The US executive order on AI

Executive Order 14110, signed on 30 October 2023, was at 111 pages the longest of Joe Biden's presidency, and the first binding American rule aimed at frontier models. Its most concrete provision used the Defense Production Act to require any company training a model with more than 10²⁶ floating-point operations, or acquiring a computing cluster large enough to do so, to tell the government and to share the results of its safety tests. Beyond that it directed agencies: standards for red-teaming from NIST, guidance on watermarking synthetic content, rules on AI in housing, hiring and healthcare, and a programme to bring in foreign talent.

The compute threshold, a number rather than a description of capability, was the order's lasting idea. The EU adopted 10²⁵ for its own definition of systemic risk, and California's vetoed SB 1047 borrowed the figure.

The order was revoked on 20 January 2025, the first day of the Trump administration, as a barrier to innovation, and replaced in July by an action plan that kept the export controls and dropped the safety requirements. The reporting threshold survived in practice because the companies had already built the process. The order is on this timeline as the high-water mark of the American government's attempt to regulate the models directly.