Musk and Zuckerberg convince Trump to kill AI safety review

On the afternoon of May 21, 2026, the White House was set to host a signing ceremony for a new executive order on AI safety. Major AI and cybersecurity company executives had been invited; some based in California were already traveling to Washington when aides began calling to cancel. Trump told reporters in the Oval Office that he postponed the order because he didn't like certain aspects of it, adding that the U.S. is 'leading China, leading everybody' on AI and he didn't want to do anything to 'get in the way of that lead.'

The official explanation contradicted what was happening behind the scenes. Between Wednesday night, May 20, and Thursday morning, May 21, Elon Musk, Mark Zuckerberg, and former White House AI adviser David Sacks all called Trump directly to argue against signing the order. Musk and others also worked to bring officials at the National Economic Council on board with opposition to the measure.

The executive order had two components. The first, on cybersecurity, would have expanded the government's capacity to defend critical infrastructure (hospitals, banks, power grids) against AI-enabled cyberattacks and mandated a hiring surge at the Cybersecurity and Infrastructure Security Agency. The second, on frontier AI models, would have created a voluntary 90-day pre-release review window: labs could choose to submit their most powerful AI systems to federal agencies before public release, giving the government a first look at potential security vulnerabilities.

The frontier model component did not create a licensing regime, mandatory delays, or pass/fail gates. Labs retained control over whether to participate. The NSA would have been responsible for classified testing of submitted models, and CISA, the Office of the National Cyber Director, and the White House Office of Science and Technology Policy would have jointly determined which systems qualified as 'covered frontier models.'

Elon Musk is the founder and CEO of xAI, which develops the Grok series of large language models. Grok is a direct competitor to OpenAI's GPT series, Google's Gemini, and Anthropic's Claude. In early 2026, xAI faced significant internal turbulence: nine of the company's 11 co-founders had departed by February, and the company was under pressure to ship a competitive next-generation model. A voluntary review process that applied to Grok would have created a pre-release scrutiny window for xAI's core product.

Mark Zuckerberg is the founder and CEO of Meta, which develops the open-source Llama series of frontier models. Meta had recently debuted a major new AI model in April 2026 after a $14 billion investment to bring in Alexandr Wang. Any voluntary review framework that included Llama as a covered frontier model would have put Meta's open-source release strategy under federal scrutiny before those models shipped.

David Sacks served as Trump's AI and crypto czar from January 2026 until March 2026, when he hit the 130-day limit for special government employees and transitioned to co-chair the President's Council of Advisers on Science and Technology (PCAST). As a special government employee, Sacks received two separate ethics waivers: one covering his crypto holdings and a second covering his AI holdings, permitting him to shape federal AI policy despite retaining financial stakes in the industries he oversaw.

An analysis of Sacks's financial disclosures found that among his roughly 708 tech investments, 449 are AI companies that stood to benefit from the permissive regulatory posture he advocated. Sacks divested from Amazon, Meta, and xAI directly, but his venture firm Craft Ventures maintained more than 400 tech investments with AI ties. Government ethics expert Kathleen Clark called the waivers 'sham ethics waivers' that function 'like a presidential pardon in advance' for conduct that would ordinarily violate criminal conflict-of-interest law. Democrats launched a formal ethics investigation into Sacks in September 2025.

Sacks argued to Trump that a voluntary pre-release review would function as a de facto licensing regime. He told Trump he feared the voluntary procedure would slow AI companies' model releases and that a future administration could easily convert it into a mandatory one. The 'accelerationist' framing, that any friction in AI development costs the United States its competitive advantage over China, was the rhetorical lever Sacks used to bring National Economic Council officials on board with the opposition push.

The argument was effective even though the order contained no mandatory requirements, no fines, no pass/fail tests, and no legal obligation to participate. Critics pointed out that 'voluntary' by definition means labs could ignore the review entirely, and that Musk and Zuckerberg fought it anyway.

Not all major AI companies opposed the order. OpenAI, which develops GPT-series models and is one of xAI's direct competitors, reportedly supported the executive order, according to people familiar with the matter. Anthropic, which develops Claude, had been negotiating with the administration and was in discussions about the order. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman both indicated they couldn't attend the signing ceremony, a logistical wrinkle that contributed to last-minute disorder, though the primary cause was the lobbying campaign.

OpenAI and Anthropic had already been cooperating with government AI safety reviews. xAI and Meta had not. The companies that lobbied to kill the order are the companies least subject to existing oversight arrangements.

Both Musk and Meta disputed published accounts of their role. Musk posted on X: 'This is false. I still don't know what was in that executive order and the President only spoke to me after declining to sign.' Meta said Zuckerberg spoke with Trump only after the order was rescinded, not before. Semafor, the Washington Post, and Fortune all reported, based on people familiar with the matter, that conversations between Musk, Zuckerberg, and Sacks with Trump took place between Wednesday night and Thursday morning, prior to the postponement announcement.

Both companies benefit from the narrative that billionaires don't directly intervene to shape federal policy that affects their products. The conflicting accounts have not been independently resolved, but the weight of contemporaneous reporting from multiple outlets places the conversations before the cancellation.

The European Union's AI Act entered force in August 2024 and is fully applicable by August 2026, requiring frontier model developers to register systems, submit to conformity assessments, and maintain post-market monitoring. The United States, by contrast, now has no federal pre-deployment oversight mechanism for frontier AI. A December 2025 Trump executive order directed the Justice Department to challenge conflicting state AI regulations, further narrowing the regulatory patchwork.

The Center for AI Safety expressed disappointment at the postponement. Without a federal first-look mechanism, the government's ability to assess whether the most powerful AI systems can assist with cyberattacks, bioweapon design, or other dangerous capabilities before those systems ship is limited to the goodwill of the companies building them.

Trump's stated rationale shifted across the day of May 21. His initial public comment — that he didn't like certain aspects of the order — gave way to a broader framing: that regulation of any kind could cost the United States its lead over China. That second framing matches the argument Sacks and Musk made in their calls. The Washington Post reported that one of the main reasons for the delay was simply that Trump 'just hates regulation.'

Axios published the full text of the stalled executive order on May 22, 2026, allowing the public to read the measure that was pulled. The order's text confirmed there was no mandatory review requirement, no licensing gate, and no penalty for non-participation, making the industry's last-minute lobbying campaign against it a fight to prevent even voluntary government access to frontier AI systems.