Demis Hassabis is urging the creation of a U.S.-led standards body to independently test frontier AI models before public release.
Google DeepMind CEO Demis Hassabis has proposed creating a U.S.-led standards body to independently test the world’s most advanced AI models before they are released.
The proposal would cover both open and closed frontier models developed in the U.S. and abroad, focusing on cybersecurity risks, dangerous capabilities and whether built-in safeguards can be bypassed.
Hassabis argues that artificial general intelligence could arrive within a few years, leaving a limited window to establish consistent oversight.
Rather than introducing a separate licensing process for every model, the proposed organization would create a common testing framework that developers could use before deployment. The body would be modeled on the Financial Industry Regulatory Authority (FINRA), with independent experts evaluating models against evolving safety benchmarks.
The proposal builds on the White House’s recent voluntary pre-release testing framework for frontier AI models, but goes a step further by calling for a permanent standards organization.
The debate is now moving beyond whether frontier AI should be tested and toward who should conduct those evaluations, how testing should be standardized, and when governments should have the authority to delay deployment of high-risk models.
