Google DeepMind CEO calls for US to lead regulation as threats from AGI nears

Demis Hassabis, the cofounder and CEO of Google DeepMind, wants the US government to set up a new watchdog that will be tasked with screening the world’s most powerful AI models. The watchdog will also have the power to order the industry to slow down if the danger grows.

He detailed this in his manifesto, which was published on July 14, and said that he wants the body running before year-end.

Hassabis is one of the leading voices in the AI space. He shared the 2024 Nobel Prize in Chemistry for his protein-folding work and runs the lab behind Gemini.

In an interview, he stated that the industry has reached the point where AI oversight needs to be systematic, funded by the labs themselves, staffed by technical experts, and answerable to the US government.

What will the US AI watchdog do?

Hassabis is modeling the body on FINRA, the private, industry-funded regulator that polices Wall Street brokers under the Securities and Exchange Commission’s authority.

In his plan, frontier labs would first hand over their models voluntarily, up to 30 days before release, so the body could test them for dangerous cyber, biological, and “deception” capabilities.

He stated that once that testing proves reliable, formal rules “could quickly follow,” meaning a model would have to pass before it could be deployed in the US market.

The standards would apply to every frontier-class system, and Hassabis added that “no matter their country of origin or whether they are open or closed,” with the benchmarks updated as the technology advances.

He also wants the board to be majority-independent, filled with Turing Award winners and other credentialed researchers, alongside seats for industry, government, and open-source developers. He stated that the “frontier” label would become a mark of status, adding, “I think that’s a pretty nice, prestige kind of asset to have.”

Why is Hassabis calling for an AI watchdog now?

Hassabis says that today’s AI-driven cyberattacks are “warning shots.” According to him, within 18 months, more serious biological and nuclear risks could sit inside open-source models that no government can pull back. He added that the threat is not limited to open systems and that the labs’ own future proprietary models will carry the same dangers.

He said he wants the watchdog operational in months and has spent recent weeks briefing the Trump administration, rival lab leaders, and European officials before going public. “The noises I’ve been hearing are very positive,” he said of the White House talks.

That administration had favored a hands-off approach until last month, when it placed Anthropic’s Mythos and Fable models under an export-control order, limiting their access. The order was later removed, with Anthropic releasing the models with more controls on them.

OpenAI, on the other hand, agreed to limit GPT-5.6 to government-vetted partners before releasing it publicly last week after testing with the Commerce Department, a move that was seen as it trying to avoid the same outcome as Anthropic’s.

Hassabis says the incident is “a bit of a wake-up call.”

A rare point of agreement

Anthropic CEO Dario Amodei is also of the opinion that there should be an agency that has the power to rein in unsafe AI models, calling for binding rules and an FAA-style agency.

Amodei’s position is in agreement with Hassabis’; however, the latter favors the lighter, industry-run FINRA structure.

Hassabis’s belief that the stakes are rising fast is worth noting. He has said AGI, a system matching the full range of human cognitive ability, is “probably only a few short years away,” and at a Stanford event, he put the arrival near 2030, “plus or minus a year.” In the manifesto, he wrote that researchers have “found a way to make sand think,” calling it “miraculous.”

If you're reading this, you’re already ahead. Stay there with our newsletter.