Policy
DeepMind Standards Plan Puts Frontier AI Regulation On A FINRA Track
Demis Hassabis has called for an independent standards body to review frontier AI systems before release, sharpening the debate over whether powerful models should be governed by ad hoc government pressure or a standing technical institution.
By Michael C ·
Google DeepMind chief executive Demis Hassabis has moved the frontier AI regulation debate from a question of emergency intervention to a question of institutional design. In a proposal reported by TechCrunch and discussed across the AI policy community, Hassabis called for a U.S.-led standards body that could review the most capable AI models before release, develop technical assessment protocols and eventually become a gate that frontier labs would have to pass before deploying their strongest systems in the American market.
The proposal lands after a series of opaque release fights around powerful models. TechCrunch previously reported that OpenAI's Sol release and Anthropic's Fable access dispute drew government attention without a clear public account of who evaluated the models, which tests were used or what standard counted as safe enough. Hassabis is arguing that those decisions should not be improvised each time a new model is ready. They should be routed through a technically staffed institution with predictable procedures.
The model Hassabis invoked is the Financial Industry Regulatory Authority, the self-regulatory organization that oversees broker-dealers in U.S. securities markets under federal supervision. The analogy is revealing. It is not a call for a traditional agency built entirely inside Washington. It is a call for a body funded by industry, watched by government and staffed with enough technical expertise to evaluate fast-moving systems before they are used at scale.
That structure is meant to solve a real weakness in AI governance. General-purpose agencies often lack the compute access, model-security expertise and evaluation staff needed to test frontier systems. But a purely industry-run board would face an obvious trust problem. The largest companies most likely to be regulated would also have the strongest influence over funding, standards and timing. Hassabis' version tries to balance those forces by including independent experts and open-source representatives, while leaving the United States government in a supervisory role.
Business Insider and The Verge both reported that Hassabis framed the moment around artificial general intelligence arriving within a few years, not decades. That timeline is contested, but the policy implication is clear: if labs believe models are approaching economically and strategically disruptive capabilities, voluntary safety cards and private red-team reports will not satisfy governments forever. A standards body would be a way to create a repeatable process before a crisis forces a more blunt response.
The immediate challenge is defining the threshold. A frontier regulator has to know which systems count as frontier systems. Model size is no longer enough. Capability can come from training scale, test-time compute, tool use, agentic scaffolding, model combinations, access to sensitive data or downstream fine-tuning. A body that keys only on benchmark scores could miss dangerous systems and over-regulate harmless ones. A body that claims too much authority could become a bottleneck for research and open development.
The politics are also awkward. The Trump administration has supported AI infrastructure expansion and has been skeptical of creating an FDA-style AI regulator. At the same time, national-security officials have shown they are willing to pressure labs over model access when they believe a system creates strategic risk. A FINRA-like approach offers a middle path: not a conventional new agency, but not laissez-faire release governance either.
Open-source advocates will watch the details closely. If the standards body mainly reflects the preferences of large closed labs, it could become a compliance moat that entrenches incumbents. If it gives open-model builders meaningful representation, it could help build evaluation science that applies beyond a small circle of private API providers. Hassabis has gestured toward that broader participation, but the charter, funding rules and board design would decide whether the promise survives contact with institutional politics.
For now, the proposal is not law. It is a sign that frontier AI governance is leaving the statement-of-principles phase. The next fight is over who gets to test the most powerful models, who pays for the testing, who sees the evidence and who has authority to say no. Hassabis has offered one answer. The harder question is whether governments, labs and civil society can build an institution trusted enough to use it before the next release fight arrives.
Topics: DeepMind, AI regulation, frontier models, Demis Hassabis