The October 2023 executive order on AI was the most comprehensive US federal AI policy action in living memory, which says more about the previous bar than the order itself. Tracking what actually happened through implementation is a more useful exercise than re-litigating the politics. The headline coverage overstates enforcement and understates the durable institutional changes, which is roughly the standing tradition for US tech policy coverage.
What moved: the NIST AI Safety Institute got funded and staffed, and the AI RMF generative-AI profile is a direct output. Federal agencies produced AI use inventories, which is the first time anyone systematically counted how federal agencies are actually using AI. Spoiler: more than they admitted, less than the worried op-eds claimed. Several agencies published procurement guidance that, for the first time, requires AI vendors to document model provenance, training data, and evaluation results. Boring on paper, load-bearing in practice.
What stalled: the reporting requirements for frontier model developers. The computation thresholds in the original order were set conservatively and follow-on rulemaking has been slow in the way that “voluntary commitments from major labs remain voluntary” usually means.
What changed at the transition: the Biden order was revoked early in the Trump administration and replaced with a narrower one prioritizing “AI dominance” framing over the original’s safety-first framing. The practical effects are smaller than the press release implied. NIST’s AI Safety Institute keeps its technical work. The voluntary commitments quietly faded from administration communications. The regulatory appetite for mandating frontier-lab safety evaluations dropped.
What’s still moving: state-level AI regulation (California, Texas, Colorado, in various states of coherence), the EU AI Act timeline, and international coordination through the G7 and OECD AI Policy Observatory. The action is in the states and Brussels. The federal level is a coordination problem.