Demis Hassabis, the CEO of Google DeepMind, has put forward a detailed plan for a new US agency that would test the world's most advanced AI models before they are released. The proposal, published on his personal channels and backed by months of private meetings with governments and industry leaders, is the most concrete governance plan to come from inside a major AI lab.
The idea is modeled on FINRA, the private US organization that polices stockbrokers under government supervision. FINRA is funded by the financial industry itself, governed by independent experts, and operates faster than a traditional government regulator because it sits inside the industry it oversees. Hassabis wants the same structure for AI: paid for by the big labs, staffed by credible independent scientists, and answerable to Washington.
The testing would happen before release, up to 30 days in advance, and would check for specific dangers: whether the model could help someone build a biological weapon, carry out a cyberattack, or deceive users in ways that could cause serious harm. At the start, participation would be voluntary. Once the process is proven to work, it would become a formal requirement for any lab wanting to sell in the US market.
The body would also have the authority to call a slowdown across the whole industry if a model is judged too dangerous to release. That is a meaningful power, and one that would apply equally to Google, OpenAI, Anthropic, and any other lab building at the frontier.
The timing of this proposal matters. The Trump administration recently froze Anthropic's most powerful models overnight using export controls, with no established rules or process. Anthropic spent two and a half weeks negotiating their release. OpenAI restricted its own latest model to government-vetted partners at launch to avoid the same fate. The absence of any formal system made things messy and unpredictable for everyone. Hassabis is essentially arguing that a clear process, even a strict one, is better for the industry than ad hoc government interventions.
Hassabis has reportedly been building support for this plan for months, briefing the Trump administration, fellow lab leaders, and European officials before going public. He says the early signals from Washington are positive. At the G7 summit last month, both Hassabis and Anthropic CEO Dario Amodei pushed for a US-led international coalition on AI standards.
There is a legitimate concern worth naming. A regulatory regime that requires expensive pre-release testing and independent review would, by design, be easier for large, well-funded labs to navigate than for smaller competitors or open-source developers. Hassabis addresses this by proposing that smaller companies and academic researchers are exempt, covering only the handful of labs building the most powerful models. Whether that line holds in practice is a separate question.
For business operators, the practical takeaway is this: formal AI regulation at the most powerful level appears to be coming, and the companies building the most advanced systems are now actively helping design it. That is a signal about where the industry's own risk assessment stands. If the people building these systems are asking for oversight, it is worth paying attention to why.