Demis Hassabis, CEO of Google DeepMind, is pushing for the creation of an independent standards body specifically designed to govern frontier AI models — and he has a concrete analogy in mind: FINRA, the Financial Industry Regulatory Authority that oversees broker-dealers in the United States.

The Proposal

Hassabis's pitch centers on establishing an organization that would:

  • Test frontier AI models before and after deployment
  • Develop best practices for how and when those models are released to the public
  • Operate independently from both governments and the AI companies themselves
  • Serve as a credible, technical authority — not a political one

The FINRA comparison is deliberate. FINRA is a self-regulatory organization (SRO) — it's not a government agency, but it operates under SEC oversight and has real enforcement authority. Hassabis appears to be advocating for a similar structure: industry-adjacent, technically rigorous, and insulated from the slower cycles of legislative action.

Why This Matters Now

Frontier AI governance is currently a patchwork. The EU AI Act is in phased rollout. The UK's AI Safety Institute has been rebranded and restructured. The US under the current administration has pulled back from Biden-era executive orders on AI safety. Meanwhile, labs like DeepMind, OpenAI, Anthropic, and Meta are releasing increasingly capable models with no unified framework for pre-deployment evaluation.

The absence of a shared technical standard creates real risks — not just safety risks, but reputational and liability risks for companies building on top of these models. If a frontier model causes harm, the question of who tested it, how, and against what criteria becomes immediately consequential.

The FINRA Model — and Its Limits

FINRA works because financial products are, at some level, legible. Risks can be quantified. Disclosures can be standardized. AI models — especially large multimodal systems — are substantially harder to audit in any consistent way.

Critics of the SRO model for AI will likely point out that self-regulatory bodies in finance have a mixed track record, and that an industry-funded standards organization could be captured by the largest labs. The companies with the most resources to participate in standard-setting are the same ones with the most to gain from standards that favor their existing approaches.

That said, the alternative — waiting for national legislatures to develop technically coherent AI law — has its own obvious failure modes.

What It Means for Founders and Builders

For startups and developers building on frontier models, a credible standards body would have significant practical implications:

  • Compliance costs would become more predictable — you'd know what a "safe" model looks like according to a defined rubric
  • Liability exposure could shift — if a model passed independent evaluation, that changes the legal calculus for downstream builders
  • Procurement decisions for enterprise customers would likely begin referencing standards-body certification, similar to how SOC 2 or ISO 27001 function in security
  • Fundraising narratives may increasingly need to address regulatory readiness, especially for companies in sensitive verticals

Hassabis hasn't outlined a specific governance structure, funding model, or timeline — this is a call to action, not a fully formed proposal. But coming from the CEO of one of the world's most influential AI labs, it carries weight and is likely to accelerate ongoing conversations at bodies like the OECD, the UN's AI advisory body, and national AI safety institutes.

The question is whether the industry's biggest players will back an organization with genuine independence and teeth — or prefer one they can more easily influence.