The United States has no federal artificial-intelligence statute. What it has instead is a contest over who is entitled to write one: state legislatures that have gone ahead without Congress, and a federal executive branch that has set about contesting them in court without a law of its own to point to. No court has yet held that any state AI law is preempted, and the instruments doing most of the work on the federal side are executive orders.

The states moved first, because they could. New York City began enforcing the first AI hiring-bias audit rule in July 2023. Colorado passed the first comprehensive state AI statute in May 2024, over its own governor's written reservations that it created a complex compliance regime and should be replaced by a federal one. California's governor vetoed a frontier-safety bill that September while signing a narrower disclosure law, AB 2013, requiring developers to publish summaries of the data their models were trained on. Texas followed with the Responsible AI Governance Act in June 2025, California with the frontier-AI transparency statute SB 53 that September, and Illinois with a safety-and-audit law in July 2026. Washington, meanwhile, kept changing direction. Executive Order 14110 of October 2023 imposed reporting duties on frontier developers; it was revoked on the first day of the next administration, replaced within three days by an order directing an "AI dominance" agenda, and elaborated in July 2025 into an action plan. None of it was legislation.

The confrontation began in December 2025, when an executive order directed federal agencies to contest state AI laws. Its first practical use came in April 2026 in Colorado: xAI sued the state attorney general to enjoin the Colorado AI Act, arguing that designing a model is protected speech and that the law's carve-out for algorithms advancing diversity violated equal protection; on 24 April the Justice Department moved to intervene against the statute, its first such intervention anywhere; and on 27 April the court suspended the law before it had ever applied, expressing no view on the merits. Colorado did not wait for the ruling. On 14 May Governor Jared Polis signed a bill repealing and re-enacting the Act as a narrower regime covering automated decision-making, dropping the mandatory risk-management programmes and impact assessments he had objected to two years earlier and deferring the obligations to 2027; a fortnight later he signed a second AI statute governing chatbots, and in August his attorney general filed the first draft rules ever written under a comprehensive state AI law. In July xAI took its speech argument to Minnesota, suing to stop a first-in-the-nation ban on nudification tools days before it applied. In parallel the federal executive has been building a gate of its own: an order of 2 June 2026 creates a review mechanism for the most capable models, voluntary by its own terms and triggered by a cyber-capability benchmark that is classified and that the director of the National Security Agency alone applies.

The developer's constitutional argument has not yet won in court. xAI has brought three of these suits, and the one that has run furthest went against it: on 4 March 2026 a federal judge in Los Angeles refused to block California's training-data disclosure law, which has applied ever since, and the Ninth Circuit has had the appeal since that month. A Minnesota judge has twice declined to freeze that state's nudification ban — in July before it took effect, and again on 4 September — each time on procedural grounds, without reaching the speech question, which now goes to the state's motion to dismiss. So the pattern that held through the spring — a private developer freezing state AI laws, with the United States joining in support — no longer describes the record, and the federal government has still not asked any court to hold a state law preempted on its own account. What remains unresolved is whether an appeals court will accept that disclosure duties are compelled speech; what the state regimes cost, since most do not bite until 2027; and how a developer is meant to know whether the federal gate applies to it, when the threshold rests on a benchmark nobody outside an intelligence agency can inspect. Polis's call for Congress to settle the matter with a preemptive federal statute, made when he signed the first state law in 2024, is still unmet.