Google DeepMind CEO Demis Hassabis posted on X Tuesday morning calling for a new regulatory body to oversee frontier model releases, according to TechCrunch AI. The post, titled “A Framework for Frontier AI and the Dawning of a New Age,” proposes a standards body modeled on FINRA, the self-regulatory organization that polices US broker-dealers. It would test frontier models before launch and set best practices for how they ship.
One of the most powerful people in AI just asked to be regulated. That alone makes this worth your attention.
The Proposal, in Five Points
- Voluntary first. Frontier labs would share models with the standards body for review up to 30 days before release.
- Mandatory later. “Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market,” the post reads.
- Post-release duty. Labs would keep working with the body to patch critical vulnerabilities found after launch.
- Industry-funded, government-backed, independently run. The AI labs pay for it. Washington stands behind it. Neither one operates it.
- Staffed by people who actually understand the tech. Open source representatives and technical experts from inside the industry, with the option to outsource specific evaluations to specialized AI safety groups.
Why FINRA and Not the FDA
The model choice is the whole argument. FINRA is a self-regulatory organization: the industry funds it, the government authorizes it, and it operates at arm’s length from both. That structure sidesteps the fight nobody in this administration wants to have.
And that fight is live. TechCrunch AI notes that White House AI advisor and a16z general partner Sriram Krishnan recently shut the door on an executive-branch regulator, saying “there will not be an FDA for AI.” Hassabis isn’t asking for one. He’s routing around it. That’s not an accident, it’s the design.
The Status Quo He’s Trying to Replace
Right now, frontier oversight in the US is improvised. The US government ran ad hoc reviews on Anthropic’s Mythos and OpenAI’s Sol, and those reviews got hammered on two fronts: the reviewers lacked the technical depth to evaluate what they were looking at, and the decisions on when a model could ship were opaque.
That’s the gap. A lab spends billions training a model, then waits on a process with no published criteria, no clear timeline, and no reviewers who can read the eval results. Neither safety advocates nor the labs are happy with that arrangement. Hassabis’s pitch is that a technically staffed, industry-funded body fixes both complaints at once.
What Stands Out
The 30-day pre-release window is the sharpest detail here. It’s short enough that labs can live with it and long enough to run real evaluations. Compare that to drug approval timelines measured in years. Hassabis is explicit about the tradeoff: “The strength of this approach is it would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour.”
He also builds in a throttle. The framework “is designed to keep up with the field’s acceleration and adapt to the biggest risks as they are identified, and could be ratcheted up if the seriousness of the situation demands.” Start light. Tighten if the risk picture changes.
The obvious critique writes itself: an industry-funded regulator staffed by industry experts is a regulator with a conflict of interest. FINRA has caught that same criticism for decades. Whether “operated independently” survives contact with the labs writing the checks is the question that decides if this works.
What to Watch
- Whether OpenAI, Anthropic, and Meta publicly back the framework or stay quiet. Voluntary participation means nothing without them.
- Whether the White House engages at all, given Krishnan’s position.
- Whether the 30-day window holds once labs run the math on competitive delay.
If you’re building on frontier models, the practical takeaway is simple: pre-release review is coming in some form. The fight is over who runs it, not whether it happens. Plan your release timelines with a review window in mind, even before one is required.
Full details are at the original source.