Google DeepMind's CEO Wants a FINRA-Style Watchdog for the Most Powerful AI Models
Demis Hassabis has proposed an independent standards body that would review frontier AI models before they reach the public. Here is what he is calling for, and why it matters.
Key points
- Google DeepMind CEO Demis Hassabis published a proposal on 14 July 2026 calling for an independent body to review the most powerful AI models before release.
- The proposed body is modelled on FINRA, the Financial Industry Regulatory Authority, a self-funded, independent watchdog that oversees US financial brokers without sitting inside government.
- Under the plan, AI labs would voluntarily share new models with the body up to 30 days before launch, with mandatory compliance potentially following once the process is proven.
- Earlier US government reviews of Anthropic's Mythos and OpenAI's Sol models drew criticism for lacking technical expertise and clear decision-making rules.
- White House AI advisor Sriram Krishnan said in 2026 that the administration would not create "an FDA for AI," making a self-regulatory structure the more politically viable path.
Demis Hassabis, the CEO of Google DeepMind, posted a detailed proposal on Tuesday calling for a new organisation to act as a gatekeeper for the world's most powerful AI systems. He calls them "frontier models," meaning the cutting-edge AI systems that push the boundaries of what the technology can currently do.
The idea is simple enough. Before a frontier model goes public, its developers would hand it to an independent standards body for up to 30 days of safety testing. Pass the checks, ship the product. Fail, fix the problems first.
Hassabis wants the body structured like FINRA, the Financial Industry Regulatory Authority. FINRA is not a government agency. It is funded by the financial industry itself and operates independently, setting and enforcing conduct rules for stockbrokers. The AI version would work the same way: backed by cash from the big AI labs, staffed by technical experts and open-source researchers, and free to contract out specialist evaluations to independent AI safety groups.
The proposal fills a gap that has become embarrassingly obvious. The US government already ran ad-hoc safety reviews of two high-profile models, Anthropic's Mythos and OpenAI's Sol, a large language model and a multimodal AI system respectively. Both are among the most capable AI tools built to date. Those reviews were widely criticised, first reported by TechCrunch AI, for thin technical expertise and murky rules about what would actually trigger a block on release.
A standing body with full-time specialists would fix both problems.
The political angle is tricky. The Trump administration has shown little appetite for new AI regulators. Sriram Krishnan, the White House AI advisor and a partner at venture capital firm Andreessen Horowitz, flatly said there would be no "FDA for AI." Hassabis's self-regulatory model is a direct response to that objection: no new government agency, no new bureaucracy, just the industry policing itself with government backing.
Should ordinary users care about this?
Yes, in a practical sense. Right now, the most powerful AI tools arrive with no consistent safety check. A standards body would mean that the next wave of AI products you use, whether for writing, healthcare, or financial advice, has cleared at least one independent technical review before reaching your phone or browser.
Hassabis frames the system as designed to grow with the technology, tightening its requirements if the risks grow more serious. That is a reasonable structure. The detail that will matter most is who actually sits on the body and whether the labs accept its findings when those findings are inconvenient.
