For months now, the White House has hinted that it may try to rein in the AI industry. Just two weeks ago, the nation’s top tech executives—including Sam Altman and Dario Amodei—were invited to attend a ceremony for the signing of a long-anticipated executive order on AI. But just hours before the ceremony, Donald Trump scrapped it. America is leading the world in the AI race, the president told reporters at the time, “and I don’t want to do anything that’s going to get in the way of that lead.”

Apparently, Trump has changed his mind again. Earlier today, the president signed an executive order that will create a process for top AI companies to voluntarily share certain upcoming models with the government for safety testing up to one month before wider release. OpenAI, Anthropic, and the like will also be asked to work with the government to shore up federal, state, and local cyberdefenses. The White House spokesperson Liz Huston told us that the policy reflects a “common-sense approach of collaborating with industry to balance innovation and security.”

The order itself is relatively toothless: Even before today, the major AI firms already had agreements in place that allowed the government to preemptively test their models for safety risks. The new rule “effectively formalizes what has already been happening between the US government and the leading AI companies,” Daniel Remler, an AI expert at the Center for a New American Security, told us.

But the executive order is meaningful in that the president is doing something—anything—about AI. At the start of his second term, Trump signaled to tech companies that he would stay out of the way. Last January, he rescinded a set of modest Joe Biden–era policies, calling the rules “dangerous” and a “barrier” to American AI leadership. Even the preamble of today’s executive order celebrates that Trump “unleashed tremendous technological growth” by “slashing the bureaucratic constraints that the prior administration placed on America’s AI developers.” Yet core components of those supposedly dangerous Biden-era AI regulations—voluntary agreements to share information about advanced AI models with federal agencies, for instance, as well as federal programs to leverage AI for cyberdefense—are strikingly similar to today’s new AI executive order. Dean Ball, a former AI adviser to the Trump administration, wrote that the policy “is considerably more intrusive” than Biden’s executive order.

Today’s order still could have been much more forceful. When the White House first started previewing the possibility of regulatory action in May, one administration official suggested that AI models would be reviewed “just like an FDA drug.” Even the leaked draft text of the version that Trump had originally planned to sign last month would have been more burdensome for tech companies. After David Sacks, the White House’s former AI czar, reportedly called the president to complain, Trump canceled the signing ceremony. Today, after the new order was announced, Sacks declared the watered-down provisions a “game changer” on X—despite the fact that the new government-review process is not so different from what he had originally opposed. This means that two former libertarian AI advisers to the White House—Ball and Sacks—disagree about whether this order is a good thing.

At the same time, joining Sacks in praising the rule is Steve Bannon, Trump’s former chief strategist and a leading critic of AI on the right. “It’s not perfect,” he told us. “But directionally, it is pretty damn good.” As Bannon sees it, despite the fact the order is weaker than earlier versions, codifying rules is a step in the right direction.

The entire, chaotic saga—a wishy-washy White House, confused statements from populist and tech-elite Trump whisperers—is only the latest in a long string of strange, often contradictory AI-policy positions. Trump’s approach to AI has been inconsistent, if not incoherent, almost since the day he retook office. Consider that, for all the talk of cybersecurity, this administration has also gutted the Cybersecurity and Infrastructure Security Agency, the government agency that aims to protect the nation against hackers. CISA also happens to be one of the main federal agencies tasked with implementing today’s executive order.

Or take the White House’s relationship with Anthropic. On the one hand, Anthropic likely triggered the executive order in the first place. In April, the company announced Claude Mythos Preview, a new model with advanced hacking capabilities that has ignited concern over the growing power of AI companies. Ever since, the president has seemed to cozy up to Anthropic. Dario Amodei, the firm’s CEO, visited the White House that same month for conversations over the future of the government’s relationship with the company. “I like high-IQ people, and they definitely have high IQs,” Trump later told reporters of Anthropic’s leadership.

On the other hand, the Trump administration appears to be fighting in court to bar Anthropic from doing most national-security work. In February, the Pentagon designated Anthropic a “supply chain risk” after a high-profile contract dispute over the use of AI in warfare, essentially declaring it a national-security risk for the military to even touch Anthropic products. In late April, when Anthropic tried to grant Mythos access to more companies for cyberdefense—very in line with today’s executive order—the White House appears to have, inexplicably, blocked the move. (An Anthropic spokesperson pointed us to a post on X in which the company called today’s executive order “an important step in strengthening America’s leadership in AI.”)

Then there’s the administration’s attitudes toward China. Trump has repeatedly emphasized the need to deregulate the AI industry in order to stay ahead of China. Meanwhile, he has also permitted Nvidia to sell some of its most advanced AI chips to Chinese companies, lifting an export control the Biden administration put in place precisely to waylay Chinese AI development. (Anthropic, by the way, denied a Chinese think tank access to Mythos.) Trump has, in the name of beating China, pushed to remove regulatory constraints on data-center construction: “Build, baby, build,” he said last July. But once uproar emerged about data centers hiking up electricity bills, the White House announced a voluntary pledge for AI companies to take a number of measures that would prevent everyday people from paying for data-center electricity. Build, baby, but prudently.

Indeed, at least some of the vacillations seem to be driven by public opinion. Over the past several months, as AI models have improved, attitudes toward the technology have soured. Today’s order allows the administration to look as if it is undertaking more robust AI regulation—but it doesn’t actually require the industry to do very much, if anything. Trump is trying to score points with both the public and Silicon Valley. But in doing so, he’s not saying or doing anything substantive at all.

AI spending is consuming the U.S. economy, people are afraid of losing their jobs to AI, and communities across the nation are gathering to protest data centers. Political figures as divergent as Bannon and Bernie Sanders are expressing concern over AI and the concentration of power among the industry’s executives. This would seem to be a clarion call for the president of the United States, and a populist one at that. Instead, the White House spent weeks prevaricating on an executive order that rests on the voluntary cooperation of the AI industry. With Anthropic, OpenAI, and their competitors becoming major economic and geopolitical powers, the window for any one government to seriously regulate AI is rapidly closing. Hopefully, it is not already gone.