Exclusive: Google's Hassabis calls for U.S.-led global AI watchdog

Demis Hassabis, the co-founder and chief executive of Google DeepMind, has called for the United States to lead the creation of a global artificial intelligence watchdog. In a personal manifesto titled “A Framework for Frontier AI and the Dawning of a New Age,” published Tuesday, Hassabis outlines a proposal for a regulatory body that would oversee the development and deployment of the most advanced AI systems worldwide. According to the manifesto, Hassabis argues that current approaches to AI governance are inadequate. He warns that the risks posed by AI-driven cyber threats are already serving as “warning shots.” If left unchecked, these risks, alongside more severe biological and nuclear threats, could soon reside within open-source models beyond the reach of any nation’s regulatory framework. Hassabis emphasizes that the greatest danger comes not from open-source models but from the more powerful proprietary models developed by leading laboratories. Hassabis has spent several months working behind the scenes to build support for this initiative. He has engaged with members of the Trump administration, senior figures at other AI research organizations, and European officials. He reports that the feedback he has received has been largely favorable. The Trump administration, which previously favored a hands-off approach to AI regulation, appears to have shown openness to the idea, particularly following recent developments involving Anthropic’s models. The proposed regulatory body would function similarly to the Financial Industry Regulatory Authority (FINRA), a self-regulatory organization in the financial sector. Under Hassabis’s plan, leading AI research labs would be required to submit their most advanced models for safety evaluation up to 30 days before public release. This process would involve rigorous testing for potential dangers related to cybersecurity, biological threats, and deceptive capabilities. Once the testing procedures are deemed effective and reliable, the body could transition into a formalized regulatory entity. The structure of the proposed watchdog would include a majority-independent board composed of top-tier technical experts, including recipients of the Turing Award. The board would also incorporate representatives from both the industry and government sectors, as well as those from the open-source community. Hassabis envisions that the “frontier” designation, applied to the most advanced AI models, would become a mark of distinction, indicating that the model has undergone thorough scrutiny. The proposal applies to all AI models classified as “frontier class,” regardless of their country of origin or whether they are open-source or proprietary. Hassabis suggests that the criteria for classification would be periodically reviewed and adjusted as technological capabilities continue to evolve. Recent actions by the U.S. government have underscored the urgency of such measures. The Trump administration’s abrupt halt of Anthropic’s Mythos and Fable models last month served as a wake-up call for many in the AI community. Anthropic faced an unexpected export-control order that effectively froze its most powerful models. After nearly three weeks of negotiation, the models were eventually released, though without clear guidelines or protocols. OpenAI, seeking to avoid a similar situation, restricted access to its GPT-5.6 model to vetted government partners during its initial launch. The model was later made available to the public after extensive discussions with the Commerce Department. Other prominent AI researchers, including Anthropic’s CEO Dario Amodei, have also advocated for stricter regulations. Amodei has proposed the establishment of an agency akin to the Federal Aviation Administration (FAA), capable of blocking the deployment of unsafe AI models. While there is broad agreement among leading AI researchers that stronger regulation is necessary, there remain differences over who should hold the ultimate authority in shaping these policies. Hassabis remains optimistic that the proposed regulatory framework can be implemented within months, with the new body potentially becoming operational before the end of the year. As the race to develop increasingly powerful AI continues, the need for coordinated, globally recognized oversight mechanisms grows ever more pressing.