Federal preemption was designed to prevent fifty different AI rules. By May 3, 2026, it had produced something stranger: a regulatory environment with no federal floor and dozens of subnational ceilings, where the binding constraints on what an AI system can legally do were being set by a Maryland grocery law and a labor court in Beijing.

Two governments acted on the same Sunday. Maryland became the first US state to ban surveillance pricing in grocery stores — the practice of using a shopper's data to charge them more than the customer in front of them. A Chinese court issued the country's second ruling in six months that companies cannot terminate workers solely to replace them with AI. Different political systems. Different economic systems. Opposite ideological premises. Identical structural position: each government decided not to wait for the national legislature.

This is not what the architecture was built for.

What the system was designed to do

In November 2025, the AI industry asked President Trump for a single national framework. They got it on December 11, when Trump signed an executive order aimed at preempting state AI laws. The phrasing in the order was unambiguous: "one minimally burdensome national standard," not "50 discordant state ones." David Sacks said the EO would focus on "excessive and onerous" state laws. The Justice Department was directed to create an AI Litigation Task Force to challenge state laws that conflicted with the federal approach.

The premise was the standard preemption argument: a single national rule prevents a patchwork. Without it, an AI company has to comply with the strictest state's rules everywhere, because models are deployed nationally and the floor becomes the ceiling. With it, innovation routes through the federal level and the states are silenced.

Two weeks later, New York's governor signed the RAISE Act anyway — the strongest state AI safety law in the country, requiring 72-hour incident disclosure and fines of up to $3 million. Policymakers behind other state bills scoffed at the EO. Some Republican governors said they were going ahead. The Axios headline that week read "States Defiant."

By April 4, 2026, the White House's effort to pass preemption legislation through Congress had stalled. The proposal had become too obviously partisan; Democrats wouldn't carry it; Republican senators with state AI bills in their own legislatures wouldn't either. The EO remained on paper. The states kept moving.

What the system actually did

Walk forward five months from the EO and the inventory of substantive AI rules in force is almost entirely state and foreign. California SB 53 required AI safety testing disclosure starting September 2025. New York's RAISE Act compelled incident reporting. Illinois banned AI as the sole provider of mental health therapy. Colorado passed a landmark anti-discrimination law strict enough that xAI filed a free-speech challenge against it in April. The Maryland law signed in May 2026 was the first explicit ban on algorithmic price discrimination in groceries, with bills pending in Colorado, California, Massachusetts, Illinois, and New Jersey.

Federal substantive AI rules passed in the same span: zero. The single national standard exists only as the absence of one.

May 2026
Maryland becomes the first US state to ban surveillance pricing in grocery stores, as other states including CO, CA, MA, IL, and NJ consider similar bills
The Guardian

The federal AI framework was designed to silence the states. The states became the framework.

The China side ran on a parallel track. A Chinese court ruled in December 2025 that companies could not fire workers solely to replace them with AI. On May 3, a second Chinese court issued the same ruling in a different case — a pattern, not an outlier. Bloomberg had already reported in March that Beijing was facing a structural dilemma between AI productivity gains and labor displacement, and that "dark factories" requiring no human workers were upending the country's labor market. The national government did not legislate. The courts did. Each ruling binds only the parties to that case, but the precedent is now established in two jurisdictions, and Chinese employment law tends to converge faster than US state law because the court system is centrally administered. The constraint exists. It just doesn't come from the legislative organ that would normally produce it.

The phases

Each phase felt like the resolution. The EO was supposed to end the state push. The state push continued. The legislative framework was supposed to harden the EO. The framework stalled. The xAI lawsuit was supposed to invalidate Colorado's law. The case is still being argued — and in the meantime, two more state actions arrived.

The same Sunday, the case for the rules

The day Maryland banned algorithmic discrimination in grocery stores, the most credible empirical defense of AI deployment in years also landed. A Harvard study found OpenAI's o1 correctly diagnosed 67% of emergency room patients from electronic records and a few notes from nurses, against 50–55% for triage physicians. This is the case for letting the system run: it is, on the evidence presented, better than the people it replaces at one of the highest-stakes decisions in medicine.

The same Sunday, the Wall Street Journal reported that OpenAI employees had raised internal alarms about ChatGPT's failure to alert law enforcement when users described plans for real-world violence to the chatbot. The company is now facing wrongful-death suits tied to a mass shooting in Canada. This is the case against letting the system run unsupervised: even the people who built it disagreed with how it was deployed.

Neither finding was anticipated by the federal framework. The EO did not contemplate clinical superiority. It did not contemplate liability for chatbot output. Maryland's surveillance-pricing law happens to apply to AI systems that personalize prices, but it was drafted around grocery stores, not models. The Chinese ruling happens to constrain AI-driven layoffs, but it was issued under existing labor law, not new AI legislation. The rules binding AI deployment in May 2026 were almost all written for something else, by jurisdictions the EO was supposed to silence.

Why the floor became the ceiling

Preemption only works when the federal layer produces a binding rule that displaces the state layer. The Trump EO produced no rule — only a prohibition on state rules. Without a federal floor, there was nothing to preempt to.

For the AI companies, the structure is now worse than what they asked for. They wanted one rule. They got the absence of one rule plus the active resistance of states that interpreted the EO as a federal abdication. Industry counsel now has to track state legislation in California, New York, Colorado, Illinois, Maryland, and a half-dozen others — and increasingly, foreign labor courts. The cost of compliance is higher than under a single rule. The certainty is lower. The reputational floor is whatever the strictest state decides this quarter.

The federal courts are the one path back. If the xAI Colorado lawsuit succeeds on First Amendment grounds, it could invalidate a significant slice of state AI law, and other suits will follow. But the litigation timeline is years. The state laws are in force now, and an AI company facing a Colorado fine or a Maryland enforcement action cannot defer compliance until the case resolves. The economics of the wait are simple: comply with the strictest state rule, or stop operating in that state. Almost no one stops operating.

The reversal

The Arm precedent is the closest structural analogue. Arm spent 35 years designing chips for everyone and competing with no one — a neutrality protected by regulators who blocked Nvidia's acquisition specifically to preserve it. The neutrality survived the external threat. It did not survive Arm's own incentive to capture the AI market. The system designed to prevent the threat from outside was undone by a force the regulators had not modeled: a change in what the regulated entity wanted to be.

Federal AI preemption was designed to prevent fragmentation from outside the federal layer — to stop the states from imposing fifty rules. It did not survive a force the architects had not modeled: that the federal layer, having declared itself the source of authority, would then produce nothing for the states to be preempted to. The states moved into the vacuum because there was a vacuum.

The dynamic plays out in another system this week. The Chinese court ruled the same way it had in December — not because anyone harmonized the doctrine, but because the structural condition (worker displacement, social stability concerns, no national AI labor law) pressed two unconnected courts toward the same answer. The same forces that produced one ruling produced the second. Both governments built systems intended to handle AI from the top down. Both ended up regulating it from below.

The crystallization

In December, the executive order declared that the United States must have one minimally burdensome national standard for AI, not fifty discordant state ones. On the first weekend of May, a Maryland grocery law and a Chinese labor court ruling were doing more to shape what AI companies could legally do than anything signed in Washington or passed in Beijing in the five months between them.

The framework designed to silence the states became a vacuum the states filled. The systems that were supposed to be governed from the top are being governed from below — case by case, statehouse by statehouse, court by court. The federal layer is still where the press conferences happen. The floor is somewhere else.