← Back to feed

Google DeepMind CEO Demis Hassabis has published a framework paper arguing AGI is “probably only a few short years away” and calling for a US-led international body with authority to slow or halt frontier AI development.

The paper marks a significant escalation in Hassabis’s public positioning. Unlike prior calls for voluntary safety commitments, the framework explicitly proposes that frontier models require approval before deployment — and that development itself could be slowed if deemed necessary.

Key claims

• AGI, defined as a system exhibiting “all the cognitive capabilities the brain has,” is likely within a few years
• The coming transition will be “10x of the Industrial Revolution at 10x the speed”
• “Nobody in the world knows for sure what is going to happen from here, and even the experts disagree”
• A global oversight body should be established, led by the US “given its economic and technical standing”
• Humanity has a “precious window” to establish governance before that window closes

The proposed oversight structure goes beyond anything currently in place. The US CAISI framework — which Google, Microsoft, and xAI recently joined — covers voluntary pre-release testing. Hassabis is describing something closer to an international regulatory mandate with mandatory pre-deployment approval gates.

The competitive subtext

The timing is notable. DeepMind trails Anthropic and OpenAI on the latest public benchmarks — Fable 5 leads the AA Intelligence Index at 64.9, GPT-5.6 Sol follows — and Google is managing the fallout from a classified Pentagon AI deal that prompted an internal DeepMind researcher to publish a resignation essay. A framework advocating mandatory oversight would constrain rivals at least as much as it constrains Google.

Hassabis has been more consistent on safety governance than most frontier lab CEOs. But the specificity of this paper — pre-deployment approval, sanctioned slowdowns, a US-led watchdog — moves the argument from aspiration to institutional design.

Whether the White House would lead such a body is an open question. The US has favored voluntary frameworks, and its response to the Mythos export order — granting 100 trusted partners access while blocking broader use — suggests ad hoc gating rather than systematic pre-approval is the current model. Hassabis is arguing that is not enough.

What it would require

A genuine pre-deployment approval body would need to:

• Define what constitutes a “frontier model” subject to review
• Set technical thresholds (METR, CAISI classifications, AISI evaluations)
• Agree on what failure conditions warrant a delay or halt
• Build enforcement capacity across jurisdictions

None of those are simple. The paper signals Hassabis is willing to advocate for them anyway.