The head of Google's artificial intelligence unit proposed that the United States establish a regulatory organization to evaluate emerging AI systems for potential threats to national security, such as vulnerabilities in cybersecurity and biological domains. On Tuesday, the head of Google DeepMind, who received a Nobel Prize, published an article on X advocating for immediate action to address risks associated with artificial general intelligence—systems matching or surpassing human intellectual capacity—and observed that current advanced AI poses cybersecurity threats while future iterations may create vulnerabilities in nuclear and biological domains.
Hassabis recommended establishing a collaborative initiative between government and private sector entities, with federal oversight, to address these challenges. Requests for comment were sent to the White House, State Department, and Commerce Department.
The remarks follow reports from approximately one month earlier that Hassabis and Anthropic's Amodei advocated for standards and regulations guided by a coalition at a G7 gathering attended by tech executives and government leaders, including Trump. Earlier this month, a well-regarded news outlet published an article in which Sam Altman of OpenAI advocated for establishing a comparable regulatory organisation.
While industry figures increasingly advocate for dedicated AI oversight bodies, disagreements between government authorities and technology companies over the supervision of advanced AI systems have intensified. Anthropic held discussions with government representatives following a brief period when the Trump administration placed restrictions on exporting one of its sophisticated models. The company encountered limitations after federal authorities asked it to delay releasing a new artificial intelligence model.
Hassabis noted that America's economic and technical position offers advantages for developing an AI framework, suggesting the possibility of creating a standards organisation through federal oversight with public-private collaboration or self-regulation similar to FINRA, featuring a board with independent technical specialists and open-source community representation.
Hassabis noted that the proposed organisation would require considerable financial support to recruit top-tier technical specialists and supply the computational infrastructure needed for extensive testing, while FINRA oversees brokerage operations and securities exchanges across the United States. He noted that the financing would probably be sourced from the sector.
Frontier laboratories would at first voluntarily submit models to the organisation for assessment as much as 30 days prior to public availability, with this practice transitioning to a requirement for deployment within the United States market once demonstrated to function effectively, and Hassabis noted that particular agentic artificial intelligence evaluations might examine efforts to circumvent safety protocols or indicators of misleading behaviour, while verifying adherence to recommended practices including digital watermarking of AI-produced imagery and creation of human-interpretable output tokens for understanding model logic.
Demands for stronger regulation of artificial intelligence accompany intensifying competition between Washington and Beijing in model development and implementation. U.S. businesses are increasingly adopting recent artificial intelligence models from Chinese developers such as DeepSeek and Z.ai, which observers view as comparable in performance to top-tier systems produced by Anthropic and OpenAI, particularly as expenditure on AI technology continues to grow. Legislators in the United States are now examining measures to limit domestic companies' increasing use of artificial intelligence models from China, with the State Department indicating concerns to a news outlet.
