Demis Hassabis pushes for a US-led global AI watchdog as frontier risks mount

Updated July 14, 2026 3:20 pm

Google DeepMind chief Demis Hassabis is calling for a global artificial intelligence watchdog, led by the United States, that could slow or block frontier AI systems if they are judged too dangerous to release. His proposal matters because it would create a new layer of international oversight at a moment when the most advanced AI models are rapidly approaching capabilities that many experts say could outpace existing rules.

In a new blog post, Hassabis argues that the US is best positioned to anchor such an institution because of its technical depth, economic influence, and central role in the AI industry. He says the system should bring together independent experts and open-source representatives, with the power to review frontier models before launch and trigger an industry-wide pause when risks rise too high.

The idea arrives as governments struggle to keep up with the pace of AI development. There is still no dedicated global framework for AI governance, and even the US does not have a comprehensive national rulebook covering the most powerful systems. Hassabis is urging policymakers to move before the technology becomes even more difficult to control.

What Hassabis is proposing

Hassabis wants a standing international body that would function like a sector-specific regulator for advanced AI, with the authority to evaluate cutting-edge models before they are released to the public. The institution would not simply issue guidance after the fact; it would be able to intervene early if a system appears too risky.

The proposal, as described in his essay, is aimed at frontier models — the largest and most capable AI systems being trained today. Those models can already produce text, code, images, and strategic reasoning at a level that has raised questions about misuse, safety testing, and concentration of power.

He envisions an oversight structure that could resemble existing financial regulators in spirit, drawing on independent specialists rather than leaving decisions solely to the companies building the systems. Hassabis says the body should include researchers, safety experts, and open-source community voices so that the review process is not dominated by one constituency.

Why the watchdog would be US-led

Hassabis says the United States should take the lead because it has the strongest combination of AI talent, capital, and global influence. In his view, a US-backed institution would have the best chance of setting norms that other countries and companies would follow.

That argument reflects a broader reality in AI governance: the country home to many of the leading labs is also the one most likely to shape international standards. A US-led framework could carry more weight than a purely advisory global forum, especially if it is tied to the firms and researchers driving frontier development.

At the same time, the proposal does not envision a purely American system. Hassabis is calling for a global watchdog, even if the operational center of gravity sits in the US. That balance is meant to address a common criticism of AI governance efforts: rules built by one nation may not be sufficient for a technology that crosses borders instantly.

Why this debate is intensifying now

The push for stronger oversight comes as AI models become more capable and more widely deployed. Hassabis argues that general-purpose systems are progressing toward a stage where the stakes will no longer be limited to misinformation, copyright disputes, or workplace disruption.

In his essay, he warns that artificial general intelligence may be only a few years away. He frames the current moment as an early phase in a much larger technological transition, suggesting that today’s decisions could determine how safely the next generation of systems is built and deployed.

That warning has become increasingly common among AI leaders, but Hassabis’s proposal stands out because it offers a concrete governance mechanism rather than a broad appeal for caution. Instead of simply asking companies to self-regulate, he is suggesting a formal body with real authority.

Hassabis writes that the world is entering a historic technological shift and that society is still only at the beginning of it, making stronger oversight urgently necessary.

How frontier AI risk is being framed

Experts have long debated what counts as an unacceptable AI risk. The list now often includes cyber abuse, bioweapon assistance, autonomous decision-making failures, and model behavior that cannot be reliably predicted or contained.

Hassabis has recently aligned himself with that concern. Last month, he signed a statement calling for stricter safeguards against the use of AI in bioweapons development, underscoring that his warning is not limited to abstract future scenarios. It reflects a growing consensus among some top researchers that certain capabilities should be tightly monitored before they spread.

The concern is not only that individual models could be misused. It is also that multiple companies may be racing to release increasingly powerful systems without a shared standard for when testing is enough, what red lines should exist, or who gets to decide when a model is too dangerous to ship.

How would the watchdog work?

Hassabis’s framework would give the new organization three main responsibilities: review frontier models before public release, assemble qualified experts to assess safety and capability, and coordinate pauses if a system crosses a risk threshold. That would make the body more active than most current advisory groups.

The model is meant to be preventive rather than reactive. In other words, the idea is to identify a serious risk before a system reaches mass deployment, rather than waiting for real-world harm to occur and then trying to contain it.

While Hassabis has not published a detailed legal blueprint, the concept suggests a hybrid of standards-setting, technical auditing, and emergency intervention. It would likely need buy-in from governments, labs, and cloud providers to have practical force.

Who is backing the idea behind the scenes?

According to Axios, Hassabis has spent months quietly building support for the concept. That reportedly included briefings for the Trump administration, other AI laboratories, and European officials, suggesting he is trying to assemble a transatlantic coalition before pushing the idea into the open.

He told Axios that the signals he has heard from the Trump administration have been encouraging. That is significant because any watchdog with real influence would likely require support from the White House and buy-in from the companies building the most advanced systems.

The reported outreach also shows that Hassabis is not treating the proposal as a purely theoretical thought experiment. He appears to be testing whether the political and industrial climate is favorable enough for a formal institution to emerge this year.

Why the US political moment matters

The timing is important because the US remains central to AI development, but Washington has not settled on a clear regulatory regime. Federal agencies have issued guidance and held hearings, but lawmakers still have not enacted a comprehensive national framework for frontier AI.

That leaves a gap Hassabis is trying to exploit. If the US is already the de facto center of AI innovation, then a US-led watchdog could become the default mechanism for managing the most powerful systems before they spread globally.

Whether that would be politically feasible is another question. Any body with enforcement power would have to navigate competition concerns, debates over free speech and innovation, and resistance from companies that move quickly and value autonomy.

What makes this different from existing AI governance efforts?

This proposal is different because it imagines an institution with teeth. Many current AI governance initiatives focus on voluntary commitments, safety pledges, or general principles. Hassabis is instead describing something closer to a standing authority with the ability to review and, if needed, slow deployment.

That is a meaningful shift. Voluntary frameworks can encourage common practices, but they often fail when commercial incentives push firms to race ahead. A watchdog with review powers would create friction in that race, forcing companies to justify release decisions more publicly.

It would also create a more durable structure than ad hoc government summits or one-off policy statements. If established properly, the body could become a long-term institution that grows alongside the technology it monitors.

How the proposal fits into a wider AI safety movement

Hassabis’s comments are part of a broader push among influential figures to think seriously about AI’s long-term effects. Recently, economists and technology leaders — including Anthropic cofounder Jack Clark and former Google CEO Eric Schmidt — urged governments to pay closer attention to the economic shockwaves that advanced AI may create.

That conversation has expanded beyond job disruption and market concentration to include national security, model misuse, and the possibility that AI systems could become more capable than the institutions designed to oversee them. Hassabis’s watchdog idea is one response to that widening list of concerns.

It also reflects a shift among some industry leaders away from pure optimism. Many still believe AI will drive major scientific and economic gains, but they increasingly argue that the upside will only be sustainable if governments and labs create meaningful guardrails first.

The broader stakes for AI companies and governments

If a global watchdog were created, it could reshape how AI companies plan launches, conduct safety evaluations, and communicate risk. Firms might need to share more technical detail, submit models to external review, and wait longer before deploying highly capable systems.

That could slow product cycles, but it could also reduce the risk of a serious failure that damages public trust. For governments, a watchdog could offer a centralized way to monitor a fast-moving field that no single agency currently understands in full.

Still, any such system would face difficult questions:

• Who chooses the experts?
• What qualifies a model as frontier AI?
• How is risk measured consistently across different systems?
• What happens if a company refuses to comply?
• How can oversight remain international without becoming too slow or politicized?

Those questions suggest that the hardest part of Hassabis’s proposal may not be the idea itself, but the institutional design. A watchdog powerful enough to matter would need legitimacy, technical credibility, and clear enforcement mechanisms — all while operating in a competitive global market.

How likely is a year-end launch?

Hassabis reportedly hopes the organization could be operating before the end of the year, but that timeline would be ambitious. Creating a new global institution usually requires negotiations over mandate, membership, funding, and authority, and those discussions can take far longer than a few months.

Even so, his push is notable because it suggests momentum is building inside the AI elite for something more formal than the current patchwork of voluntary commitments and government hearings. If major labs and policymakers take the idea seriously, the conversation could move quickly.

For now, the proposal is best understood as a warning and a blueprint at once: a warning that AI progress is accelerating toward more serious risks, and a blueprint for how industry and government might try to catch up.

Timeline of the watchdog push

The following timeline captures the key moments surrounding Hassabis’s latest call for global AI oversight.

What happens next?

The next test is whether governments and AI leaders treat Hassabis’s proposal as a serious starting point or just another warning about the future. If his private outreach has already opened doors in Washington, the idea could gain traction faster than similar calls for regulation have in the past.

But if the industry remains divided over how much power any watchdog should have, the plan could stall. The tension between innovation and restraint has defined AI policy for years, and Hassabis is now asking the world to move from debate to institution-building.

For an executive who has helped shape some of the most advanced AI systems in the world, the message is clear: the age of informal oversight may be ending, and the next phase may require a global referee with real authority.

Quick facts

• Who: Demis Hassabis, CEO and cofounder of Google DeepMind
• What: A proposal for a global AI watchdog led by the US
• Why now: Frontier AI models are becoming more capable, while regulation remains fragmented
• Core power: Pre-release model review and the ability to coordinate slowdowns
• Target timing: Hassabis reportedly hopes it could be running by year-end