Demis Hassabis wants a US-led AI watchdog with brake power

The U.S. government spent the past month improvising its way through bans on Anthropic's Mythos and OpenAI's Sol, earning widespread backlash for technically inept, closed-door decision-making. Now, one of the leading figures in AI is proposing a permanent playbook.

In an essay titled “A Framework for Frontier AI and the Dawning of a New Age,” Google DeepMind CEO Demis Hassabis called for the creation of an independent, U.S.-led standards body to oversee frontier model releases. He wants the watchdog up and running before the end of the year.

A FINRA playbook for frontier models

Hassabis is taking inspiration directly from Wall Street. Instead of an FDA-style government bureaucracy, he wants an industry-funded, independently operated self-regulatory organization modeled after the Financial Industry Regulatory Authority.

That framing is strategically tailored. White House AI advisor and a16z partner Sriram Krishnan recently dismissed any federal agency expansion, declaring that “there will not be an FDA for AI.” Hassabis's pitch sidesteps that political roadblock by keeping the body outside the executive branch while giving it federal backing.

Under the proposal, frontier labs would initially submit their models voluntarily for safety testing up to 30 days before public release. Independent technical experts and open-source representatives would screen systems for deceptive tendencies, bioweapons risks, and cyberattack capabilities, with specialized audits potentially outsourced to safety groups.

Once the testing protocol proves effective, Hassabis envisions codifying it into law. Models would be legally required to pass the evaluation before deploying in the U.S. market, with labs coordinating on post-release vulnerabilities. Most notably, the body would hold the authority to coordinate an industry-wide slowdown if a release is deemed too dangerous.

The ticking clock toward AGI

Hassabis has spent months quietly pitching the plan across Washington, other frontier labs, and European governments. Speaking to Axios, he noted that feedback from the Trump administration has been very positive.

The urgency stems from his aggressive timeline for artificial general intelligence. Hassabis warned that AGI is likely only a few short years away, claiming humanity is standing in the foothills of the singularity. He told Axios that open-source models could venture into truly dangerous territory within the next 18 months.

The push follows a separate call Hassabis co-signed last month targeting AI-assisted bioweapon development, alongside warnings from figures like Eric Schmidt and Anthropic's Jack Clark regarding economic disruption. The consensus among lab chiefs is increasingly clear: informal oversight is no longer sustainable.

Why it matters

For everyday developers and end users, this proposed body will not disrupt typical API calls or application building anytime soon. The scrutiny is trained specifically on apex frontier systems. However, if the 18-month projection for open-source risk holds true, open weights could soon face heavy compliance hurdles that slow down the open ecosystem.

Across the industry, this is a clear preemptive strike. Rather than waiting for lawmakers to draft clumsy, heavy-handed restrictions, DeepMind is seizing the pen to draft the rulebook itself. Framing the organization around technical expertise is clever, offering labs an escape from arbitrary political crackdowns.

Yet the structural conflicts are hard to ignore. An oversight body funded directly by the very corporations it monitors will always face skepticism around genuine independence. Furthermore, the power to mandate industry-wide slowdowns could easily become a defensive moat, allowing established giants to throttle the pace whenever competitive pressure mounts.

We think self-regulation modeled after financial watchdogs is probably the most realistic compromise on the table right now. But unless its evaluation benchmarks and testing results are subject to rigorous public transparency, it risks turning into a private club where tech monopolies decide what the public is allowed to use.