DeepMind CEO argues that advanced AI models should be “reviewed” by an international regulatory body before they can be launched, and must clear the process before going public

Google DeepMind CEO Demis Hassabis published a proposal this week, arguing for an international oversight body led by independent experts, modeled after the U.S. Financial Industry Regulatory Authority (FINRA), and requiring frontier AI models to undergo mandatory inspections for up to 30 days before release.
(Backgrounder: If you’re also afraid AI will destroy humanity, you have to read this report《AI 2040: Plan A》)
(Additional context: One more person in the safety camp is gone—OpenAI’s chief futurist resigns, and the mission-alignment team has also been disbanded.)

Table of Contents

Toggle

• A warning sign that puts the proposal on the table early
• Copying Wall Street’s homework: independent, funded by the industry, and upgradable
• The industry nods—what’s hard is getting governments to nod

This week, Google DeepMind CEO published a manifesto titled《A Framework for Frontier AI and the Dawning of a New Age》on his personal Substack, calling for the establishment of an international oversight body independent of both government and corporations. Modeled after the U.S. Financial Industry Regulatory Authority (FINRA), it would conduct pre-release checks—like a “medical exam”—for the most cutting-edge AI models.

The “frontier models” he refers to, put simply, are the newest generation of AI systems with the greatest computational power and whose risk boundaries have not yet been fully understood. Under the proposal, future frontier models that meet the threshold—whether open-source or closed-source, and regardless of which country they were developed in—must be submitted for review first. The review can last up to 30 days, and the board would reserve most seats for independent technical experts at the level of Turing Award winners.

OpenAI CEO Sam Altman called the idea “very thorough,” and Musk also said it is a “good starting point for discussion.” The two rival competitors who usually can’t stand each other have, unusually, sided with the same proposal this time.

A warning sign that puts the proposal on the table early

Hassabis has been brewing this framework for months. The real trigger for publishing it earlier is the Mythos model, released in limited quantities by Anthropic. After it debuted in April this year, it demonstrated cybersecurity offense-and-defense capabilities far exceeding expectations, sparking concerns in the industry and policy circles about a new era of “AI-driven cybercrime.” Hassabis directly called it a “warning sign,” warning that regulators can’t wait for them to catch up slowly. He also plans to fly to Washington in person next week to meet face to face with U.S. policymakers about the framework—clearly not just trying to post an article and be done with it.

In addition to the cybersecurity risks at hand, Hassabis also mentioned another, longer-term concern in interviews: biosecurity risks. Put simply, it refers to the ability of AI models to help design pathogens, biological weapons, and other capabilities that harm public health. He acknowledged: “The models are not strong enough to that extent yet, but if we extrapolate from current progress, we may see it within a few years.”

At the same time, the U.S. government’s stance toward reviewing new models has tightened nerves across the entire industry. Under pressure from the Trump administration, both Anthropic and OpenAI have delayed large-scale releases of their latest products. This move caught other countries that rely on U.S. technology off guard and also led to negotiations between the European Union and the U.S. government over model access rights. Meanwhile, AI company CEOs also seized the occasion of the France G7 summit to meet directly with leaders from different countries and discuss the matter.

Copying Wall Street’s homework: independent, funded by the industry, and upgradable

The reference model Hassabis chose is the U.S. Financial Industry Regulatory Authority (FINRA). In simple terms, it is a self-regulatory organization funded and run by Wall Street broker-dealers themselves, but which accepts oversight by the U.S. Securities and Exchange Commission (SEC). It audits whether financial institutions follow the rules, and when something goes wrong, it can issue fines and suspend privileges. He wants to transplant the same logic into AI: funded by the industry, but with decision-making power entrusted to people who are not affiliated with any one company.

In terms of specific design, the new organization’s board would have independent technical experts as the majority. Members would include heavyweight figures on the level of Turing Award winners, and additional seats would be reserved for representatives from industry, government, and open-source communities. The funding would come entirely from for-profit top AI labs. The entities subject to review would be all models that meet the frontier threshold—regardless of where a company’s headquarters is, and whether the model is open-source or closed-source—treated equally. The review would take up to 30 days, and the threshold itself would also be updated periodically as model capabilities evolve.

Hassabis emphasized that the strength of this mechanism can be scaled up depending on how severe the situation is. In the most extreme case, it could even coordinate multiple labs together to slow down research efforts. His ideal timeline is that he hopes the organization can start operating as early as possible before the end of this year.

The industry nods—what’s hard is getting governments to nod

Before the public announcement, Hassabis had already privately shown the draft to nearly all major AI labs, including people he calls friends—such as Anthropic CEO Dario Amodei.

Amodei had suggested earlier that the U.S. should set up a dedicated agency similar to the Federal Aviation Administration (FAA) to regulate AI, and Altman has also proposed a similar international oversight initiative. In a Bloomberg interview, Hassabis said: “We talked about a lot, and I believe it is highly compatible with the ideas that Dario and Sam have been putting forward all along. We just need a concrete plan—not an abstract concept.”

As a Nobel Prize winner who has spent more time than most in AI research, Hassabis indeed has fewer enemies than Musk or Altman, and it is easier for him to build consensus within the industry. But an unnamed AI industry leader reminded that industry consensus alone isn’t enough—the real problem is political: to date, the U.S. Congress has not passed any meaningful federal AI legislation, and the Trump administration’s stance toward AI has been swinging between laissez-faire and intervention. The regulatory paths among the U.S., the European Union, and China also often diverge. Hassabis is optimistic that the momentum right now may be enough to truly carry this all the way past the finish line. But he also knows: “What comes next is to figure out how to actually implement it—not just publish a blog post.”