Google DeepMind's Boss Wants a U.S. Government-Backed Body to Vet AI Before It Ships

Key points

• Google DeepMind CEO Demis Hassabis published a proposal on Tuesday calling for a U.S.-led, federally overseen body to review AI models before public release.
• Hassabis warned that frontier AI models, the most powerful systems available today, already pose cybersecurity risks and could soon enable nuclear or biological threats.
• The proposed body would require AI companies to submit new models for review up to 30 days before launch, starting voluntarily, then becoming mandatory.
• Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have each made similar calls for an AI oversight body in recent weeks.
• Funding for the watchdog would "likely" come from the AI industry itself, Hassabis said.

Demis Hassabis, the CEO of Google DeepMind and a 2024 Nobel laureate in chemistry, published a detailed proposal this week for a new American institution: a standards body, essentially a watchdog agency, that would test and approve powerful AI models before they reach the public.

Think of it as the Food and Drug Administration, but for AI software.

Hassabis said the body should be modelled on FINRA, the Financial Industry Regulatory Authority, which is the organisation that licenses and polices stockbrokers and brokerage firms in the United States. FINRA is federally overseen but run as a public-private partnership, meaning government and industry share the table.

The reason Hassabis wants this now: he believes the stakes are rising fast. "We've already seen the challenges frontier models pose for cybersecurity, and other threats including nuclear and bio risks may soon emerge," he wrote.

His proposal says AI companies would voluntarily submit new models to the body for review up to 30 days before launch. Once the body proves it works, that submission would become mandatory for any model sold in the U.S. market.

The testing would look at whether a model can be tricked into bypassing its own safety rules, whether it shows signs of deception, and whether its outputs can be traced back to an AI source. That last point matters for ordinary users: it is how your bank, your doctor's office, or your employer would one day verify whether a document was written by a person or generated by software.

Money is the obvious question. Hassabis said the body would need "substantial" funding to hire top technical staff and run the computing infrastructure needed for large-scale testing. He expects the AI industry to pay for most of it.

This is not a lone voice. As first reported by CNBC Tech, Hassabis and Anthropic CEO Dario Amodei raised the idea at a G7 meeting that included President Donald Trump last month. OpenAI CEO Sam Altman published a similar argument in the Financial Times earlier this month.

Yet the political path is bumpy. The Trump administration recently imposed temporary export controls on an advanced Anthropic model, then asked OpenAI to slow the rollout of a new one. Hassabis is essentially asking the same government that is already restricting AI to now build a formal, funded, staffed institution around it.

What does this mean for people who use AI at work?

If this body is ever created, models you use through work software or consumer apps would have to pass a safety review before launch. That could mean slightly slower release schedules, but it also means an independent set of eyes checked whether the tool can be manipulated into producing dangerous output. For most users, the practical effect would be invisible, unless it quietly prevents the next major security incident involving an AI product.

The U.S.-China dimension adds pressure. Chinese AI models from companies like DeepSeek are winning customers among American businesses, partly on cost. U.S. lawmakers are now considering restrictions on their adoption. A domestic standards body would almost certainly apply to foreign models sold in the U.S. market too.

Honest takeaway: An AI watchdog is not law yet, and this proposal faces real political headwinds. Watch whether Congress picks it up, not whether tech CEOs keep endorsing it.