DeepMind's Demis Hassabis proposes a U.S. body to evaluate advanced AI models before release, citing potential AGI risks.
Demis Hassabis, the CEO of Google DeepMind, has proposed the creation of a standards organization in the United States to assess advanced AI models prior to their public release. He predicts that artificial general intelligence (AGI) could emerge in the next few years, with implications akin to those of electricity or fire.
"When we look back on this time in the coming decades, I think we will realize that we stood at the foothills of the singularity — nothing less than the beginning of a new era for humanity," Hassabis stated.
Hassabis describes AGI as a system that would replicate all human cognitive functions, suggesting that its potential impact would far exceed that of the internet or mobile technologies. He emphasized that while AGI could lead to breakthroughs in medicine, clean energy, and new materials, the rapid evolution of AI capabilities is outpacing our understanding of their risks.
The current generation of advanced AI models already presents cybersecurity threats, and as their capabilities expand, they could also introduce biological and nuclear risks. Hassabis expressed concern about the rise of autonomous systems that could execute complex tasks and improve themselves without oversight. He called for additional safeguards to ensure proper control over these technologies.
Proposal for a New Standards Body
Hassabis suggested establishing a Frontier AI Standards Body, inspired by the Financial Industry Regulatory Authority (FINRA). This organization could function as a public-private partnership or a self-regulatory body under federal supervision, primarily funded by the AI sector.
Leadership would include independent technical experts and members from the open-source community, with the goal of developing evaluation protocols in conjunction with U.S. federal agencies and national laboratories.
The proposed body would implement specialized assessments to determine if a model qualifies as advanced, focusing on cybersecurity, biological threats, and other critical areas. It would also evaluate attempts by models to bypass safety measures or mislead users.
Voluntary to Mandatory Evaluation Transition
Initially, companies would voluntarily submit their advanced models for evaluation at least 30 days before release. Once the evaluation system proves effective, it may become a legal requirement for models to pass assessment prior to entering the U.S. market. Hassabis also recommended establishing a market for independent auditors and using the U.S. framework as a basis for international standards.
In June, an executive order from U.S. President Donald Trump initiated voluntary evaluations of advanced AI models' cybersecurity capabilities, allowing companies to grant authorities access to their models for a limited period before sharing them with external partners.
Hassabis acknowledged the uncertainty surrounding the future development of AI technology, noting that even experts have differing opinions on various issues. He referenced earlier reports about the Trump administration requesting OpenAI to delay the public release of GPT-5.6 over safety concerns.
Additionally, U.N. Secretary-General António Guterres has called for global regulations on advanced AI systems to protect vulnerable populations and restrict autonomous weaponry. In a related movement, activists in San Francisco have urged major AI companies to pause the development of more powerful models to focus resources on safety, while former OpenAI researcher Daniel Kokotailo has suggested halting the AI race until 2040.
