Deepmind CEO Hassabis proposes new US standards body for advanced AI oversight

Google Deepmind CEO Demis Hassabis has published a detailed framework proposal for governing advanced AI, arguing that the development trajectory demands immediate action despite—or perhaps because of—the uncertainty surrounding where the technology is headed. Hassabis maintains that artificial general intelligence is likely just a few years away, repeating a claim he made in April that the impact could be ten times greater than the Industrial Revolution and arrive ten times faster. At the same time, he acknowledges that progress is outpacing humanity's understanding of the technology; in May, he described humanity as already "in the foothills of the singularity," a statement that provoked debate.

At the center of Hassabis's proposal is a new US standards body modeled after FINRA, the financial industry regulator. The agency would develop evaluation protocols for frontier AI models, beginning on a voluntary basis with the expectation that participation would later become mandatory. Industry would fund the body, which would employ regularly updated benchmarks to assess frontier models. Critically, Hassabis proposes that the international community would then need to follow suit and reach consensus on the most vital governance points. If circumstances demand it, the standards body could also coordinate a slowdown in development—an option that Anthropic, another major AI company, has recently considered. To preempt accusations of regulatory capture (where established firms use regulation to suppress smaller competitors), Hassabis explicitly exempts non-frontier models created by startups and academic institutions from the proposed oversight regime.

Hassabis situates this proposal within a landscape of deep expert disagreement. "Nobody in the world knows for sure what is going to happen from here, and even the experts disagree," he writes. "When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy." That expert disagreement was on vivid display in December when Hassabis himself clashed publicly with Yann LeCun, a deep learning pioneer, over whether general intelligence based on language models is achievable. LeCun called the idea "complete BS" and "completely delusional," while Hassabis responded that LeCun was "just plain incorrect." Other researchers have offered nuanced positions: Oriol Vinyals, a Gemini co-lead, suggests that today's models excel in narrow domains but lack the ability to truly innovate. Richard Sutton, another deep learning pioneer, holds a similar view and has announced his startup Oak Labs to pursue that goal. Meanwhile, Deepmind co-founder Shane Legg believes that a "minimal AGI" could be possible as early as 2028—a timeline that adds weight to Hassabis's argument that governance frameworks should be in place now. The timing of Hassabis's proposal follows a letter from prominent AI researchers and economists warning of potentially sweeping consequences from widespread AI-driven job losses, underscoring that questions of AI governance and societal impact are now front and center in both industry and research circles.

Hassabis's proposal arrives at a moment of heightened debate within the AI research community about the trajectory and consequences of advanced AI development. His statement that "nobody in the world knows for sure what is going to happen from here" reflects genuine uncertainty among experts—a point underscored by the public disagreement between Hassabis and prominent AI researcher Yann LeCun over whether language models can achieve general intelligence. LeCun called the concept "complete BS," while Hassabis pushed back publicly. Other voices, including Gemini co-lead Oriol Vinyals and deep learning pioneer Richard Sutton, have offered middle-ground positions acknowledging that today's models excel in narrow domains but lack true innovation capability. Deepmind co-founder Shane Legg's assertion that a "minimal AGI" could arrive as early as 2028 adds urgency to the governance question. The timing of Hassabis's framework proposal follows a letter from prominent AI researchers and economists warning of sweeping job-loss consequences from massive AI deployment, suggesting that governance and safety considerations are now part of mainstream industry and research discourse.