Trump’s AI executive order was inevitable
Sufficiently capable models force national security responses — turning even the most ardent opponents of regulation into begrudging regulators
On the very first day of his second term, President Donald Trump revoked Joe Biden’s AI executive order, which required AI companies to disclose details of their internal safety testing.
A month later, Vice President JD Vance railed against the notion of “AI safety” at a summit in Paris. “The AI future is not going to be won by hand-wringing about safety,” he said.
Sixteen months into the presidency, the White House is acting very differently. After a chaotic few weeks, on Tuesday President Trump signed an executive order on AI — and with it, crushed the dreams of regulation opponents.
The new EO establishes a voluntary pre-deployment evaluations regime to tackle catastrophic cyber risk to America’s “vital functions.” When companies develop new frontier models, they’ll share them with the government for testing. If the model meets a certain (classified) threshold for cyber capabilities, the government will have exclusive access to the model for up to 30 days — the intention seemingly being that the government can use its head start to secure critical infrastructure before attackers get similar capabilities.
This was inevitable. For years, the AI safety argument for government regulation has rested on a simple principle. Models will become powerful enough to pose national-security-relevant risks, and we will need regulation to deal with them.
The most vocal opponents of regulation, for all their talk of “acceleration,” denied that such risks would materialize. They thought models would never get this powerful, and so no special regulation would be needed. The idea of frontier AI regulation was complete anathema to this group — “a form of murder” which, Marc Andreessen later argued, would “impose tyranny far beyond anything even imagined by the Communists and Fascists of the 20th Century.”
For anyone expecting continued and rapid improvement in AI, that position was untenable. The release of Mythos — a model which very clearly does present national security-relevant risks — made the entire argument collapse. That is why, despite David Sacks’ best efforts, Trump is finally taking action. His executive order shows what AI safety advocates have argued all along: you can’t not regulate AI. Sufficiently capable models force national security responses, turning even the most ardent opponents of regulation into begrudging regulators.
This is not the end of the debate. As the tortured run up to yesterday’s signing underlined, Trump is capricious; he could easily change his mind once again. And the devil, as always, is in the details. The executive order only establishes a voluntary framework (though companies will face immense pressure to comply), and it does not lay out what will happen in the event models are found to be unacceptably risky. It also only applies to cyber capabilities; a similar regime to tackle biorisk will no doubt be needed soon. Much is left to be decided, and there are plenty of ways that what comes next will be either insufficient or actively counterproductive (including, as accelerationists have long warned, by being too strict).
But it is still a watershed moment. The nature of Trump’s cult of personality is that others must now fall in line. Senator Ted Cruz, long dismissive of AI risks and the need for regulation, is now urging Congress to “address catastrophic risk.” David Sacks, no doubt speaking through gritted teeth, is trying to spin this as a win for him and his faction. (It isn’t.) And the pro-industry super PAC has begun to endorse regulation, seemingly accepting its inevitability.
The question of whether the government should regulate the most powerful AI models is, simply put, no longer a question. The accelerationists got the most sympathetic administration imaginable, and a direct line to POTUS. But what the models could do mattered more than anything they could.
The poll-watcher in chief is looking ahead to November and it isn't a good sight. I wonder how shapes AI's pr shaped problem.
Eleven days between the killed version and the signed version. 90 days became 30. Safety became cybersecurity. The text gained a clause explicitly barring any mandatory pre-release requirement. The actual language reads more like a guarantee against preflight checks than an implementation of them.