Google DeepMind Chief Executive Demis Hassabis unveiled a proposal for a new international watchdog that would do “rigorous” tests and reviews of cutting-edge AI models before release.The plan, unveiled in a blog post last week, was quickly praised by rivals such as OpenAI’s Sam Altman, who called it “thoughtful,” and Elon Musk, who described it as “a good starting point for discussions.” Next week, Hassabis intends to hold meetings about his proposal in Washington with U.S. policymakers, according to a person familiar with the matter.Hassabis, who had been preparing his framework for months, said he decided to publish it following concerns around the advanced cybersecurity capabilities of Anthropic’s Mythos model, which he called a “warning shot” for society.“I’m worried about different types of risks also coming down the line,” Hassabis said in an interview, focusing on the potential for new biosecurity threats. “The models aren’t good enough yet, but if you extrapolate out, maybe they will be in a couple of years.”His proposal coincides with heightened uncertainty in the U.S. around government scrutiny of new AI models. Anthropic and OpenAI delayed the wide release of their latest offerings under pressure from the Trump administration. Those moves appeared to come as a surprise to other countries — which rely on U.S. technology — and prompted discussions between the European Union and the U.S. government over model access, along with meetings between tech CEOs and world leaders at the recent G-7 conference in France.Hassabis said that his company has also been in talks with the government over the release of its newest offerings. “As we train our latest models and as they get certain scores and benchmarks, we’re talking to government and the AI security institutes about those things and testing them internally. That all has to continue,” Hassabis said. “But I think doing that in an ad hoc fashion isn’t very sustainable in the longer term.Altman also previously called for a new U.S.-led international group to set safety standards around AI. The suggestion from Hassabis is similar, and modeled after oversight bodies like the Financial Industry Regulatory Authority.His plan specifies who should be appointed to the new group (independent technical experts); where the funding should come from (leading AI firms); and how long the companies should submit their models for review (up to 30 days). Its approach “could be ratcheted up if the seriousness of the situation demands,” he said, including coordinating an industry slowdown in development among leading labs.There’s no guarantee such a plan will come to fruition. The U.S., EU and China have diverged at times in their playbooks for AI regulation. The Trump administration has also waffled between a hands-off and hands-on stance toward AI. And Congress has yet to pass meaningful federal legislation for regulating AI.Hassabis, for his part, sounded optimistic that there might be “enough momentum” at this moment to “actually get something over the line.”Prior to publishing the blog post, Hassabis said he also socialized his plan with virtually every major AI lab and a number of CEOs — including Anthropic’s Dario Amodei — to seek feedback and support. (Amodei recently suggested the creation of a new agency similar to the Federal Aviation Administration to oversee AI in the U.S.)“We’ve talked a lot about this,” Hassabis said of his chats with Amodei, who he considers a friend. “I think it’s very compatible with the ideas he’s been putting out, and also Sam [Altman] too, with some of his ideas. But we needed something concrete rather than abstract.”Hassabis, a Nobel laureate who has worked on AI for longer than most, may be uniquely positioned to rally support for the idea. He’s widely respected within the industry and has had fewer feuds with rival executives than, say, Musk or Altman. Still, one AI leader, who also asked not to be named, noted that industry consensus isn’t enough, referring to the political challenges of getting the U.S. government on board.“Obviously the next step is to see if we can actually make it happen,” Hassabis said, “not just publish a blog post.”Ghaffary writes for Bloomberg.
