Demis Hassabis wants the United States to stand up a global AI watchdog with the authority to screen frontier models before release and force an industry-wide slowdown if any are judged too risky to deploy. The Google DeepMind CEO laid out the proposal in a personal manifesto published Tuesday morning titled 'A Framework for Frontier AI and the Dawning of a New Age,' and he wants the institution operational before the end of 2026.
The structure Hassabis is sketching resembles the Financial Industry Regulatory Authority: industry-funded, staffed by leading independent technical experts, and including representatives from open-source communities. Membership would carry real teeth — the body could evaluate the most advanced models before they ship and coordinate a collective pause across labs when evaluations flag serious dangers.
Hassabis argues the US is the right host given its economic and technical standing, and that the timeline is tight because AI capability is compounding faster than governance. Artificial general intelligence, he wrote, is probably only a few short years away.
Key facts
- 01Demis Hassabis wants a US-led global AI watchdog operational before the end of 2026.
- 02The body would screen frontier models pre-release and coordinate industry-wide slowdowns if risks warrant.
- 03Hassabis proposes a structure modeled on the Financial Industry Regulatory Authority, staffed by independent experts.
- 04He says artificial general intelligence is 'probably only a few short years away.'
- 05The 2024 Nobel laureate has been quietly briefing the Trump administration, other labs, and European officials for months.
There is currently no global rulebook governing frontier AI, and no comprehensive national framework in the US either. That regulatory vacuum is the gap Hassabis is trying to fill before the models get materially more capable, and before an incident forces a less coherent response.
According to Axios, Hassabis has spent months quietly briefing the Trump administration, rival AI labs, and European officials to build support. He told Axios the reception has been encouraging on the US side.
That signal matters. The Trump administration has so far leaned toward a light-touch federal posture on AI, and a voluntary, industry-funded body led from Washington is a structure the White House is more likely to accept than a top-down statutory regulator.
“the noises I've been hearing [from the Trump administration] are very positive.”— Demis Hassabis, Google DeepMind CEO and co-founder
Hassabis is not the only Google DeepMind figure pushing this line. He signed a statement last month calling for tougher protections against AI-aided bioweapons production, and his framework arrives alongside a separate open letter from top economists and tech leaders — including Anthropic co-founder Jack Clark and former Google CEO Eric Schmidt — urging governments to take the economic disruption from AI seriously. That letter followed earlier coverage on this site of the 200-economist warning about a jobs 'tsunami' from AI.
Hassabis, who shared the 2024 Nobel Prize in Chemistry for AI-based protein structure prediction, is one of the more credible voices to make this pitch. He runs the lab behind Gemini, he has a scientific track record independent of the commercial arms race, and he is proposing rules that would constrain his own product roadmap. That combination is rare among frontier lab CEOs and is part of why the proposal is landing.
The counterweight is straightforward: an industry-funded watchdog staffed by industry experts is, by design, exposed to capture. Critics of self-regulatory bodies point to FINRA itself as a model with mixed results, and a pre-release model review regime run by the labs' own peers will face immediate questions about independence, enforcement authority, and what happens when a company simply declines to submit a model. Hassabis's framework does not yet spell out a legal backstop.
There is also a jurisdictional problem. A US-led body cannot compel Chinese labs, and a coordinated slowdown only works if every frontier developer participates. Absent buy-in from Beijing, the watchdog risks becoming a Western club that slows down its own members while ceding ground abroad — an argument that Washington policymakers are already alert to.
For AI companies, the practical implication is that pre-release safety evaluations are moving from voluntary lab practice toward a shared institutional layer, whether or not Hassabis's specific proposal is the one that lands. Compute-intensive red-teaming, third-party model evals, and disclosure regimes are becoming table stakes for anyone shipping a frontier system, and the labs that build that muscle early will have a lower compliance tax when a body like this — Hassabis's or someone else's — actually starts operating.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
