Hassabis Pitches a US-Led Referee for Frontier AI

Demis Hassabis, the chief executive of Google DeepMind, has laid out a plan for who should be checking the most powerful AI systems before they reach the public. In a framework published on 14 July, he called for an independent standards body to evaluate frontier models, and he named the United States as the country that should host it. TechCrunch reported the proposal, which Hassabis set out in a document titled A Framework for Frontier AI and the Dawning of a New Age.

His model is not a new government agency. Instead he points to the Financial Industry Regulatory Authority, the private body that polices Wall Street brokers under federal oversight, as the template. A frontier AI equivalent would be a public-private partnership run by technical experts rather than civil servants, with the industry funding and staffing much of the work while the government keeps a hand on the wheel.

What it would test

Under the plan, evaluations would concentrate on the risks that keep safety researchers up at night: cybersecurity, biological threats, and what Hassabis calls agentic risks, meaning a model's attempts to slip its own guardrails. Those assessments would be refreshed every quarter rather than done once and filed away, as CNBC reported. Labs building the largest systems would voluntarily hand their models to the body up to 30 days before release, and the board itself would draw on independent researchers and open-source representatives.

The framework would apply to any model that meets the frontier threshold, whether it was trained in California or Hangzhou and whether its weights are public or locked away. Smaller models from startups and university labs would be exempt, a carve-out meant to keep the compliance burden off the people least able to carry it.

A familiar tension

The proposal lands in the middle of an unresolved argument about who governs AI and from where. Hassabis is betting that a light-touch, industry-led body hosted in the United States can move faster than treaties and win the trust of the companies it oversees. Critics will note that asking the frontier labs to help design their own referee has obvious risks, and that a US-anchored body sits awkwardly next to calls for genuinely international oversight. Delegates at the United Nations first global dialogue on AI governance pressed for exactly that kind of cross-border coordination, and China has been building out its own governance agenda around events like the World Artificial Intelligence Conference.

Voluntary sharing is the other question mark. The plan leans on labs choosing to submit their models early, with no obvious penalty for those that decline. Regulators in Washington have started reaching for firmer tools, including a recent FTC move to police how AI models are steered. Whether a FINRA-style body would satisfy them, or simply become a venue the strongest players learn to manage, is the part Hassabis cannot answer on his own.