The Challenge of Regulatory Preemption in AI Governance

Since the U.S. Supreme Court overruled Chevron v. Natural Resources Defense Council in Loper Bright Enterprises v. Raimondo in 2024, much ink in legal scholarship has been spilled over the separation of federal powers. This new precedent, however, also reshapes a longstanding tension between federal administrative agencies and state tort law: preemption.

Preemption—the principle that federal laws supersede contrary state laws—is grounded in the Supremacy Clause of the U.S. Constitution, but a modern tension facing new technologies lies specifically in regulatory preemption, the process by which federal administrative agencies, through statutory interpretation and regulatory implementation, shape the federal laws that displace state regulatory and tort frameworks. In the context of the governance of artificial intelligence (AI), this tension between the federal administration and the states is existential: It threatens the information-producing function of state-level litigation. An early draft of President Donald J. Trump’s mega spending law, for example, proposed a 10-year moratorium on state-level AI legislation, although this provision was ultimately removed from the final version. This illustrates federal preemption in emerging technology governance.

For decades, Chevron called for judicial deference to federal administrative agencies. Courts were often compelled to accept an agency’s interpretation of an ambiguous statute, so long as that interpretation was reasonable. Before Loper Bright, if a federal agency had established regulations on AI, it could have not only set regulatory standards for the emerging technology but also interpreted these standards as preempting the state-level tort liability.

Such “judicial acquiescence” was criticized as depriving states of their role as laboratories of democracy able to test out different policy approaches—what Justice Louis Brandeis called “experimentation.” State-level litigation can generate information or feedback that enables regulators to calibrate policy amid shifting technological realities. Judicial deference to federal agencies stifles such information production.

So, the real democratic failure of Chevron was not just deference itself but that agencies did not need to provide robust explanations of their reasoning, particularly when agency interpretations operate to preempt state tort law. Particularly in the realm of AI governance, where information asymmetry exists between federal regulators and AI developers, such compulsory deference lessened federal agencies’ obligation to gather dynamic feedback or articulate persuasive, technically sound justifications for regulations. By cutting off the ground-truth signals of state litigation, judicial deference fostered the risk of unelected administrative agencies overriding state laws based on federal authority, and generating ill-fit or rigid, ex ante regulation without persuasive justifications. From this perspective, preemption is not just about federal supremacy, but about information suppression versus information production. When governing AI, risks are probabilistic and emergent rather than strictly predictable, so a more nuanced interpretative framework is needed.

Of course, in 2024, the Supreme Court in Loper Bright overruled Chevron, replacing “deference” with “delegation” of authority. On its face, the holding stopped federal agencies from interpreting for themselves the limits of their regulatory authority, thereby preventing them from declaring the rules to be preemptive regulatory ceilings.

The reality, however, is more complex. As Cary Coglianese and David B. Froomkin argue, the Loper Bright ruling is “disingenuous.” It does not contain a meaningful, substantive departure from Chevron. Long before Loper Bright, decisions such as United States v. Mead had applied Chevron’s holding only to statutes that indicated a congressional intent to delegate authority to an agency. Moreover, Loper Bright reflects a flawed interpretation of the Administrative Procedure Act. The Act empowers judicial courts to delineate the boundary of an agency’s interpretation authority, but the Court treats this as an authorization for courts to substitute their own policy preference for agency judgments, which may not lead to any better-informed regulations, particularly for novel and technical areas such as AI, than what agencies would have produced.

If Loper Bright is no improvement, how should we govern AI risks? The answer lies in persuasion rather than power. Fractured state tort liability regimes often cause higher transaction costs and compliance requirements as friction in horizontal federalism, generating negative economic consequences in the national market. Catherine Sharkey’s “agency reference” model provides a promising alternative. Under this proposal, federal regulations, grounded in persuasive reasoning, would establish uniform regulatory baselines while preserving room for state tort law to impose liability. Preemption as a regulatory floor, rather than a ceiling, would ensure that information flows dynamically because a regulatory floor allows state tort lawsuits to proceed, effectively using judicial discovery processes to extract useful information. Given the information asymmetry between federal regulators and AI developers regarding systemic risks, federal regulations must not operate as a total preemptive ceiling.

In today’s post-Loper Bright landscape, judges should treat agency interpretation as persuasive rather than binding, returning to the level of deference required under Skidmore v. Swift & Co, where courts afford deference according to the persuasiveness of an agency’s reasoning. This form of deference enhances rational judicial decisions. Skidmore deference supports the concept of “regulatory excellence” because it respects agencies’ professional judgement when it is supported by a well-reasoned justification. Treating agency reasoning as persuasive serves as a superior substitute for deference and realizes democratic accountability. This approach requires agencies to articulate their reasoning, allowing others to contest agency interpretations and provide iterative feedback, with state tort litigation serving as a source of feedback.

Although Skidmore-style persuasion entails uncertainty and higher deliberative costs, such costs are normatively justified in fields such as AI governance, where epistemic uncertainty and systemic risks make the reason behind decisions important.

Post-Loper Bright, the administrative shift is not just from deference to persuasion, but also from statutory interpretation to practical implementation. While courts now scrutinize how agencies interpret the law, current administrative law lacks a framework for monitoring the implicit policy choices regulators make when they implement those laws using algorithms or AI. This reliance presents a secondary issue within the administrative state: replacing judicial deference with algorithmic deference. As AI tools increasingly infiltrate administrative agencies, hallucinations and sycophancy render them neither convincing nor accurate.

Normatively, encoded policy—policy decisions produced by algorithms or AI systems—should receive no deference. But as Danielle Citron has observed empirically, encoded policy receives more than Chevron-level deference. When regulators rely on AI to automate or assist decision-making, implicit policy choices and technical biases are embedded into codes. Because these choices are hidden within opaque algorithms, they may evade meaningful public or judicial scrutiny. In this sense, the problem is not excessive deference by courts but misplaced deference by agencies to algorithms or large language models. We do not need agencies that merely command; we need agencies that can convince. This conviction requires human accountability anchored in professional judgment rather than overreliance on computer outputs.