Trump AI Watchdog Plan Reaches White House With Its Powers Unsettled

• A proposal developed with Treasury Secretary Scott Bessent would create an independent AI regulator reporting to the Securities and Exchange Commission, according to a report based on unidentified sources.
• President Donald Trump had not reviewed the plan, while its assessments, funding and the SEC’s role remained unspecified.
• The reported administration plan is distinct from Google DeepMind CEO Demis Hassabis’s blueprint, which would begin with voluntary reviews and later require frontier models to pass before U.S. deployment.
• Washington already has a voluntary, 30-day federal review process. The live question is who could block a release, using which tests and under what oversight.

The Trump administration is considering a more durable system for reviewing powerful AI models after its recent interventions frustrated major labs. But the proposal now at the White House is still a political and institutional sketch, not a finished safety regime.

The proposal has reached Wiles, not Trump

A Bloomberg report republished by Business Standard, citing people familiar with the matter, says Bessent helped develop a proposal for an independent AI regulator that would report to the SEC. White House Chief of Staff Susie Wiles was reviewing it, but Trump had not yet done so, one of those people said. The framework remained under deliberation and subject to change.

That status limits what can be claimed about the plan. The report says it would vet advanced models with industry input and give Wall Street and technology companies a larger role in setting safety standards. It does not specify the assessments, the funding model or precisely how the SEC would oversee privately held AI developers. A White House official said the administration was considering multiple proposals to strengthen U.S. leadership in AI and cybersecurity.

The immediate constituency is split. Financial firms want protection from AI-enabled cyber risks. Frontier labs want predictable rules after the administration imposed export controls that temporarily disabled Anthropic’s Fable 5 and Mythos 5 models and sought changes before OpenAI released Sol, according to the report. Anthropic and OpenAI called those interventions excessive relative to the safety issues officials identified.

Trump has also resisted rules he thinks could weaken the U.S. position against China. When he postponed an earlier signing ceremony in May, he said he did not want to obstruct the country’s lead. He later signed a narrower executive order establishing voluntary federal review of national-security risks in advanced systems for no more than 30 days before public release. The White House described that process more specifically as voluntary sharing of cutting-edge cyber models to protect critical infrastructure and improve government defenses.

The administration draft and the Hassabis blueprint are not the same plan

The reported proposal broadly resembles a framework from Hassabis, the Google DeepMind co-founder and CEO whose company would be among the best-resourced participants in any frontier-model regime. He had spent months briefing administration officials, other lab leaders and European officials, and said he wanted a new body operating within months, ideally before the end of 2026, in an interview.

Hassabis’s published proposal supplies details absent from the administration reporting. A federally overseen, substantially industry-funded standards body would define “Frontier-class” models through changing capability thresholds. It would work with federal agencies and national laboratories on tests for cyber capabilities, biological threats, attempts to bypass safeguards and signs of deception.

Frontier labs would initially submit qualifying models voluntarily, up to 30 days before release. Only after the protocol had proved “effective and robust,” Hassabis wrote, would passing become mandatory for deployment in the U.S. market. His framework could apply to foreign and domestic models, whether open or closed; systems below the frontier threshold, including those from startups or academia, would be exempt.

That sequence matters. The administration has not publicly adopted Hassabis’s mandatory gate, his threshold or his proposed scope. Treating his endpoint as current government policy would turn an industry executive’s recommendation into an official decision that has not been made.

Other lab leaders support outside testing but offer different centers of authority. A comparison of their proposals described Anthropic CEO Dario Amodei’s preference for an FAA-like federal agency able to block a release from the start, and OpenAI CEO Sam Altman’s preference for a U.S.-led international system. Their convergence on review does not resolve whether government, an industry-funded body or an international forum should have the last word.

Tests can inform a gate before they can justify one

The federal government already has model-evaluation capacity. In an April evaluation, the National Institute of Standards and Technology’s Center for AI Standards and Innovation tested DeepSeek V4 Pro on nine benchmarks spanning cyber, software engineering, natural sciences, abstract reasoning and mathematics. Two tests were held out and uncontaminated, according to the agency’s methods and results.

That exercise showed why independent test selection matters. DeepSeek’s own reported results placed V4 Pro around models released roughly two months earlier. CAISI’s suite placed it near GPT-5, released about eight months earlier. Those comparisons measured capability, not a model’s overall safety, and CAISI disclosed that prompts, scaffolding, token budgets and scoring methods can affect individual benchmark results.

The same evidence cautions against converting a benchmark score directly into permission to enter a market. In January, NIST said practices supporting valid, transparent and reproducible AI evaluation were “only beginning to emerge.” Its draft guidance covered automated benchmarks, which the agency said cannot meet every evaluation objective.

Hassabis proposes quarterly revisions at first, retirement of saturated benchmarks and, eventually, tests developed independently of the labs. Yet his blueprint leaves the pivotal judgment undefined: who determines that the protocol has become robust enough to control access to the U.S. market?

FINRA is a governance model, not a shortcut

The FINRA comparison can obscure as much as it explains. FINRA is industry funded, but its authority sits inside an established public oversight system. The SEC monitors its operations and programs, reviews changing risks and inspects its work.

From fiscal 2021 through 2023, the SEC initiated 21 program inspections of FINRA. Those reviews covered eight of 10 areas specified by the Dodd-Frank Act, including conflict-of-interest management and governance transparency; the other two were covered in reviews outside that period, according to a federal audit.

That record does not show that the same arrangement will work for AI. It shows that “self-regulation” requires continuing government inspection, performance measures and channels for complaints and referrals. The reported AI proposal has not disclosed comparable machinery.

Industry money could pay for the compute and specialists that Hassabis says large-scale testing needs. It would also make regulated companies the main funders of a body that might ultimately determine whether their products reach U.S. customers. No public budget, fee schedule or insulation from funder pressure has been specified.

The competitive trade-off runs through the threshold design. Exempting non-frontier systems would spare most startups and academic projects. But once a smaller or open-source developer crossed the threshold, it would face a process initially shaped by labs that already possess security teams, government relationships and technical staff. Critics therefore warn that certification could improve safety while also entrenching the largest companies.

The next decision is whether the body advises or blocks

The first signal will be whether Wiles advances this proposal for Trump’s review, and whether the administration chooses it over the other options the White House says are under consideration. Until then, the FINRA analogy is a direction of travel, not settled policy.

Any fuller plan would need to answer the questions the current reporting leaves open: who sets and audits the tests, who appoints the independent members, how laboratories can contest a result, how sensitive model information is protected, and what authority the government retains over the regulator.

Most important, the administration must decide whether the body would advise on a voluntary release process or hold mandatory market-access power. Evidence that independent evaluation can change a model comparison supports outside testing. It does not by itself establish that today’s tests, funding structure or oversight are ready to decide which models Americans may use.