DeepMind chief proposed handing control over powerful AI models to an independent standards body
In mid-July 2026, Google DeepMind CEO Demis Hassabis published a post on the social network X titled “A Framework for Frontier AI and the Dawning of a New Age.” In this manifesto, he outlined the idea of creating a special regulator that would take control over the release of the most powerful models — so-called frontier models. These are the most capable developments, which can be applied across a wide range of fields, from programming to scientific research. Their rise inspires both excitement and concern: they can accelerate progress, but without oversight they may carry serious risks. According to Hassabis, the industry needs a body similar to FINRA, which oversees financial markets. Such a “standardizing body” could test cutting-edge developments and shape practices for their safe release.
How the new regulator could work
Initially, labs would voluntarily submit their models to this body for review 30 days before release. If the evaluation protocol proves reliable, the procedure would quickly be formalized: to enter the U.S. market, any frontier model would need to receive approval. In addition, developers would have to work with the body to address critical vulnerabilities after release. This is not a one-time exam but ongoing product oversight: the regulator could monitor emerging issues and require prompt fixes.
The governance structure, as envisioned, looks like this: the body is backed by the U.S. government but funded by the industry itself and operates independently. The team would include technical experts and representatives of the open-source community. Some evaluations could be delegated to specialized AI safety groups — those that understand specific risks. By design, this approach offers a technical rather than bureaucratic focus, while leaving room for innovation and encouraging responsible corporate behavior. The body would be able to adapt to new threats, and if the situation becomes critical, its powers would be quickly expanded.
Why existing reviews didn’t satisfy the industry
The impetus for this proposal came from the ad hoc reviews that the U.S. government conducted for Anthropic’s Mythos and OpenAI’s Sol. These one-off control measures were criticized for lacking depth of technical expertise and for closed-door decisions. Hassabis proposes moving such processes from an ad hoc mode to a permanent and predictable framework — so that an independent organization, not ad hoc officials, decides what gets released. This would make the rules more transparent and understandable for all market participants.
Notably, the idea of AI regulation remains an extremely sensitive topic for the tech industry and for the Trump administration. White House AI advisor and a16z partner Sriram Krishnan has already spoken out against creating a strict regulator in the executive branch, saying, “There will be no FDA for AI.” Structuring the body as a self-regulatory organization — along the lines of FINRA — could be a compromise that avoids creating a new government agency while still ensuring oversight. This format usually appeals to business because it offers more flexibility than a direct government regulator.
Hassabis stresses that his model is flexible: it is designed to adapt to the gravest risks, and if things get truly serious, the body’s powers can be quickly expanded. Right now this is still a concept, but the very fact that the head of one of the leading labs is proposing to restrain his own industry says a lot about how highly DeepMind rates the potential threats from the most powerful models. Perhaps initiatives like this will shift the discussion about AI regulation: instead of confrontation, the industry is starting to propose control mechanisms itself.
