Hassabis proposes U.S.-led AI standards body for frontier models

The proximate cause of Hassabis's proposal is concrete: the Trump administration froze Anthropic's Mythos and Fable models with no established protocol, forcing Anthropic to spend 2.5 weeks negotiating release. Hassabis explicitly called this 'a bit of a wake-up call,' and the episode directly triggered OpenAI to delay GPT-5.6 pending government approval. ^[4]The industry's real fear is not abstract safety risk but arbitrary, uncoordinated government intervention with no rulebook - the proposal is as much about creating legal predictability for labs as it is about protecting the public. By proposing a structured 30-day window with defined criteria, Hassabis is essentially asking the government to trade unpredictable freezes for a process the industry helped design. ^[1]

Hassabis explicitly chose the FINRA model - a federally-overseen self-regulatory organization funded and largely governed by the industry it regulates - over the FAA model proposed by Anthropic CEO Dario Amodei approximately one month earlier. The distinction is structural and consequential. FAA-style oversight means direct government authority with deployment-blocking power; FINRA-style means the industry sets its own standards under federal oversight. ^[3]Critics like Mark Daley at Noetic Engines argue the proposed body dangerously conflates four distinct functions: scientific lab, standards organization, industry regulator, and security council - a concentration that makes it an 'epistemic chokepoint' with authority over which models can enter the global market. ^[7]That Hassabis chose the lighter-touch model while pushing for it to eventually become mandatory for all frontier-class models regardless of country of origin reveals the underlying bet: self-regulation now to prevent harder government mandates later. ^[2]

The structural logic behind Hassabis's proposal is a coordination problem: competitive pressure makes individual safety investment commercially irrational. A lab that voluntarily slows down for safety testing loses market position to one that does not. The proposed Standards Body is designed to solve this coordination failure by making the slowdown universal and the rules identical for everyone - including foreign models seeking U.S. market access. ^[1]This is why the 'slowdown mechanism' - authority to coordinate a halt in AI development across all frontier labs if critical risks are identified - is the most consequential and controversial element. It is the only provision that actually changes the competitive calculus, because it removes the first-mover advantage for unsafe releases. ^[4]Benchmarks refreshed quarterly and exemptions for startups and academics are design choices that reflect this competitive-dynamics framing, not just technical safety considerations.

Hassabis grounded the urgency in a specific technical claim: within 18 months, biological and nuclear capabilities could emerge in open-source AI models, with current cyber-risks described as 'warning shots.' ^[8]This framing - AGI is 'probably only a few short years away' and the impact will be '10x of the Industrial Revolution at 10x the speed' - is doing political work as much as predictive work. ^[6]The proposal's implementation window (end of 2026) maps directly onto that 18-month risk horizon. The timeline argument is what converts the proposal from optional good governance into an emergency measure - and emergency measures attract emergency political support, as the White House's prior executive order alignment suggests. ^[5]