Executive Order 14179: AI Safety Reversal & Action Plan

What the order actually revoked

Executive Order 14179 was not the administration's first attempt to void the 2023 order. On his first day back in office, January 20, President Trump signed Executive Order 14148, "Initial Rescissions of Harmful Executive Orders and Actions," which listed EO 14110 among 78 Biden-era actions revoked in a single sweep. EO 14179, three days later, named EO 14110 again, this time on its own and paired with an affirmative mandate rather than buried in a list of 78. [3][1]

The provision that had generated the most compliance work was Section 4.2 of EO 14110: developers training a "dual-use foundation model" above 10^26 floating-point or integer operations had to report training details, red-team results, and the physical and cybersecurity measures protecting model weights to the Commerce Department, using authority under the Defense Production Act. A companion NIST mandate to issue red-teaming guidelines and the standing of the U.S. AI Safety Institute, created under the same order, were both put in question the moment the order fell, though neither the Institute nor its underlying NIST work was actually shut down. [4][8]

EO 14179's own text framed the trade explicitly. Its Section 1 purpose language calls for AI development "free from ideological bias or engineered social agendas," and Section 2 states the new policy as sustaining "America's global AI dominance in order to promote human flourishing, economic competitiveness, and national security." Risk management, the organizing word of EO 14110, does not appear. [1]

The Safety Institute's second act

The infrastructure EO 14110 built did not disappear, it was repurposed. On June 4, 2025, Commerce Secretary Howard Lutnick announced that the AI Safety Institute would become the Center for AI Standards and Innovation (CAISI), keeping most of its model-evaluation and industry-liaison functions while dropping the word safety from its name. Lutnick's stated reason was that "censorship and regulations" had been used under the guise of national security, and that CAISI would instead evaluate "demonstrable" risks such as cybersecurity, biosecurity, and chemical weapons misuse. [6]

Atlas interpretation: So the reporting mandate lapsed, but the compute threshold that defined it did not. Eight months later, California's SB 53 adopted the identical 10^26 operations figure to define its own "frontier developer" category, this time as a state disclosure law rather than a federal one. The number EO 14179 stripped of federal teeth resurfaced as a state law's trigger point within the same year. [4]

The 180-day deadline

The order gave three officials inside the White House: the Assistant to the President for Science and Technology, the Special Advisor for AI and Crypto, and the Assistant to the President for National Security Affairs, 180 days to develop and submit an action plan for achieving the order's stated policy. That put the deadline at roughly July 22, 2025. The order's formal text, including that deadline, was published in the Federal Register on January 31, 2025. [1][2]

The White House published "Winning the Race: America's AI Action Plan" on July 23, 2025, a day past the informal deadline by most countings and within it by others depending on how the 180 days are counted from signature. It listed more than 90 federal policy actions across three pillars: accelerating innovation, building AI infrastructure, and leading in international diplomacy and security. Individual items ranged from expedited permitting for data centers and chip fabs to government procurement standards requiring vendors' models be "free from top-down ideological bias" and a directive for Commerce and State to package hardware, models, and standards together for export to allied governments. [5]

Atlas interpretation: Meeting the 180-day mark is notable given how often executive-order deadlines slip, but the deliverable itself was a list of actions to take, not the actions themselves. Most of the 90-plus items in the plan carried no deadline of their own, so the July 23 date measures when the administration finished deciding what it wanted to do, not when any of it took effect. [5]

Reactions at the time

Coverage in the weeks after EO 14179 was dominated by law firm client alerts and industry trade press rather than public statements from AI safety organizations. The dominant framing, including from legal commentary published within 48 hours of the first, broader rescission on January 20, was uncertainty: what would replace the reporting requirement, and whether the AI Safety Institute would survive at all, rather than direct objection to the revocation itself. OpenAI and Google had both filed public comments in the weeks before pushing for exactly this kind of deregulatory shift, including relief from what they characterized as burdensome copyright and transparency obligations. [8]

Atlas interpretation: The sharper public criticism landed later and elsewhere: at the CAISI rename in June, when commentators read the dropped word "safety" as a deliberate signal, and at the Action Plan and December preemption order, when the deregulatory direction became concrete enough to fight over. A memo revoking a framework is a smaller target than the machinery built on top of it months later, and that is roughly how the response was staged. [6][5]

What it set up

The clearest line from EO 14179 runs to the December 11, 2025 preemption order, which cites EO 14179 by name and describes it as having "revoked" the prior administration's approach and directed the removal of barriers to U.S. AI leadership, framing the new order as continuing that same project against state law rather than federal law this time. [7]

Atlas interpretation: Read together, the three documents trace one continuous argument rather than three separate decisions: revoke the federal risk-reporting regime in January, spend 180 days turning "remove barriers" into a 90-item list in July, then in December turn the same language against the state laws, including SB 53's frontier-developer disclosure rules, that had filled the gap the January revocation left. The target moved from a federal executive order to state statutes, but the stated policy, a single national floor rather than fifty different ones, stayed the same across all three. [7][1]