- Frontier AI Proposal: Google DeepMind CEO and cofounder Demis Hassabis proposes a US-led, industry-funded standards body for frontier AI.
- Prelaunch Tests: Labs would initially submit models voluntarily up to 30 days before release, with mandatory US deployment tests only after the protocol proves robust.
- Slowdown Power: The proposed body could coordinate an industry-wide slowdown, but it has no government approval or binding authority.
- Year-End Target: Hassabis may want the body operating before the end of 2026, but Washington must first grant formal authority.
Google DeepMind CEO and cofounder Demis Hassabis is proposing a US-led standards body to test frontier AI models before release and coordinate a slowdown if risks rise. Frontier models are leading-edge, general-purpose systems whose capabilities may create new safety or national-security risks. Hassabis is proposing a regulator, not announcing an operating body or enacted US policy.
AI labs would initially volunteer models for a 30-day prelaunch evaluation. Mandatory tests for US deployment would follow only after the assessment process proved effective and robust. Voluntary review would let labs and government evaluators develop common tests before review could become a market-access condition affecting which advanced systems launch in the United States.
Hassabis has spent months briefing officials and technology firms, including the Trump administration and other AI labs. Washington has not formally endorsed the proposal, and the body has no binding authority.
Hassabis recommends to create a body funded by AI companies and operated independently, with US government backing and a majority-independent board. Separating technical judgments from the companies providing the money is central to the design.
How the Proposed AI Referee Would Work
Hassabis models the institution on the Financial Industry Regulatory Authority (FINRA), an industry-funded Wall Street regulator overseen by the US Securities and Exchange Commission. Independent technical experts, government officials, and representatives from open-source communities would join its board. AI labs would initially submit frontier models voluntarily, with a common mandatory test for US deployment following only after the assessment protocol proved effective and robust.
Evaluators would receive models before public deployment and test for dangerous autonomous capabilities, guardrail bypasses, deception, cyberattack ability, and biological or nuclear risks. Hassabis estimates that severe capabilities could reach openly available models within 18 months. His estimate helps explain the emphasis on prelaunch access, when testers can probe capabilities and developers can still delay deployment.
Earlier in 2026, Hassabis warned that competitive pressure could push labs to release systems faster than they can verify safety. Prelaunch access would move some scrutiny ahead of deployment, but a common test would matter only if rival developers accepted the same evaluation and consequences.
The proposed rules would cover foreign and domestic frontier models, whether open or closed. Once the protocol became robust, an independent standards body would test frontier models before US deployment. A model judged too risky could also trigger a coordinated slowdown among participating labs.
Both powers remain conditional because the standards, enforcement route, and legal basis for restricting market access have not been settled. Representatives from open-source communities would need to prevent a gate designed around large private labs from placing different practical burdens on openly released models.
Who Would Watch the Watchdog
Washington already uses a narrower mechanism. The White House early-access plan gives federal evaluators voluntary pre-release access to covered frontier models for cybersecurity review. Hassabis suggests broadening that approach to additional risks and eventually add mandatory deployment tests.
Moving from early access to a binding gate would require governments and competing labs to accept common standards, independent judgments, and consequences for a failed evaluation. AI-company funding creates the proposal’s hardest governance problem because the body would police the companies paying for it.
A majority-independent board could reduce that conflict only if tests were transparent, outcomes enforceable, and participation broad enough to stop one lab gaining an advantage by rejecting a slowdown. US oversight would also need to constrain funders without turning technical judgments into political instructions.
Hassabis argues that artificial general intelligence, a still-hypothetical system matching the breadth of human cognitive abilities, may be only a few years away. He recommends the proposed body operating before the end of 2026. Washington would have to formally authorize it before voluntary review could become a binding gate for US deployment.
