Demis Hassabis pushes for international AI oversight body

Hassabis's proposal is more structurally specific than most AI governance calls. The body he describes is modeled on FINRA - the Financial Industry Regulatory Authority, which regulates U.S. broker-dealers as a federally-overseen but industry-funded self-regulatory organization. Under the proposed framework, frontier labs would voluntarily share their most capable models with an independent technical panel up to 30 days before release ^[1]. That panel - staffed with Turing Award winners, open-source representatives, and government voices - would run probes specifically targeting cyber, biological, nuclear, and deception capabilities ^[1]. The 'voluntary' framing has a planned escalation path: once the testing regime is proven effective, passage becomes mandatory for U.S. market deployment ^[1].

The body would also carry unprecedented authority: the power to coordinate an industry-wide development slowdown or pause across all frontier labs if risks cross a threshold ^[2]. Hassabis's manifesto explicitly states it 'could be ratcheted up if the seriousness of the situation demands' ^[1]. Non-frontier startups and academics are exempt, which means the framework's compliance burden falls entirely on the handful of labs already operating at the capability frontier. Hassabis aims for the body to be operational before year-end 2026 ^[3].

The most revealing part of Hassabis's proposal is not the governance mechanism - it's the economic logic he uses to justify it. In his manifesto and public statements, he explicitly frames the AI safety crisis as a structural prisoner's dilemma: any lab that slows down unilaterally to improve safety cedes competitive ground to rivals who don't ^[4]. This is not an observation about bad actors. It is a CEO of one of the three most powerful AI labs publicly confirming that the market structure of his own industry makes voluntary safety impossible.

Hassabis describes it directly: 'We are currently caught in an extraordinarily intense, multi-layered commercial and geopolitical race' ^[4]. The ChatGPT launch in late 2022 was the trigger - it forced a 'code red' at Google and locked the entire industry into product cycles that, in Hassabis's own words, diverted AI from solving foundational science problems like cancer toward quarterly commercial competition ^[4]. A binding standards body would be the escape hatch: if every lab is required to submit to the same pre-release vetting, no single lab loses competitive ground by complying. The governance proposal is structurally inseparable from the market failure argument - without the latter, the former has no justification.

The most alarming specific claim in Hassabis's manifesto is not about AGI. It is about a much nearer-term threat. He warned that current AI-driven cyber risks are only 'warning shots,' and that within 18 months, serious biological and nuclear capabilities could sit inside open-source models beyond any government's control ^[5]. This is a concrete, time-stamped claim from the CEO of a frontier lab, made to U.S. policymakers in private discussions before going public.

The claim carries specific implications for the governance design. Open-source models sit outside any proprietary lab's control once released - meaning a mandatory pre-release vetting regime that applies only to closed frontier models would miss the risk vector Hassabis is most concerned about. His own proposal acknowledges this: it explicitly covers all frontier-class models 'regardless of country of origin or whether they are open or closed' ^[6]. Google DeepMind co-founder Shane Legg separately estimates a 50% probability of 'minimal AGI' by 2028 ^[7], providing a technical data point behind Hassabis's urgency framing. The governance window he describes is not measured in decades - it is measured in months.

The central design tension in Hassabis's proposal has been flagged from two different directions. Tom Wheeler, former FCC chairman and senior fellow at Brookings, argues that company-only standards processes give governments 'cover to appear to be doing something' while lacking substance on marketplace conduct or public interest protections ^[8]. Wheeler recommends expanding participation to include government and civil society, citing FATF and Basel III as enforcement models. Sam Altman, whose own governance proposal runs along parallel lines, warned G7 leaders directly: 'Do not cede your responsibilities to AI labs like mine' ^[8].

Both critiques point at the same structural problem: a body funded by the labs it regulates, tasked with detecting deception in models those labs profit from, faces a structural incentive to avoid recommending a development pause. FINRA itself - the model Hassabis explicitly names - has been criticized throughout its history for industry capture, with enforcement that critics argue systematically under-penalizes large broker-dealers. The former White House AI advisor Sriram Krishnan, while calling the proposal 'a very thoughtful' one, simultaneously stated 'there will not be an FDA for AI' ^[1]- a signal that even sympathetic observers inside the administration frame this as an advisory body, not an enforcement one. Whether the FINRA model produces a genuine watchdog or an industry-legitimized rubber stamp will depend almost entirely on the composition of its independent technical panel - and on whether the 'voluntary then mandatory' escalation pathway ever actually triggers.