a model dangerous, would be extremely burdensome for smaller labs and startups. It would be considerably less burdensome for Google, which has the legal teams, the compliance infrastructure, and the engineering resources to sail through any review process while a two-hundred-person competitor drowns in paperwork. Regulatory frameworks have a long history of being championed by incumbents who benefit from the barriers they create.
The Slowdown Clause Is the Most Interesting Part
What's genuinely novel in what Axios reports is the coordination mechanism. Hassabis is proposing that this watchdog have the authority to call for an industry-wide slowdown if dangers reach a certain threshold. Not a ban. Not a recall. A coordinated pause.
This is actually a more serious proposal than it sounds. The core problem with unilateral safety commitments in AI is that they don't work if your competitors don't make the same commitment. If OpenAI decides a particular capability is too dangerous to ship and Google doesn't, OpenAI has just handed Google a market advantage for being responsible. Nobody stays in that position for long.
A regulator with genuine coordination authority could theoretically solve that problem. If everyone has to stop, stopping doesn't cost you the race. That's a real insight. The question is whether you can get there through a body that the industry itself funds and staffs, or whether that just produces a very expensive working group that publishes reports nobody has to follow.
The Dingo Take
Give Hassabis credit for one thing: he's at least showing up. The alternative to an AI CEO calling for oversight is an AI CEO saying the whole thing is fine and please leave us alone, which is what most of the industry has been doing since approximately 2023. A Nobel laureate going on record that the technology he's building requires systematic oversight is not nothing.
But the industry-funded structure is a tell. Real regulatory independence costs something. It means inspectors who can show up unannounced. It means fines that actually sting. It means disclosure requirements that force you to publish findings you'd rather keep internal. None of that is compatible with a model where the regulated parties hold the purse strings. What Hassabis is describing sounds less like the FDA and more like a very serious-sounding advisory council that meets quarterly and issues white papers.
The charitable reading is that Hassabis genuinely believes some oversight is better than none, and he's trying to build something that the current U.S. political environment might actually accept, because a fully independent agency with subpoena power isn't happening under this Congress. The uncharitable reading is that Google would very much like to lock in a regulatory framework before a competitor builds something that blows up and forces Congress to write one in a panic. Both things can be true. They probably are.
