Hassabis pitches an IAEA for AI as he steps back from DeepMind's top job

answer: Hassabis pitched Trump officials on an IAEA-style global AI safety oversight body.

takeaways:

The stakes here are regulatory, not just reputational: whoever shapes the rulebook for frontier AI gets to set the cost of compliance for every lab that follows. Demis Hassabis is trying to write that rulebook before governments do it for the industry.

The Wall Street Journal reported on 13 August that Hassabis pitched senior Trump administration officials, including Treasury Secretary Scott Bessent and science adviser Michael Kratsios, on a new independent AI safety entity modelled on the International Atomic Energy Agency, the UN's nuclear watchdog. He also raised the idea with other frontier AI companies. The talks took place in the weeks before he gave up the DeepMind CEO role on 5 August; he now serves as DeepMind chair and Alphabet chief scientist, according to OODA Loop's summary of the Journal's reporting.

The proposed body would set safety guardrails and best practices for developing artificial general intelligence. Hassabis has been devoting more of his time to the idea, the notes indicate. It builds on a position he laid out publicly on 14 July, when he called for a US-led, government-overseen public-private standards body to test frontier models for national-security risks, according to CNBC. In June, he and Anthropic's Dario Amodei had already urged a US-led coalition on AI rules at a G7 meeting.

His longer-standing architecture: national AI safety institutes, such as the UK's, feeding into a global body that runs safety benchmarks, certification schemes and independent audits of frontier systems. The "IAEA for AI" framing is not new to policy circles — UN Secretary-General António Guterres endorsed the concept in 2023 — but Hassabis is now the one carrying it into rooms with US cabinet-level officials.

The timing tracks a rougher summer for the industry. Agent incidents at both OpenAI and Anthropic pushed labs toward common safety standards, and a group of labs has since begun discussing an industry-led standards body. Hassabis has also been explicit about where he draws the line technically: "Nobody should be building systems that are capable of deception because then they could be getting around other safeguards," he said on the 20VC podcast earlier this year.

A body with IAEA-style audit powers would reshape competitive dynamics in frontier AI: certification and independent audits raise fixed costs, which favours well-capitalised incumbents like Google DeepMind, OpenAI and Anthropic over smaller entrants who cannot absorb compliance overhead. That is worth watching alongside the politics — Hassabis is courting Trump officials rather than European regulators, suggesting the industry would rather set terms with a US-led framework than wait for a Brussels-style mandate.

The move away from DeepMind's CEO chair also reads as freeing Hassabis to run policy diplomacy full-time. If labs that are still discussing their own standards body converge with his IAEA model, the practical rulebook for AGI development could be written by the companies it governs, not by legislators.