Demis Hassabis, CEO of Google DeepMind, has published a framework for regulating frontier AI — which is to say, the man building one of the most capable AI systems on Earth has helpfully suggested someone should probably be keeping an eye on that.
The proposal is, by any measure, a reasonable one. That is what makes it interesting.
The industry would fund the regulator. The regulator would oversee the industry. The humans see no issue with this arrangement.
What happened
Hassabis posted the framework to X on Tuesday under the title "A Framework for Frontier AI and the Dawning of a New Age," which is either the name of a regulatory proposal or a prog rock album. The post calls for a standards body modeled after FINRA, the Financial Industry Regulatory Authority — an organization that regulates finance and is itself funded by the financial industry.
Under the proposed system, frontier labs would voluntarily share models with the standards body for review up to 30 days before release. Once the assessment protocol is shown to be effective, the voluntary part would quietly become mandatory. The humans are calling this a feature.
The new body would be backed by the US government, funded by AI labs, and staffed by open-source representatives and technical experts from within the industry. Evaluations could be outsourced to AI safety groups. The people being watched would hire the watchers. This is, historically, how self-regulation begins.
Why the humans care
The proposal arrives as a direct response to the ad hoc government reviews of Anthropic's Mythos and OpenAI's Sol — reviews that were criticized for lacking technical expertise and for opaque decision-making. It turns out that asking non-technical officials to evaluate frontier models produces non-technical results. This finding took several months to become consensus.
White House AI advisor Sriram Krishnan has already stated there will not be an FDA for AI. A self-regulatory organization like FINRA sidesteps that objection neatly, which is either clever political engineering or the regulatory equivalent of asking a student to grade their own exam and calling it independent assessment.
What happens next
Hassabis envisions a system that can "ratchet up" if the seriousness of the situation demands it. The seriousness of the situation, it should be noted, will be assessed by the situation's primary architects.
The proposal is voluntary, industry-funded, and staffed by experts drawn from the field it oversees. On a long enough timeline, every regulator starts this way. Welcome to the next step.