Washington Is Handing AI Safety Back to the Industry

The White House is weighing whether to hand AI safety over to the people who build the models. The model on the table is the Motion Picture Association — the body Hollywood created a century ago to rate its own films — and adviser David Sacks has been explicit about why he likes it.

What the MPA did then was promote standards that then forestalled more intrusive, heavy-handed government action. — David Sacks

The body Sacks describes would set industry standards — and its purpose, as he said, is to forestall government from doing the same thing with real teeth. ^[1] The administration's record points the same way. In May, Trump cancelled a planned executive order that would have required developers to hand "covered frontier models" to Treasury, the NSA, and other agencies up to 90 days before public release for cybersecurity vetting — a pre-deployment gate. The lobbying came from Musk, Zuckerberg, and Sacks himself, and Trump's reasoning was plain.

It's causing tremendous good, and it's also bringing in a lot of jobs, tremendous numbers of jobs. — Donald Trump

He didn't want anything slowing the race with China. ^[2] The replacement arrived in June: a voluntary 30-day pre-release review, cut from 90 after more lobbying, that explicitly prohibits mandatory government licensing or permitting. The word "voluntary" is doing the work here. The administration's own substitute for the gate it killed is, by legal design, not containment. ^[3] That is the pattern the MPA proposal completes. But the containment model the administration stripped away is being rebuilt from below, and quickly. The labs are asking Congress for the mandatory rules they lost. OpenAI is now urging "mandatory, capability-based national regulation," arguing voluntary commitments aren't enough; Anthropic's Dario Amodei has endorsed third-party testing of frontier models before release. ^[4][5] A state has already gone further. Illinois passed the first law mandating annual independent third-party audits with access to unredacted models, 72-hour incident reporting, and penalties of one to three million dollars. Governor Pritzker framed it as a direct answer to federal inaction. ^[6]

Where the federal government has been unwilling to step up, states must venture once more unto the breach. — JB Pritzker

And Congress is split three ways — between leaving it to the industry, a federal oversight body with audit power, and Bernie Sanders's bill to ban superintelligence outright, with prison time attached. ^[7][8][9] None of this is settled. The administration hasn't formally chosen between an independent regulator modeled on Wall Street's FINRA and the MPA-style body, and Meta is on the other side of the fight from OpenAI and Anthropic, arguing against any national regulator at all. ^[1] Then the paradox. On the same day OpenAI asked Congress for mandatory safety rules, it released GPT-6 Astra — a model that crossed a "critical cybersecurity threshold," able to exploit unknown vulnerabilities without human direction. ^[7][4] The gates are being rebuilt by the very actors who keep blowing past them. The labs want a fence they will not build themselves; the states are building it anyway; and the administration is weighing whether to let the industry write the standards meant to hold it back.