Demis Hassabis, CEO of Google DeepMind, proposed creating a U.S.-led AI regulatory body.
He said on X on the 14th (local time) that only a few years remain until the arrival of AGI. He described the technology’s impact as 10 times greater and faster than the Industrial Revolution, and said it would be more akin to the discovery of electricity or fire than to the internet or mobile devices.
The idea is specific. The model he cited as a template is the U.S. Financial Industry Regulatory Authority (FINRA). Like a private body that regulates Wall Street under the supervision of the Securities and Exchange Commission (SEC), the proposed organization would evaluate advanced AI models and, in cooperation with federal agencies, handle testing in security-related areas such as cybersecurity and biological risks.
Initially, frontier AI labs would voluntarily submit their models for review 30 days before release, and if the system proves effective, submission would become mandatory. It is reported that operating costs would be borne by the industry, and that startups and academic non-frontier models would be excluded from the scope.
Hassabis is said to have coordinated in advance for several months with the Trump administration, European officials, and industry peers, with the goal of launching the body within the year.
The proposal came shortly after prominent AI researchers and economists issued a statement warning about AI-driven mass unemployment. Signatories included Anthropic co-founder Jack Clark and former Google CEO Eric Schmidt. Hassabis did not sign, but his concerns about the technology’s impact are in the same vein.
He created AlphaGo, the AI that defeated Go champion Lee Sedol in 2016, and in 2024 he shared the Nobel Prize in Chemistry for his work on AlphaFold, the AI system that predicts protein structures. In effect, one of the most cautious voices in the industry regarding AGI timing has now publicly put a number on it: a few years.
The striking point of this proposal is that the head of a company likely to be regulated is arriving with the blueprint for regulation itself.
The AI industry has generally been defensive about government intervention. The variable that changed the dynamic was security. The U.S. Commerce Department issued export-control guidance restricting Anthropic’s high-performance model, Claude Mythos, from serving foreign users for a period, citing national security. OpenAI also went through a process of pre-releasing its latest model, GPT-5.6, to government-approved entities before its formal launch.
The trigger was Mythos, which emerged in April. Concerns grew among U.S. security authorities and the financial sector after the model was reported to have the ability to identify security vulnerabilities in computer code and potentially exploit them. That was the backdrop for the Trump administration’s shift from a relatively passive stance on AI regulation.
The executive order signed by President Trump on the 2nd of last month follows the same logic. It requires the government to conduct security checks 30 days before the release of high-performance models, while making corporate cooperation voluntary in principle.
The original draft set the review period at 90 days. Industry opposition, arguing that U.S. companies’ development pace could slow and leave them behind China in the race, reduced it to 30 days. Hassabis’s proposal would institutionalize this voluntary framework through a permanent private body. The logic is clear: industry wants to seize the initiative on setting standards before government regulation fully takes hold.
The Trump administration’s reaction is favorable. Hassabis told Axios that the response he has heard from the administration has been highly positive.
The design, which excludes small developers and academia from regulation, is seen as an effort to avoid criticism of regulatory capture, in which large companies use regulation to squeeze out competitors.
Still, the criticism is serious. There are concerns that a body funded by the industry itself would be judging the risks of industry models, creating an inherent conflict of interest. That leads to questions about whether it could actually force a project to stop once a risk is identified.
Some in U.S. politics continue to argue that a framework relying on voluntary participation cannot adequately control dangerous models and are calling for mandatory regulation. Skepticism also remains over whether U.S.-only leadership can replace the European Union’s regulatory framework or an international agreement.
The signal here is that the center of gravity in AI governance is shifting toward the United States. With executive orders, export controls, and proposals for private standard-setting bodies coming one after another, the U.S. is effectively locking in a situation where it alone designs the standards for frontier model verification.
Hassabis himself said the U.S.-led body would lead to the creation of international standards. If the path from voluntary submission to mandatory compliance becomes reality, AI companies in other countries may have to pass U.S. standards in order to access the market. In that sense, the verification process itself becomes a barrier to entry.
The first priority is to block conflicts of interest. If the new body is to earn trust, the industry may fund it, but the power to make decisions must sit outside the industry.
The starting point would be to form a board that includes independent experts on the level of Turing Award winners, the open-source community, and government officials, while making evaluation criteria and results public so they can be externally verified.
Evaluation benchmarks should be updated regularly to prevent models from learning the tests, but the party responsible for updates must be separate from the companies being evaluated. The body also needs a basis for enforcing recommendations to halt development when necessary.
Just as FINRA operates under the SEC’s legal oversight, a new body would need federal supervision and a link to sanctions to have power beyond mere recommendations. This is an issue the U.S. government and participating companies must resolve at the launch stage.
The next step is legislation. Turning voluntary submission into mandatory compliance requires a legal basis, not just an executive order. Congress should begin reviewing the foundational law, including criteria for models subject to review, penalties for refusing submission, and protections for corporate trade secrets.
That is because an executive-only measure has clear limits if the goal is to build a system that endures regardless of changes in administration. Legal provisions must also spell out confidentiality safeguards and intellectual property protections for model information accessed by the government and the body during the evaluation process if companies are to be brought on board.
South Korea’s response falls into two tracks. In the short term, it should participate in the formation stage before U.S. standards harden into international norms.
The Ministry of Science and ICT and the Ministry of Foreign Affairs should enter the discussion on the proposed body, even as observers if necessary, so they can reflect the interests of domestic companies in the evaluation criteria and procedures, and put mutual recognition of assessment results between South Korea and the U.S. on the agenda to reduce the burden of duplicate verification for Korean firms.
Structurally, the key is to secure its own evaluation capabilities. To gain leverage in mutual recognition negotiations, South Korea must expand the AI Safety Institute’s staffing and computing resources for frontier model evaluation and bring its cyber and biological risk assessment methodologies up to a level compatible with the U.S. body.
It is also necessary to revise subordinate regulations under Korea’s Basic AI Act so they align with the U.S. verification framework. The gap will be determined here between countries that merely copy verification standards and those that can use them together.