Both remaining AI oversight plans differ on how much self-regulation to allow

For anyone writing next year's model-release checklist, the question that matters is whether a government office will need to clear a launch before it happens. None of the surviving proposals creates one. The two paths still on the table are a body funded by the firms it would review and a voluntary ratings scheme, and their backers are arguing over how much self-regulation should apply, with self-regulation itself no longer in question ^[2].

The heavier of the two is sold as pre-deployment review. Google DeepMind's Demis Hassabis popularised it in a July essay and briefed White House officials over the summer, proposing a standards body modelled on FINRA that would test advanced models for risk before wide release ^[6]. Look at the model being copied: FINRA is industry-led, funded by the more than 3,000 firms it oversees, seats member-firm representatives on its board, and reports to the SEC ^[7]. And as Politico notes, FINRA has little involvement in signing off new investment products, which is the exact pre-deployment function the AI body would exist to perform ^[8]. FINRA, in other words, is being borrowed for a job it doesn't do.

The lighter path drops the pretence. David Sacks, Trump's former AI and crypto czar, has proposed a body modelled on the Motion Picture Association's voluntary film ratings; he called the regulator idea a "DMV for AI" where models queue for a test, argued on his August podcast that the MPA's standards forestalled more intrusive government action, and said Elon Musk backs the approach ^[9][10].

For planning purposes, the direction of travel matters more than the current draft. A 90-day mandatory review has already become a 30-day voluntary window, with formal government evaluation authority swapped for a collaborative framework, after industry pushed back on competitiveness grounds ^[11]. That is a two-thirds cut to the review window ^[15], and it moved in one direction.

The mechanism doing the moving is a phone call. A senior White House official, granted anonymity, said Zuckerberg opposed the proposal ^[3], though another person familiar said he did not ask Trump to reverse himself, only that appointees reflect the president's light-touch approach, and that Trump called him first ^[4]. Meta declined to comment, and TNW, which relayed Politico's report, has not independently verified the call ^[5]. Sacks delayed Trump's first AI executive order in May with a single call on the morning of the signing ^[12]. Politico reports these one-on-one conversations have become an informal part of policymaking, able to slow or redirect proposals after officials have developed them ^[13]. Zuckerberg's public reasoning is on record: he wrote in August that a policy delaying a release even by a month would "add significant risk to American leadership" over China ^[14].

So the forcing question for a compliance plan is what the body can actually stop a team from shipping, whatever it's called. Sort the options that way: a reviewer with authority to withhold a launch sits in one box, a ratings sticker you can print and ship past sits in another. Every proposal that has survived contact with industry sits in the second. The workable plan is self-certification against a ratings queue you can clear on your own timeline, with the review file kept ready in case a body with teeth arrives later. And "we passed review" is worth reading closely, because on the FINRA precedent it need not mean a launch was ever cleared ^[8].