The Trump administration has circulated a draft executive order in recent weeks that would create a self-regulatory organization for leading artificial intelligence companies, according to two people familiar with the document, marking a step toward establishing the new body but facing significant obstacles before it can be finalized.
The order, which would need approval from senior officials including President Donald Trump, has seen its momentum stall recently, according to three people familiar with the matter. The concept stems from a proposal last month by Demis Hassabis, then chief executive of Google DeepMind, who suggested a public-private body to oversee the AI industry.
Hassabis called for a standards body modeled on the Financial Industry Regulatory Authority (FINRA), a self-regulatory organization that has long overseen broker-dealers and securities professionals in the United States. Two sources said the executive order for the new AI body originated from multiple proposals by senior administration officials. Bloomberg reported in July that Treasury Secretary Scott Bessent had proposed establishing such a regulator.
The specific reasons for the executive order's stagnation remain unclear, but the idea of creating an AI regulator modeled on FINRA has faced strong opposition from David Sacks, the former White House AI and crypto czar. On his podcast last week, Sacks called the proposal a "terrible idea," specifically criticizing mandatory testing of models before their release and the requirement for the self-regulatory organization to report to the government. He instead advocated for a model like the Motion Picture Association, where the industry independently manages movie ratings without needing to report to a government agency.
Details of the executive order and the self-regulatory body have not been made public, and the timeline for implementation remains uncertain. Another AI-related executive order issued by the Trump administration in June, which set up a voluntary evaluation framework for frontier models, took months of negotiations to complete. A White House spokesperson said all policy announcements would be made directly by the president, adding that discussions about potential executive orders were purely speculative.
Model evaluation has been a politically charged issue during the Trump presidency. The administration renamed and downsized the U.S. AI Safety Institute established under Biden. Pushed by Sacks, the Trump administration has generally taken a deregulatory approach, rolling back many Biden-era rules. However, as AI models like Anthropic's Mythos continue to become more capable, administration officials have begun calling for additional safety measures for the technology.
The June AI executive order established a voluntary testing framework, driven largely by concerns over cybersecurity capabilities of new models. But the tech industry still has many unresolved questions about how this framework will actually operate. Although the framework was finalized in early August, the White House has not yet publicly released it.
Meanwhile, several industry leaders including Hassabis have proposed their own regulatory approaches. In his article, Hassabis wrote that the FINRA-style AI standards body would need to develop evaluation procedures and work with "appropriate federal agencies." The body would classify models as frontier-level if they meet thresholds it sets, while encouraging model developers to adopt industry best practices such as publishing model cards, implementing cybersecurity protections, and conducting personnel background checks.
Frontier AI labs would need to submit their models to the body up to 30 days before release. Much of the language in Hassabis's proposal closely aligns with the White House's earlier voluntary framework, raising questions about how the two systems would interface. While Hassabis's article received broad support, public opposition from figures like Sacks who still hold influence within the Trump administration could potentially derail the entire initiative.
Sacks posted on X last week criticizing the "AI version of FINRA" concept, comparing it to the DMV, and mistakenly attributed the plan to Anthropic CEO Dario Amodei rather than Hassabis. Sacks wrote that as rules continue to pile up to guard against theoretical potential harms, the regulatory process will only become more convoluted.
Comments