DeepMind's Hassabis Wants an AI Referee — Modeled on Wall Street's Watchdog

Demis Hassabis, CEO of Google DeepMind, has published a detailed proposal for how to govern advanced AI — and his model is borrowed from finance. He wants a new US standards body patterned on FINRA, the industry-funded organization that supervises stock brokers, to act as a referee for the most powerful AI models.

The body would develop evaluation protocols for frontier models — the handful of most capable systems from the biggest labs — starting voluntary and later becoming mandatory. It would be funded by the industry itself and rely on regularly updated benchmarks, meaning standardized tests of what models can and can’t do safely. If those evaluations turned up something alarming, the body could coordinate an industry-wide slowdown in development. Two design choices stand out. First, startups and academic research models would be exempt, which heads off the accusation of “regulatory capture” — big companies using regulation as a ladder they pull up behind themselves. Second, Hassabis frames the whole thing around uncertainty rather than doom: “Nobody in the world knows for sure what is going to happen from here, and even the experts disagree,” he writes. When stakes are high and certainty is low, “cautious optimism” — moving forward while building guardrails — is his prescription.

What’s behind this? The timing is notable. Days earlier, more than 200 economists and AI researchers, including 16 Nobel laureates, signed a letter warning that the window to prepare for AI’s economic impact is closing. Hassabis didn’t sign it — but his proposal reads like a concrete answer to its abstract alarm. He has also repeated his view that artificial general intelligence, AI that matches humans across most tasks, is likely just a few years away. Worth keeping expectations grounded, though: experts genuinely disagree here. Yann LeCun has called language-model-based general intelligence “complete BS,” and Turing Award winner Rich Sutton just founded a lab premised on the idea that current methods need fundamental rework. A proposal from one lab CEO is a starting bid, not a plan with political momentum — and a body that could slow down development would need buy-in that doesn’t exist yet.

What this means for you: Nothing changes today. But this is a useful signal: the people building the most powerful AI systems are now publicly designing brakes for their own industry, and arguing about which institutions should hold them. If AI ends up with a FINRA-style watchdog, decisions about what models can do will be made by test protocols and standards bodies — largely out of public view. Knowing these debates exist is the first step to having a say in them.

Sources

• Hassabis via X
• The Decoder