Demis Hassabis, the Nobel laureate running Google DeepMind, is urging the U.S. government to build a new regulatory body tasked with screening the world's most advanced AI systems and potentially coordinating industry slowdowns if risks spike.
Hassabis laid out the vision in a manifesto published Tuesday and detailed his thinking in conversations with news outlets. The core argument is straightforward: current regulation is ad hoc and fragile, leaving the U.S. vulnerable as AI capabilities accelerate toward systems that could pose biological, nuclear, or cyber threats.
The timing is urgent. Hassabis warned that within 18 months, dangerous capabilities could spread into open-source models beyond government reach. He has spent months behind the scenes building support, briefing the Trump administration, rival AI labs, and European officials. Early signals from Washington have been encouraging, he said.
The proposed structure mirrors FINRA, the private watchdog that oversees Wall Street under federal supervision. Frontier AI labs would voluntarily submit their most powerful models up to 30 days before release for safety testing that examines cyber, biological, and deception risks. Once the testing regime proves reliable, Hassabis writes, requirements could shift from voluntary to mandatory.
A majority-independent board stacked with Turing Award winners and credentialed technical experts would oversee the work, with seats for industry, government, and open-source representatives. The watchdog would maintain power to update benchmarks as capabilities evolve, and the rules would apply to frontier-class models regardless of origin or licensing model.
Hassabis envisions the frontier designation itself as a mark of status. Labs that pass the tests would carry a prestige credential in the market.
His timeline is tight. Ideally the body would be operational before year-end, with major lab leaders already aligned on the direction in principle. The recent Trump administration crackdowns on Anthropic's models and OpenAI's compliance deals served as wake-up calls, Hassabis said. Those incidents exposed how Washington lacks established protocol for regulating breakthrough AI, forcing each case to be negotiated from scratch.
Anthropic's CEO Dario Amodei has made his own push for binding regulation modeled on the FAA, which has authority to block unsafe aircraft. The two tech leaders now agree the U.S. must regulate advanced models. Their disagreement centers on which agency should hold that power.
Hassabis believes artificial general intelligence, a system matching human cognitive ability across all domains, is probably only years away. His core conviction is that the window to shape how this technology unfolds is closing fast.
Author James Rodriguez: "Hassabis is right that ad hoc crackdowns are no substitute for real rules, but the speed he's demanding and his faith that labs will self-regulate voluntarily before mandatory testing kicks in feels optimistic."
Comments