Google DeepMind’s Demis Hassabis Wants A New AI Watchdog. He Says The U.S. Should Lead It.
As governments race to keep pace with increasingly powerful artificial intelligence systems, one of the industry’s most influential executives is urging the United States to establish a formal oversight body to evaluate the world’s most advanced AI models before they reach the public.
The proposal comes amid heightened scrutiny of frontier AI following recent government intervention in advanced model releases and growing geopolitical competition between the United States and China over leadership in artificial intelligence. AI governance has also become a larger national security issue as governments examine the technology’s implications for cybersecurity, biological research and critical infrastructure.
Google DeepMind co-founder and Chief Executive Demis Hassabis has called on the United States to create a new AI watchdog capable of reviewing frontier AI models before deployment and coordinating industry-wide action if significant risks emerge, Axios detailed on Tuesday.
Hassabis outlined the proposal in a manifesto titled “A Framework for Frontier AI and the Dawning of a New Age.” In an interview with the outlet, he said the industry has reached a point where a more systematic approach to AI oversight is needed. He proposed an industry-funded organization staffed by technical experts that would operate under U.S. government oversight.
The executive said the current AI-related cybersecurity risks should be treated as an early warning, adding that more advanced cyber, biological and nuclear capabilities could become available through future frontier AI systems. Hassabis told Axios that his concerns extend primarily to increasingly capable proprietary frontier models, although the proposal would also cover qualifying open-source systems.
Hassabis has spent several months discussing the proposal with officials in the Trump administration, executives from other leading AI laboratories and European policymakers before making the plan public. He said feedback from those discussions had been positive and argued that the new oversight body should become operational within months.
The proposed organization would resemble the Financial Industry Regulatory Authority (FINRA), the self-regulatory body overseeing much of the U.S. securities industry. Under Hassabis’ proposal, frontier AI developers would initially submit their models voluntarily up to 30 days before public release so independent experts could evaluate cybersecurity, biological and deception-related capabilities.
Once the testing framework is established, frontier AI systems would be required to pass those evaluations before they could be deployed in the United States, according to the proposal. Hassabis said the governing board should include independent technical experts, Turing Award recipients, government representatives, industry participants and members of the open-source community.
The proposal comes after several high-profile AI regulatory developments in Washington. Last month, OpenAI delayed the broad public rollout of GPT-5.6 while working with the U.S. Department of Commerce on safety testing and government review, a process the company completed before releasing the model more widely. CNBC previously reported that the Commerce Department had cleared the rollout following those discussions.
Hassabis also cited the Trump administration’s temporary restrictions on Anthropic’s Mythos and Fable models as evidence that the United States needs a more structured oversight framework instead of relying on case-by-case government action. Axios reported that Anthropic spent more than two weeks negotiating with federal officials before restrictions were lifted.
Hassabis also said that any oversight framework should apply equally to frontier AI models regardless of whether they are developed in the United States or overseas, or whether they are open-source or proprietary. The proposal also calls for regularly updated benchmarks to determine which systems qualify as frontier models.
