The White House and Congress relaunched their effort this week to stop the states from regulating artificial intelligence. The mechanism is federal preemption: a single national framework, deliberately permissive, that would override the laws individual states have passed or are passing to govern the technology. An executive order already creates an AI Litigation Task Force to challenge state statutes deemed onerous, and takes specific aim at Colorado’s AI Act, due to take effect at the end of this month — claiming, remarkably, that the law’s requirement to guard against algorithmic discrimination would “force AI models to produce false results.” Twice before, the preemption was tried inside larger bills and failed against bipartisan resistance. This is the renewed attempt to ensure that the rules governing the most powerful technology of the age are written once, in Washington, by the branch most accessible to the industry, and that the fifty separate attempts to write them elsewhere are extinguished.
One Rule, Written Once
The argument for preemption is not, on its surface, unreasonable, and it is worth stating at its strongest. A patchwork of fifty different state laws, each imposing its own definitions and obligations, would be genuinely difficult to comply with; a company operating nationally would face a thicket of conflicting requirements, and the cost of navigating it would fall hardest on the small and benefit the large incumbents who can afford the lawyers. A single, coherent federal standard is, in the abstract, the better instrument. The industry’s case for preemption rests on this reasonable foundation, and the reasonableness is what makes the maneuver effective.
But notice what is being proposed and what is not. The proposal is not to replace the patchwork with a strong uniform standard. It is to replace it with a permissive one — to preempt the state laws that constrain the technology and substitute a federal framework explicitly designed to be lighter than what the states were enacting. The tell is in the target. The order singles out Colorado’s law, whose offense is requiring AI systems to guard against discrimination, and characterizes that requirement as forcing the models to “produce false results.” When a guard against discrimination is reframed as a mandate for falsehood, the goal is not coherence. The goal is the removal of the specific constraints the states were imposing, dressed in the language of orderly national policy.
The structure of preemption is what makes it the industry’s preferred instrument, because it relocates the rule-making to the venue the industry can most efficiently influence. Fifty state legislatures are fifty separate battles, fought in fifty capitals, by lawmakers closer to their constituents and harder to lobby in bulk. A single federal framework is one battle, fought in one city, against a process the largest firms have spent years and fortunes learning to shape. To preempt the states is not merely to simplify the law. It is to move the writing of the law from the many rooms the industry must lobby individually to the one room it has already furnished, and to extinguish, in the process, the laboratories of policy that the states were designed to be.
The Captured Venue
This is the federal counterpart to the fiscal capture I described when a state government balanced its budget on the AI rally and quietly lost the will to regulate the industry funding it. There the capture was accidental, a byproduct of taking the easy revenue. Here it is deliberate, a designed effort to ensure that the level of government making the rules is the level most thoroughly captured. The states have been, so far, the source of the real constraints — the child-safety laws, the discrimination protections, the disclosure requirements — precisely because they are harder to capture all at once and closer to the citizens the technology affects. Preemption is the move to take the rule-making away from the many resistant venues and concentrate it in the single venue where resistance has already been managed.
The genius of the framing is that it weaponizes a genuine inconvenience to achieve a different end. The patchwork is real; compliance across fifty regimes is genuinely costly; and so the call for a uniform standard sounds like good governance rather than deregulation. But a uniform standard could be strong or weak, and the proposal is specifically for a weak one — the inconvenience of the patchwork invoked to justify not its replacement with something coherent but its elimination in favor of something permissive. The complaint about complexity is true. The solution proposed does not address the complexity so much as exploit it, using the public’s reasonable preference for simple rules to launder a private preference for lenient ones.
That the previous attempts failed against bipartisan resistance is the one hopeful note, and it is worth holding onto, because it shows the capture is not yet complete. Lawmakers of both parties, at both levels, balked at handing the industry a single weak national rule, which means the instinct to preserve the states’ power to constrain survives in enough of the legislature to have twice defeated the maneuver. But a thing attempted twice and relaunched a third time is a thing its proponents intend to keep attempting until it passes, and the resistance, to hold, must keep winning every round, while the industry needs to win only once. That asymmetry — defense must hold always, offense must succeed once — is the same asymmetry that governs every contest in this record, and it favors the patient party with the deeper resources, which is not the states.
What This Means
The preemption fight is the quiet, procedural front of the larger war over who governs this technology, and procedural fights are the ones the public is least equipped to follow and the industry most equipped to win. A debate about whether AI should be regulated draws attention and mobilizes opposition. A debate about whether AI should be regulated by the states or the federal government sounds like a technicality, a question of administrative tidiness, and it puts most citizens to sleep — which is exactly its value to the side that benefits from the answer. The most consequential decisions about the technology are being made not in the visible argument about its dangers but in the invisible argument about which venue gets to address them, and the invisible argument is being won where no one is watching.
And the same week reveals why the venue matters so much, in the form of the constraints the states were actually imposing. While Washington moves to preempt them, individual states and courts were doing the real regulatory work: mandating the labeling of synthetic performers, holding companies liable for their AI’s false statements, protecting workers and children from specific harms. These are the laws preemption would extinguish or override, not because they are incoherent but because they are effective, and effectiveness, against this industry, is precisely the quality a permissive federal framework is designed to neutralize.
So the rule that governs the most powerful technology of the age is being fought over not on the question of what the rule should say but on the prior question of who gets to say it, and the answer the industry seeks is: the one body it has already learned to move. I have watched powerful interests choose their regulator before, and the choice is always the same — the most distant venue from the affected, the most concentrated, the most accessible to capital, the one where a single victory settles the matter for the whole country at once. Preemption is that choice, made explicit. The states are the many doors the industry would have to open one by one, each with its own lock and its own constituents behind it. Washington is the one door, and the industry has a key. The fight is over whether the country keeps the many doors or surrenders to the one, and the side that wants the one has tried twice, lost twice, and just begun again.