Demis Hassabis, the chief executive of Google DeepMind, has laid out a detailed plan for how governments and companies should govern the most advanced artificial intelligence systems, and the centerpiece is a new kind of American regulator built to move at the speed of the technology it watches. In an article published in mid July, Hassabis argued that the arrival of frontier AI models, the largest and most capable systems now under development, demands oversight that legislation alone cannot deliver quickly enough, and he proposed a standards body modeled on an institution most people associate with Wall Street rather than machine learning.
That institution is FINRA, the Financial Industry Regulatory Authority, a private, nonprofit organization funded by the brokerage industry it polices. Hassabis suggested a similar structure for AI, according to reporting from CryptoBriefing, with a dedicated body that would set testing standards for frontier systems and, like FINRA, would most likely be paid for by the companies it oversees. The proposed watchdog would develop evaluation protocols for these models and, in his framing, could even coordinate a slowdown or a pause in development across the field if the risks at a given moment appeared to outrun the safeguards.
The framework Hassabis sketched rests on a few working parts. He called for periodic independent evaluations of what AI models can actually do, repeated as the systems evolve rather than performed once and filed away, so that regulators and labs keep a current picture of capabilities that change from one training run to the next. He paired that with sector specific rules for the domains where a mistake carries the highest cost, singling out areas such as autonomous driving and medical applications, where an AI system’s behavior touches physical safety and human life directly and where a single general set of AI rules would be too blunt to fit.
Behind the proposal sits a view of the stakes that Hassabis has voiced repeatedly. He has described the current moment as a “species level transition,” one that in his words leaves “little margin for error” over the coming decade, language that casts the governance question as something closer to managing a general purpose scientific revolution than regulating a single product category. Earlier in 2026 he warned about “race conditions,” the competitive pressure that can push labs to ship new systems faster than they can verify those systems are safe, a dynamic he has argued the field needs outside structure to resist. A body able to call for a coordinated pause is, in that light, an attempt to give the industry a brake that no single company can pull on its own without falling behind its rivals.
The appeal of the FINRA analogy is speed and expertise. Supporters of that model, Hassabis among them, argue that a specialized, industry informed organization can write and revise technical standards far faster than a legislature can pass and amend a statute, and that the people who understand frontier systems well enough to test them tend to work inside or close to the labs building them. A self regulator staffed with that expertise, the argument goes, can keep pace with a technology that shifts month to month, where a law fixed in place risks being outdated the day it takes effect.
The same design invites an obvious objection. A proposal from the head of one of the world’s leading AI labs, calling for an industry funded body to set the rules for that same industry, raises the question of whether the companies with the most to gain should be the ones writing and paying for their own oversight. Critics of self regulation have long argued that industry run bodies can drift toward rules that are too soft, or that they can be captured over time by the firms they are meant to check, with the regulated quietly shaping the regulator’s priorities. Financial oversight offers examples pointing both ways, which is part of why FINRA is a contested model as much as an admired one. Hassabis’s proposal does not settle that tension so much as put it on the table, and how much independence such a body would keep from its funders is the detail on which the whole idea turns.
What is not in dispute is that the debate is now being led, in part, by the people building the systems. Coverage from The Decoder and Let’s Data Science frames the proposal as part of a broader push by Hassabis to move the conversation from abstract worry toward concrete institutions while the technology is still taking shape. Whether policymakers adopt a FINRA style regulator, a government agency, or something else entirely, the questions Hassabis has raised, about who tests frontier models, how often, and who holds the authority to slow things down, look set to shape AI governance debates well beyond this one article.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →