Post-Chevron AI Governance: Why Persuasion Beats Deference in Regulatory Preemption

The Supreme Court's 2024 decision in Loper Light Enterprises v. Raimondo overturned a forty-year-old doctrine that had quietly reshaped how America regulates emerging technologies. Under the old Chevron framework, federal agencies could interpret ambiguous laws as they saw fit, and courts had to accept those interpretations so long as they seemed reasonable. For artificial intelligence governance, this created a particular problem: a federal agency could write AI regulations and then declare those same regulations to be the ceiling—the maximum protection allowed—effectively wiping out state-level lawsuits that might otherwise hold companies accountable.

The tension runs deeper than simple federalism. When Chevron was in force, agencies had little incentive to explain their reasoning in detail, especially when their interpretations would preempt state tort law. Why labor over a thorough justification if courts would defer anyway? This mattered enormously for AI, where information asymmetry is severe. Federal regulators often know far less about systemic risks than the companies they oversee. State-level litigation—the messy, expensive process of people suing companies in court—generates real-world feedback about what goes wrong. It produces ground-truth signals. Chevron deference cut off those signals by allowing federal agencies to declare state lawsuits preempted without having to prove their regulations were actually adequate.

When Loper Light overruled Chevron in 2024, many legal scholars celebrated. The decision seemed to restore judicial independence and prevent unelected bureaucrats from interpreting the limits of their own power. But the reality is more complicated. Legal scholars Cary Coglianese and David B. Froomkin have argued the ruling is essentially disingenuous—it does not represent a meaningful break from Chevron's logic. Courts had already been applying more nuanced deference standards in earlier cases like United States v. Mead. Moreover, Loper Light may have created a new problem: judges substituting their own policy preferences for agency judgments, which might produce worse regulations in technical fields like AI than agencies would have developed in the first place.

The answer, according to legal analysis, lies not in power but in persuasion. Instead of courts deferring to agencies or courts overriding them, the framework should require agencies to justify their reasoning so thoroughly that courts can evaluate it on its merits. This echoes an older standard called Skidmore deference, named for a 1944 Supreme Court case. Under Skidmore, courts give weight to agency interpretations based on how well-reasoned they are, not because the law requires deference. An agency with a solid technical argument gets respect; an agency with sloppy reasoning does not. This approach forces agencies to do better work while preserving the judiciary's role as a check on administrative power.

For AI specifically, the framework should treat federal regulations as floors, not ceilings. A federal baseline would set minimum standards for safety and transparency, but state tort lawsuits could proceed on top of that baseline. Companies would face both federal requirements and potential state liability. This sounds inefficient—and it is, in some ways. Fragmented state liability regimes do create compliance costs. But those costs are justified when the alternative is regulatory silence or overconfidence. State litigation serves as a continuous feedback mechanism. When an AI system causes harm in ways regulators did not anticipate, the lawsuit generates discovery, expert testimony, and appellate decisions that inform the next round of federal rulemaking. Information flows dynamically rather than being locked down by preemption.

The stakes are high because AI risks are not fully knowable in advance. They are probabilistic and emergent. A regulator might write rules based on today's understanding of how language models work, only to discover three years later that the technology has evolved in unexpected directions. State litigation provides a mechanism for surfacing those surprises. It also preserves what Justice Louis Brandeis called the states' role as laboratories of democracy—places where different regulatory approaches can be tested and compared.

But there is a secondary problem lurking in the post-Loper Light landscape. As federal agencies increasingly use AI tools to implement their own regulations, courts face a new form of deference: algorithmic deference. An agency might use an AI system to make decisions about which companies to inspect, which applications to approve, or how to interpret compliance data. These algorithmic choices embed policy decisions that are difficult to scrutinize. AI systems can hallucinate, can reflect the biases in their training data, and can be opaque even to their creators. Replacing judicial deference to human agency reasoning with judicial acceptance of algorithmic decision-making would be a step backward. The deliberative costs of requiring agencies to explain their reasoning—including the reasoning embedded in their tools—are justified in a field where uncertainty and systemic risk make transparency essential.