Demis Hassabis Proposes New AI Oversight in the U.S.

Demis Hassabis, the co-founder and CEO of Google DeepMind, is pushing for the creation of a new regulatory body in the U.S. This watchdog would have the authority to evaluate advanced AI models and potentially slow down development if risks escalate.

In a personal manifesto titled “Frontier AI Frameworks and the Dawn of a New Era,” Hassabis outlines his ideas. He shared these thoughts in an interview, underscoring the pressing need for a comprehensive regulatory strategy for artificial intelligence. He envisions this approach being funded by the industry and staffed with top technology experts and held accountable to the U.S. government.

At DeepMind’s London office, Hassabis pointed to current AI-driven cybersecurity threats as early warnings of more significant risks to come. He predicts that within 18 months, advanced capabilities, alongside serious biological and nuclear dangers, may emerge from open-source AI systems that no government will be able to control. He noted that risks will not only come from open-source models but also from powerful proprietary systems developed by leading AI labs.

“What we do collectively now will determine how the next stage of civilization unfolds,” he stated in his manifesto.

In recent months, Hassabis has held private discussions to gather support for his proposed framework, engaging with officials from the Trump administration, other AI institute leaders, and European government representatives before going public. Interestingly, he mentioned receiving a positive response from an administration previously more hands-off regarding AI regulation, especially after the recent Mythos incident raised alarms.

Highly regarded in the AI community, Hassabis noted that leaders from other significant AI institutions generally agree with his directional vision. “This is where the industry needs to go,” he reflected on the feedback received. His goal is to have the new regulator operational by the end of the year, which seems quite aggressive.

The proposed oversight structure would resemble the Financial Industry Regulatory Authority (FINRA), which oversees Wall Street firms under SEC guidance. Initially, companies developing advanced AI systems would voluntarily submit their models to this new agency for a thorough safety evaluation, focusing on cybersecurity risks, biological threats, and deceptive practices at least 30 days prior to their release.

If the testing protocol proves effective, Hassabis intends to formalize it, shifting from a voluntary to a mandatory submission process for frontier AI models in the U.S. market. The governance model would feature an independent board of directors composed of Turing Award winners and other distinguished professionals, along with representatives from industry, government, and the open-source community.

This proposed framework would apply universally to all frontier-class AI models, irrespective of their country of origin or whether they are open or closed systems. The criteria for what constitutes a Frontier Class model would be updated regularly as AI capabilities evolve. Hassabis suggested that obtaining the Frontier designation would be seen as a mark of prestige, stating, “I think this is a very nice, prestige asset.”

This proposal follows recent emergency actions taken by the Trump administration concerning Anthropic’s Mythos and Fable AI models. Hassabis called that situation “a bit of a wake-up call,” indicating a need for stronger regulatory frameworks instead of ad-hoc directives. Anthropic faced an abrupt halt on its powerful models due to export control orders and spent two and a half weeks in negotiations without clear procedures in place.

To avoid similar issues, OpenAI proactively decided to limit access to its GPT-5.6 models to government-approved partners at launch. The public release occurred the following week after negotiations and testing with the Commerce Department concluded.

Additionally, Anthropic CEO Dario Amodei is pushing for binding regulations and has proposed an FAA-like agency empowered to block unsafe AI models. Leaders from Google’s Gemini system and Anthropic’s Claude system generally agree on the need for federal oversight, though they differ on which organization should assume this regulatory role.

Hassabis remarked that achieving artificial general intelligence—defined as an AI system with human-like cognitive abilities—could happen within the next few years. He expressed that humanity is on the brink of something revolutionary, noting, “We have basically discovered how to make sand think. It’s a miracle.”

It is crucial for leaders to manage AI effectively to ensure it does not control humanity.