Hassabis wants a FINRA for AI — an independent body to test frontier models

Demis Hassabis, CEO of Google DeepMind, posted on X on July 14, 2026, a lengthy essay titled “A Framework for Frontier AI and the Dawning of a New Age” — the most detailed public proposal for frontier AI governance from a major lab chief in at least a year. The centerpiece is an independent standards body modeled on FINRA — the Financial Industry Regulatory Authority, the self-regulatory organization overseeing US securities brokers and dealers — that would test the most powerful AI models before public deployment.

Key takeaways

Hassabis proposes a self-regulatory body modeled on FINRA, funded by the AI industry
Labs would voluntarily submit models for review up to 30 days before release
Once proven effective, review would become a legal requirement for US market deployment
The body would replace ad-hoc US government reviews of Anthropic's Mythos and OpenAI's Sol, criticized for lacking transparency
White House AI advisor Sriram Krishnan has stated “there will not be an FDA for AI”

How the mechanism would work

According to Hassabis, the standards body would employ technical experts and open-source representatives, funded by AI labs but operating independently. In the initial phase, frontier labs would voluntarily submit models for review up to 30 days before release. The body would assess risks, test safeguards, and develop best practices for model release. Once the protocol demonstrated effectiveness, compliance would become mandatory — models that fail the review could not be deployed to the US market.

The institution could outsource risk assessments to specialized AI safety organizations that would focus on specific risk categories. Hassabis argues the system must dynamically adapt to new risks as the field accelerates — unlike static legislation, which typically takes years to catch up with the actual state of technology.

Context: the ad-hoc reviews were opaque

Hassabis's proposal is a direct response to a recent precedent. The Trump administration conducted informal reviews of two of the most advanced models on the market — Anthropic's Mythos and OpenAI's Sol — before allowing their broader release. Those reviews drew widespread criticism: they lacked clear evaluation criteria, technical depth, and transparency about who decided and on what basis. Hassabis argues this is precisely why a technically competent institution is needed — rather than ad-hoc reviews conducted by generalists from the administration.

The FINRA analogy is intentional. FINRA is an industry organization — not a government agency — that regulates the activities of securities brokers and dealers in the US. It operates under authority delegated by the SEC but is funded by the industry itself. This model has built deeper expertise than typical government regulators while maintaining some independence from political pressures.

Regulators say: not this time

The proposal already faces political resistance. Sriram Krishnan, White House AI advisor and a16z general partner, explicitly rejected the concept of a federal AI regulator, stating “there will not be an FDA for AI.” Krishnan represents the faction that believes overregulation stifles innovation and could cede technological leadership to Chinese labs. Hassabis is trying to address that critique — a self-regulatory model like FINRA does not require new federal legislation or a new government agency, making it politically more achievable.

At the state level, however, signals of growing pressure are emerging. Illinois signed SB 315 in July — the first state law mandating independent safety audits of AI models. New York imposed a moratorium on new data centers above 50 MW. The absence of a federal regulator does not mean an absence of regulation — it risks a fragmented state-level patchwork that would complicate nationwide deployments.

Industry resistance

Hassabis's proposal carries unusual weight because it comes from the head of one of the strongest frontier labs — Google DeepMind is part of Google, which builds its own Gemini models and competes directly with Anthropic and OpenAI. The fact that it is Hassabis, rather than a smaller player, putting forward this initiative gives it political credibility. At the same time, it remains an open question whether OpenAI and Anthropic want an external body approving their releases — even voluntarily. Both companies have invested substantial resources in their own safety protocols and may see the proposed body as a potential regulatory trap or a tool that someone else could control.

Why this matters

Hassabis's proposal emerges at a moment when the AI regulatory landscape is exceptionally unstable. The US lacks coherent federal policy, and the White House is actively dampening initiatives that could slow technological deployments. In Europe, the AI Act is beginning to apply, but it does not cover models produced exclusively for the US market. In this vacuum, an industry proposal for a technical body funded by the labs themselves may be the only realistically achievable compromise.

More importantly, the problem Hassabis is trying to solve is real: the public, regulators, and even engineers themselves have decreasing visibility into what the most powerful AI models are actually doing. As models become capable of planning multi-step autonomous actions and generating production code with minimal oversight, the absence of a standardized risk evaluation process is a serious gap — not only politically, but from an engineering perspective.

What's next

Hassabis's proposal is currently a statement of intent — no concrete timelines or formalization mechanisms accompany it.
The existing government ad-hoc reviews show the Trump administration does not exclude engaging in frontier model evaluation — the question is whether it will agree to delegate that role to an industry body.
Illinois SB 315 takes effect in 2028 — if Congress does not preempt states with federal legislation, pressure to create some form of evaluation body will grow regardless of White House preferences.