Google DeepMind Chief proposes US lead a global agency to oversee AI models

Frontier models are defined as highly sophisticated systems capable of performing complex tasks with significant potential impact on crucial sectors such as security, economy, science, and information. The Nobel laureate in Chemistry suggested that the structure of this new body could be inspired by existing organizations, encompassing independent scientists and open-source community representatives, responsible for assessing both technical risks and potential social effects. If a system is classified as dangerous, the entity would have the power to advise a coordinated slowdown in the sector.

Current Oversight in the US

Even without the formal existence of this new agency, the US government already exercises control over the recent releases from major national AI companies, holding the final say on the distribution of these new models. Recently, the Donald Trump administration vetoed models from Anthropic and OpenAI due to the high risk of these tools being used by cybercriminals.

Sam Altman's company was forced to postpone the release of the GPT-5.6 model to the general public until just last week. Concurrently, the owner of Claude had its release of the Fable 5 and Mythos 5 models restricted to foreigners in mid-June, even for those located within the US. Authorization to expand the use of Fable 5 was only granted this month, while Mythos 5 remains limited to corporate partners, according to Anthropic's own decision.

According to the portal Axios, Hassabis has been debating this proposal with sectoral leaders and government officials for several months, including members of the Donald Trump administration. He informed the portal that he expects the organization to be structured this year, and that the reaction within the government has been described as 'very positive.'

Perspectives on Artificial General Intelligence

For Hassabis, the speed at which AI models are evolving makes the creation of this regulatory body imperative. He believes that Artificial General Intelligence (AGI) may be only a few years away from being achieved. AGI refers to systems that can match or surpass human capability in various intellectual activities. Although there is still no consensus on the timing or possibility of reaching this level, the intense competition for this superintelligence leads the industry to invest billions annually in infrastructure and hiring.

The last year was considered disappointing for companies seeking to achieve AGI, as recent models, such as Meta's Llama 4 family and OpenAI's GPT-5 itself, did not show significant advances in terms of learning.