Google DeepMind head calls for US-led testing body for powerful AI models

Demis Hassabis, the head of Google's artificial intelligence lab, called Tuesday for the United States to create an organization to test the most powerful AI systems before they are released to the public.

Hassabis, who runs Google DeepMind and won a Nobel Prize for work on protein science, wrote in an essay that artificial general intelligence, meaning AI able to match the full range of human thinking, is likely only a few years away.

Google DeepMind is the division behind Google's Gemini AI models, which compete with systems from OpenAI and Anthropic.

Hassabis proposed a new regulatory body modeled on the Financial Industry Regulatory Authority, the private watchdog that oversees Wall Street brokerage houses under federal supervision. He said the AI body should be staffed by leading technical experts, funded mainly by the AI industry and responsible for developing safety benchmarks that frontier models would have to pass before entering the market.

Under his proposal, AI developers would at first submit their most powerful models for review voluntarily up to 30 days before release. Once the testing system was shown to be reliable, compliance would become mandatory.

Hassabis wrote that no one knows with certainty what will happen next in AI and that even experts disagree. He said that, with high uncertainty and high stakes, a strategy of cautious optimism was sensible.

The proposed rules would apply to any model considered powerful enough to be 'frontier-class', whether built in the United States or abroad. Hassabis said he hoped international players would join the system after it was established.

The proposal comes amid calls from rival AI executives for binding government oversight. It also follows a June dispute in which the Commerce Department ordered Anthropic to suspend all foreign access to its Claude Fable 5 and Mythos 5 models, citing national security concerns linked to a reported jailbreak in which users bypassed safety guardrails. Anthropic took both models offline entirely for more than two weeks before access was gradually restored.

In an interview with Axios published alongside his manifesto, Hassabis described the Mythos episode as a wake-up call showing that Washington needs a more structured approach.

Anthropic chief executive Dario Amodei had also called for regulation shortly before the government action, publishing an essay in early June that advocated for a federal agency with authority to vet powerful AI models before deployment.