A Proposal From Inside the Industry
Demis Hassabis, CEO of Google DeepMind, is calling for the creation of an independent standards body dedicated to testing frontier AI models and establishing best practices for how those models get released to the public. The proposal draws directly on the structure of FINRA – the Financial Industry Regulatory Authority – as a working blueprint for what AI governance could look like at the frontier level.
The suggestion comes from one of the most prominent figures in AI development, and that positioning matters. Hassabis is not a regulator, a politician, or a policy researcher. He runs one of the most advanced AI labs on the planet. When he argues that external oversight is needed, it shifts the conversation about who is actually asking for that oversight – and why.
Why FINRA and Not Something Else
FINRA is a self-regulatory organization that oversees broker-dealers in the United States. It operates independently of the government, sets conduct standards, runs examinations, and enforces rules across the financial industry – all without being a federal agency. Hassabis’s decision to reference it specifically, rather than the FDA, the FAA, or any other regulatory analogy, suggests a preference for industry-adjacent oversight rather than direct government control.
That distinction carries real weight. A body modeled on FINRA would likely draw expertise from the AI industry itself, maintain some degree of independence from political cycles, and focus on technical standards rather than broad legislative mandates. The downside of that model, as critics of FINRA have long noted in financial circles, is that self-regulatory organizations can become captured by the very industries they oversee – setting standards lenient enough to satisfy members rather than protect the public.
What “Frontier Models” Actually Means Here
The term “frontier models” refers to the most capable AI systems being developed – the large-scale models at the cutting edge of benchmark performance, capability, and, increasingly, autonomous behavior. These are the systems that labs like DeepMind, OpenAI, Anthropic, and Meta are racing to build and deploy. They are also the systems where capability gains are moving faster than any existing regulatory framework can track.
Testing frontier models before release is not a straightforward task. Unlike a pharmaceutical drug with measurable biological effects, an AI model’s risks can be emergent, contextual, and difficult to reproduce in controlled settings. A model might behave safely in a sandbox evaluation but surface dangerous capabilities when deployed at scale across millions of users in unpredictable real-world conditions. That gap between pre-deployment testing and post-deployment behavior is exactly where a standards body would need to operate – and where the technical challenges are most acute.
Best practices for model release are also still largely undefined at the industry level. Some labs publish model cards. Some run red-teaming exercises internally. Some share safety evaluations voluntarily with governments or third parties. None of this is standardized, none of it is verified externally, and none of it is mandatory. What Hassabis appears to be proposing is a mechanism that could change at least one of those conditions.
There is a version of this proposal that has real teeth: mandatory pre-release evaluations conducted by independent technical experts, with results that are binding rather than advisory. There is also a version that ends up functioning as a public relations exercise – voluntary participation, self-reported results, and standards written by committees with strong industry representation. The distance between those two outcomes depends almost entirely on how such a body would be structured, funded, and empowered to act.
The Timing of the Proposal
Hassabis is making this call at a moment when AI regulation globally is fragmented and contested. The European Union’s AI Act is the most comprehensive framework in effect, but its provisions for frontier models remain a point of ongoing debate. In the United States, federal AI legislation has stalled repeatedly, and executive action has moved in different directions under different administrations. The UK has positioned itself as a governance hub but lacks binding authority. No international consensus mechanism for AI safety standards currently exists.
Into that vacuum, the proposal for a FINRA-style body is both timely and convenient for AI labs. Timely because the absence of standards is becoming increasingly difficult to defend publicly as AI systems grow more capable. Convenient because a self-regulatory model – if designed by industry insiders – gives labs significant influence over what those standards actually require. Whether Hassabis’s proposal is primarily a good-faith safety initiative or a strategic effort to define the terms of oversight before governments do is a question the details of any actual implementation would have to answer.
What Comes Next
A proposal is not a policy. For a standards body of this kind to take shape, it would need buy-in from multiple frontier labs, agreement on governance structure, a funding mechanism, and some form of legal authority – either granted by governments or enforced through industry norms and reputational pressure. None of that is in place. What exists right now is a high-profile CEO making a public argument, which is nonetheless how many policy processes actually begin.
DeepMind occupies a specific position in this conversation. It is a subsidiary of Alphabet, which gives it both the resources to push for institutional change and the corporate pressures that make any regulatory outcome consequential to its bottom line. Hassabis has been one of the more vocal figures in AI on the subject of long-term safety – he co-founded the field of AI safety research in many respects – but his institutional role means his proposals will always be read partly through that lens.
The FINRA analogy will likely face scrutiny from multiple directions: from those who think AI risks are too serious for a self-regulatory approach, from those who think any new oversight body will stifle innovation, and from regulators in different jurisdictions who may see it as an attempt to preempt their authority. Hassabis has put a specific model on the table. Whether anyone else is willing to build it on his terms – or whether governments move first and build something he did not design – is the actual open question.
