Google DeepMind chief executive Demis Hassabis called for the United States to create a new body to test the world’s most powerful AI models before they reach the public, in a personal essay published on July 14.
Writing that artificial general intelligence is “probably only a few short years away,” the Nobel laureate argued that capabilities are outpacing society’s understanding of them and that humanity has a “precious window” to set rules before AGI arrives. His central proposal is a Frontier AI Standards Body that would develop rigorous, regularly updated tests for dangerous capabilities in areas like cybersecurity, biological threats and deception, and could, if risks became severe enough, coordinate a slowdown in development across leading labs.
The design borrows from finance. Hassabis suggests modeling the body on FINRA, the private, industry-funded organization that polices Wall Street under government oversight, with a majority-independent board of leading technical experts alongside industry, government and open-source representatives.
Funding would come mostly from the AI industry itself, which he says is necessary to attract top talent and the computing power needed for large-scale testing. Under the plan, labs would first voluntarily share “frontier-class” models for review up to 30 days before release, and once the process proved robust, passing it could become mandatory to deploy a model in the US market. The rules would apply to any qualifying model regardless of its country of origin or whether it is open or closed, while smaller models from startups and academia would be exempt.
The timing is pointed. The proposal follows the Trump administration’s recent ad hoc interventions, when it forced Anthropic to pull its Claude Fable 5 and Mythos 5 models and limited OpenAI’s GPT-5.6 rollout, actions Hassabis casts as evidence that the current improvised approach is not working and that a systematic framework is needed instead. His 30-day voluntary-review structure closely mirrors the mechanism in Trump’s June 2 executive order, positioning the essay as an industry-shaped alternative to whatever formal regime Washington produces. Hassabis frames a US-led effort as a first step toward eventual international consensus.
The Case He Is Making
Hassabis grounds the argument in urgency and sequence. He warns that today’s AI cyber risks are “warning shots,” and that within roughly 18 months far graver biological and potentially nuclear-relevant capabilities could appear in open-source models beyond any government’s control. His prescription, which he calls “cautious optimism,” is to build guardrails now rather than after a catastrophe forces them under pressure.
He pairs this with an expansive vision of the upside, describing AGI’s potential impact as ten times that of the Industrial Revolution at ten times the speed, capable of accelerating drug discovery, clean energy and an era of abundance. The proposal’s appeal is that it is technically focused and adaptive, designed to keep pace with a fast-moving field rather than freeze it.
The Skeptic’s View
The proposal invites hard questions about who writes the rules. An industry-funded body whose evaluations are initially developed “in consultation with Frontier Labs” raises the specter of regulatory capture, letting the largest incumbents, DeepMind among them, help shape the standards they must meet and potentially locking out smaller competitors.
Critics of the FINRA model note that self-regulatory organizations can prioritize member interests over the public, and Hassabis acknowledges the body should eventually build independent, held-out tests to avoid labs gaming the benchmarks.
There is also a competitive subtext: mandatory pre-release review favors well-resourced labs that can absorb the delay and compliance cost, and the framing of open-source models as a channel for “bad actors” cuts against developers who release weights openly.
Finally, there is the messenger problem, since the CEO of a company racing to build AGI is proposing the very system that would govern that race. Whether such a body could be genuinely independent while depending on industry money and expertise is the question that will decide if the idea is credible governance or sophisticated self-regulation.
The intervention also arrives amid a wave of similar proposals from rival lab chiefs, with Anthropic CEO Dario Amodei publishing his own essay urging binding, FAA-style safety testing for frontier models.
