“I didn’t like certain aspects of it,” Trump explained to reporters at the White House on Thursday morning. “I think it gets in the way of—we’re leading China. We’re leading everybody, and I don’t want to do anything that’s going to get in the way of that.”
The order would have created a system in which AI companies could voluntarily submit their most advanced models to key national security agencies for testing and vetting up to 90 days prior to releasing them. Officials from multiple government departments and agencies had spent weeks negotiating over the executive order’s language, and leading AI companies had been briefed on its content. At least two of those companies, Anthropic and OpenAI, had indicated they were in favor of the voluntary vetting system.
The executive order was under consideration following the debut of Anthropic’s Mythos model, which possesses unprecedented cyber capabilities. The AI company has voluntarily limited Mythos’ release out of concern that those capabilities, if widely shared, could help hackers to launch devastating cyber attacks against critical infrastructure.
But David Sacks, the Silicon Valley venture capitalist who stepped down in late March as Trump’s AI and crypto czar, successfully mounted a last-ditch lobbying effort to derail the order’s signing. Sacks called Trump on Thursday to express his concerns, according to press accounts. The campaign also included similar calls to Trump from Elon Musk and Mark Zuckerberg, both of whom are developing advanced AI models. There were also, reportedly, efforts to convince members of Vice President JD Vance’s staff to voice concerns about the order with Trump.
Sacks is a prominent AI “accelerationist” who believes that any federal regulation will harm U.S innovation, hurt the business interests of U.S. technology companies, and delay the country from experiencing the many benefits he believes AI will bring. He also sees the U.S. as being in a potentially existential race with China to develop advanced AI capabilities, and believes that regulation will result in the U.S. falling behind in this geopolitical contest.
Although no longer officially serving as Trump’s AI advisor, Sacks continues to wield influence on the administration’s AI policy. Earlier this week, according to news reports, he attended official briefings on the executive order. At the time, he reportedly indicated he would not oppose the voluntary model testing framework.
But, according to a story in Politico, Sacks later told Trump that he feared the voluntary vetting would act as a de facto licensing regime, slowing down AI companies’ releases of new AI models. He also worried, Politico reported, that a future administration might easily turn the voluntary procedure into a mandatory one.
Trump’s decision to postpone signing the order leaves U.S. AI policy in a strange place both in terms of policy and politics. AI regulation is an issue which splits Trump’s base. Trump came into office supported by a cadre of “move fast and break things” Silicon Valley billionaires, including Sacks and Musk. They have expressed admiration for the president for tearing down what they see as unnecessary red tape and bureaucracy, and for embracing crypto currency. Zuckerberg, who was a critic of Trump during his first term as president, has in the past year emerged as a vocal supporter.
Prior to Trump’s decision to pull out of signing the executive order, many in Washington thought the order was a done deal and that the forces opposing the approach were in retreat. Poll after poll indicates that the majority of Americans—including a majority of Republicans—are fearful about AI’s impact on jobs and its potential negative impacts on education and children’s mental health. Many oppose the construction of data centers near them. And many religious Christians are deeply suspicious of AI, viewing the technology as a kind of “false god,” equivalent to idolatry.
Former Trump advisor Steve Bannon, for instance, was among more than 60 MAGA loyalists who earlier this week signed an open letter to Trump urging him to test and approve powerful AI models before they are released. The letter was organized by Humans First, a conservative group whose tagline is “technology should serve humans…not replace them.”
A poll of Republican voters released today by the Future of Life Institute, an AI safety group concerned with AI’s potential existential risk to humanity, found that 79% were in favor of the government testing AI models before they are released to ensure they are safe, and that 87% favored the government having the power to block the release of AI models that pose a national security threat.
“Our image and the stereotype is that Republicans are against regulation,” Michael Kleinman, head of U.S. policy for the Future of Life Institute, told Fortune. “But what we are finding instead is when people see a technology that has direct and often incredibly negative impact on their lives, their kids, and their communities, they want the government to step in and put in place common sense guardrails.”
Kleinman said that while Sacks and the Silicon Valley faction of Trump’s base have prevailed for now, public opinion and the trends in AI development were against them. “Public opinion is solidifying on both the left and the right,” he said. “Mythos won’t be the last model that these companies release that will pose significant national security threats—it is the first such model. So the pressure is only going to continue to build for the government to take common sense action.”
The Trump administration is, for the moment, continuing to exercise a kind of ad hoc licensing process for just that one AI model, Anthropic’s Mythos. Anthropic has shared the model with the U.S. government and, under what Anthropic calls Project Glasswing, with a select handful of U.S. technology companies and financial institutions who make software that underpins much of the internet and other critical infrastructure. But the White House, according to press reports, blocked Anthropic from expanding the number of companies with access to Mythos due to national security concerns.
Meanwhile, OpenAI has created an AI model, GPT-5.5, that—according to OpenAI’s own testing and that carried out by the U.K. government’s AI Security Institute—is almost as capable as Mythos. OpenAI has released the model only to partners in a “trusted access” program, although the program is less limited than Anthropic’s Glasswing. The U.S. government has not applied the same scrutiny to the expansion of OpenAI’s trusted access program for GPT-5.5 as it has to Glasswing.
A number of AI companies also voluntarily share their most advanced AI models with the Center for AI Standards and Innovation (CAISI), which evaluates them for some potential risks. But CAISI is part of the Department of Commerce and its experts do not necessarily have access to classified information to help them assess the risks AI models pose, or the expertise in advanced cyber security methods that parts of the U.S. national security establishment have.
The discussion over the now-postponed executive order, according to a story in The Washington Post, involved heated wrangling between different branches of the U.S. government over who should be in charge of testing and approving AI models. The Commerce Department wanted to hold on to its leadership on model evaluation with the CAISI, but the U.S. intelligence community was also vying for responsibility. Meanwhile, it is Treasury Secretary Scott Bessent who has been leading much of the administration’s response to Mythos so far.
There were also disputes between factions of Trump advisors. Kevin Hassett, the director of the National Economic Council, suggested last week that the administration would set up a licensing system similar to how the Food and Drug Administration reviews testing of drugs before approving them for sale. That brought a rebuke on social media from White House Chief of Staff Susie Wiles, who said the president was not “in the business of picking winners and losers.”
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。