Why DeepMind's Hassabis wants an "IAEA for AI"

You've probably heard AI companies promise to keep their most powerful systems safe. Now one of the industry's most senior figures is pushing for an outside body to check that they actually do.

The Wall Street Journal reported on 13 August that Demis Hassabis, the head of Google DeepMind, pitched senior US officials on creating an independent AI safety organisation. The idea is modelled on the International Atomic Energy Agency (IAEA), the United Nations body that monitors nuclear technology worldwide, according to a summary of the report.

Hassabis raised the proposal with Treasury Secretary Scott Bessent and White House science adviser Michael Kratsios, as well as with other leading AI companies. The talks took place in the weeks before 5 August, when Hassabis stepped down as DeepMind's chief executive to become the company's chair and Alphabet's chief scientist. He has reportedly been spending increasing amounts of time on the safety-body idea.

This builds on public comments he made on 14 July, when he called for a US-led, public-private body to test powerful new AI models for national-security risks, as reported by CNBC. In June, he and Anthropic chief executive Dario Amodei urged a US-led coalition on AI rules at a G7 meeting.

His broader vision involves national AI safety institutes, similar to one already running in the UK, feeding into a single global body. That body would set safety benchmarks, certify AI systems and carry out independent audits of the most powerful models. The general concept of an "IAEA for AI" was first endorsed by UN Secretary-General António Guterres in 2023.

Hassabis has previously warned about one specific risk such a body might guard against. On the 20VC podcast earlier this year he said: "Nobody should be building systems that are capable of deception because then they could be getting around other safeguards."

The push comes after agent-related incidents at OpenAI and Anthropic this summer prompted AI labs to seek common safety standards. A group of labs has since begun discussing forming an industry-led standards body of their own.

No such oversight body exists yet, so nothing changes for you today. But the push signals that the companies building the AI tools you use, including Google's, want outside checks on how those systems are tested and certified, rather than relying purely on self-policing. If a body like this is eventually created, it could set rules for how future AI models are vetted for safety before they reach the public.

The notes don't say whether the Trump administration has responded to Hassabis's pitch or whether other AI companies have formally backed it. The parallel industry discussion about a labs-led standards body is still at an early, informal stage.