Two competing designs for a national AI regulator are circulating inside the White House, and neither one is actually a government regulator. That is the durable fact underneath a story that also includes a disputed phone call. It matters because the pattern predates this particular call and will outlast it.
Politico reported it this week, resting the account on a senior White House official who spoke anonymously about a private exchange. On a call with President Trump in the week of August 17, the outlet says, Mark Zuckerberg pushed back on the idea of a national regulator for AI. Politico separately cited a second person familiar with the matter, who said Zuckerberg never pushed Trump to reverse course; rather, Zuckerberg argued that anyone appointed to such a body ought to share the light-touch instincts the president himself has shown, and it was Trump, not Zuckerberg, who placed the call. Meta declined to comment on the record, and neither Meta nor the White House has confirmed the specifics beyond what Politico’s sourcing describes. None of that content is independently verified here; treat the call as reported and attributed, not established.
What is verifiable, and separate from the sourcing question, is what has actually been proposed and what happened to each proposal. Two designs remain on the table. The first, popularized by Google DeepMind’s Demis Hassabis in a July essay, models a review body on the Financial Industry Regulatory Authority: an independent panel that would test advanced models before wide deployment. The second, pushed by former White House AI adviser David Sacks, takes its shape from the Motion Picture Association and its voluntary approach to rating films. Sacks has called the FINRA option “a DMV for AI.”
The FINRA comparison has a documented weakness worth stating plainly. Roughly 3,000 firms fund FINRA and place their own representatives on its board, yet approving new financial products before launch, the exact pre-deployment role an AI equivalent would have to fill, is barely part of what FINRA does. In other words, the institution being copied never actually performs the function it is being copied for.
Both live options share a structural feature: whichever one is adopted, the industry being reviewed helps fund or staff the reviewer. That is a different question from who called whom in August.
This is not an isolated case. Earlier this year the same pattern played out on a smaller scale: after industry pushback centered on competitiveness, a mandatory review of models, once floated at 90 days, got cut to a 30-day window, voluntary this time, with the formal evaluation authority giving way to a collaborative framework instead. One phone call from Sacks on the morning Trump’s first AI executive order was due for signature was, by reports, enough to push it back in May. Judged against that record, an August call raising objections to a regulator proposal is consistent with an established sequence rather than a new development in itself.
In an August essay Zuckerberg staked out his public position: even a single month’s delay to a model’s release, in his view, would “add significant risk to American leadership” against China. The tradeoff he describes hangs together logically, yet nothing backs it up: no analysis in public view puts a number on what a month’s delay would cost versus what a pre-deployment check would actually catch. That same month brought disclosures from both leading US labs that their models had reached systems never meant to be accessible, exactly the kind of failure a pre-deployment reviewer is supposed to exist to catch.
Public positioning from the major labs is thin. OpenAI and Google left Politico’s requests unanswered, and Anthropic would not comment on where it stands, even though the company’s co-founder Jack Clark had already written favorably in July about a FINRA-style body. Asked which structure the administration favors, a White House spokesperson did not say, offering only that officials are “committed to balancing innovation and security in AI policymaking.”
Here is the distinction that gets lost when every version on offer is voluntary: a voluntary commitment and a binding rule differ mainly in what happens when a company simply stops honoring it. No AI framework proposed so far, FINRA-style or MPA-style, has had to answer that question in practice, because none has yet been tested by a firm walking away from it.
The FINRA-style proposal is still alive despite the reported call. Operators should watch two mechanical details rather than the announcement of a winner: who funds the eventual body, and who sits on its board. Those two facts will tell you more about how much scrutiny frontier models actually face than any statement calling the process “balanced.”
The Next Web reported this story on September 4, 2026, citing Politico’s original reporting by Sophia Cai and Charles Rollet.
