Demis Hassabis, CEO of Google DeepMind, has proposed creating a global AI oversight institution with the power to evaluate frontier models before release and coordinate industry slowdowns if deemed too dangerous. The organization, which Hassabis wants led by the United States, would include independent experts and open source community representatives.

According to Axios, Hassabis has spent months building consensus around the proposal, meeting with the Trump administration, other AI labs and European officials, and aims to have the organization operational by the end of 2026. He said he has received “very positive” signals from the White House.

Why it matters to you. If the initiative takes shape, those developing or using frontier models will face more structured governance: mandatory evaluations before release, international safety standards and possible coordinated blocks. The key question is who decides what counts as “too risky” and by what criteria—still an open question.

Hassabis signed a statement in June calling for stronger protections against AI-assisted bioweapons production, a sign that the proposal isn’t isolated rhetoric but part of a coherent strategy.

In detail

The context

There is no global AI-specific regulation yet, nor a comprehensive national framework in the United States. Efforts so far have been fragmented: the EU’s AI Act focuses on high-risk applications but doesn’t directly govern frontier model development; US executive orders under Biden introduced reporting requirements for models above certain capability thresholds, but without a dedicated institution to enforce them.

Hassabis proposes a model similar to the Financial Industry Regulatory Authority (FINRA), a self-regulatory organization with real powers: evaluate models before release, define shared standards and coordinate industry pauses when needed. The idea is that the US leads, given its “economic and technical weight,” with involvement from other jurisdictions.

What sources say

Hassabis’s post, titled A Framework for Frontier AI and the Dawning of a New Age, frames artificial general intelligence (AGI) as likely “a few years away” and discusses “singularity” as an imminent horizon. Axios reports the proposal has already found an ear in the Trump administration and other labs, but doesn’t cite public opposition or reservations—which doesn’t mean they don’t exist.

The proposal comes after Hassabis signed, along with others, a statement in June against AI-assisted bioweapons risks, and after economists and industry figures (including Jack Clark of Anthropic and Eric Schmidt) called on governments to take AI’s economic impact seriously.

Implications

If the institution is created, anyone building models above a certain capability threshold will have to submit them for review before public release. The evaluation criteria—what makes a model “too dangerous”—haven’t been specified yet: that will be the crucial issue. The proposal includes open source representatives, but it’s unclear how it would reconcile with models released under permissive licenses and without centralized control.

Timing is relevant: GPT-5.6 just launched, Claude Fable 5 became available again after the export block, and open models like Soofi S or Ornith-1.0 are reaching frontier capabilities. A watchdog with blocking power would change the pace at which these tools reach developers.

What remains open

The proposal doesn’t explain how the institution would handle open-weight models already in circulation, nor how it would enforce slowdowns in jurisdictions that don’t participate. It’s unclear whether safety criteria would be public or confidential, or who would pay for evaluations. And the fact that Hassabis wants US leadership could make international consensus harder to achieve, especially as China and Europe develop their own AI capabilities.