When Might State AI Laws Run Afoul of Pike?

Neil Chilson is Head of AI Policy at Abundance Institute. He is a lawyer and a computer scientist, former FTC Chief Technologist, and author of the book “Getting Out of Control: Emergent Leadership in a Complex World.” He writes the Getting Out of Control substack.

Kevin Frazier is the inaugural AI Innovation and Law Fellow at the University of Texas School of Law. He also is a Senior Editor at Lawfare, the co-host of the Scaling Law podcast, and author of the Appleseed AI substack.

States have broad authority under the Tenth Amendment to exercise their police powers. This traditionally looks like laws intended to advance the health, safety, and education of their respective residents. California, New York, and Colorado are among the many states weighing or, in some cases, enforcing AI-specific laws with those local interests in mind. Earlier this week, Governor Gavin Newsom signed a major AI bill into law and he has many more awaiting his signature. Other states seem keen to follow. Yet, the police powers of a state are not boundless. Congress can expressly preempt state regulation of interstate commerce. And even absent express preemption, when state laws impact interstate commerce, the exclusive regulatory domain of Congress, they may be challenged under the dormant Commerce Clause. Though no AI law has yet to face such a challenge, a proactive analysis of how such laws would fare can help inform ongoing conversations about the proper distribution of AI regulatory authority between the states and the federal government. In particular, policymakers could benefit from a better understanding of how burdensome such state laws could be on out-of-state AI development.

This article applies the Pike balancing test to three prominent state AI laws: California's SB 53, New York's RAISE Act, and the Colorado AI Act. We argue that these laws are vulnerable to dormant Commerce Clause challenges. First, they substantially burden interstate commerce due to the unique, unitary nature of AI training—unlike other products, AI models cannot be easily segmented or customized for different state requirements. Second, the putative local benefits of these laws remain speculative and scientifically unproven, even as they impose nationwide effects on a critical emerging technology. Finally, legislators have overlooked less burdensome alternatives that could address their concerns without projecting their regulatory preferences across state lines. While states possess broad police powers, these AI training laws may well exceed constitutional limits.

Pike Balancing

To assess whether state AI laws impermissibly burden interstate commerce, courts will turn to the framework established in Pike v. Bruce Church, Inc. This test governs facially neutral state laws—those that don't explicitly discriminate against out-of-state actors but may nevertheless interfere with Congress's constitutional authority to regulate interstate commerce. Unlike laws motivated by economic protectionism that obviously favor in-state economic activities and are swiftly struck down, facially neutral state laws require courts to engage in a more nuanced balancing analysis. The Pike Court specified that:

[w]here the statute regulates evenhandedly to effectuate a legitimate local public interest, and its effects on interstate commerce are only incidental, it will be upheld unless the burden imposed on such commerce is clearly excessive in relation to the putative local benefits. If a legitimate local purpose is found, then the question becomes one of degree. And the extent of the burden that will be tolerated will, of course, depend on the nature of the local interest involved, and on whether it could be promoted as well with a lesser impact on interstate activities.

The Court applied that test to strike down an Arizona law that regulated an Arizona company that grew cantaloupes of “superior quality” which it then shipped to California for sorting, inspection, packing, and shipping. That practice ran afoul of Arizona’s mandate that cantaloupes raised in the state be packed in State-approved containers. The melon farmers filed suit challenging the law as unconstitutional. The Supreme Court agreed. The Court concluded that the law burdened interstate commerce by directing that commercial activity wholly outside the state be forced into Arizona. Justice Stewart, writing for the Court in a unanimous opinion, also dismissed the purported local interests of the law--making sure everyone knew the superiority of Arizona cantaloupes by selling them in state-approved and labeled packaging--as “minimal” in light of that burden.

Subsequent decisions have helped flesh out the key factors in that test given the Pike court’s omission of specifics.

On the first question of whether a law burdens interstate commerce, courts have relied on two considerations. First, somewhat obviously, whether the law does indeed have demonstrated broad, market-wide consequences for interstate commerce or whether its effects are predominantly local. If there’s no burden on interstate commerce, the court’s inquiry comes to a close. In such cases, a state is operating within the bounds of the Tenth Amendment and in a manner that does not conflict with Congress’s authority. This was the case in Exxon Corp. v. Maryland, in which the Court held that a state law prohibiting out-of-state petroleum producers from operating in the state did not result in a substantial burden on interstate commerce. The majority held that “[i]nterstate commerce is not subjected to an impermissible burden simply because an otherwise valid regulation causes some business to shift from one interstate supplier to another.”

Second, assuming the law addresses more than local conduct, the magnitude of the disruption or burden posed on interstate commerce. This consideration involves examination of evidence that the law disrupts the flow of interstate goods, changes the market structure, or compels an out-of-state actor to alter their business conduct. Pike serves as a prime example of a law forcing a business to upend their interstate operations. That law would have required any company growing cantaloupes in Arizona to replace existing packing facilities in other states with expensive facilities within Arizona, reducing the overall flow of interstate commerce. The state law in Exxon Corp., however, was found to have minimal to no effect on the overall interstate practices of the businesses in question; they simply had to exit retail gas station services within the state. There was no evidence that doing so would reduce the overall flow of fuel into Maryland from other states.

The next part of the test requires courts to perform a tricky balancing test: do the putative local benefits justify the burden imposed on interstate commerce? States cannot simply lean on the fact that a law is related to their police powers and move on. Courts require that they satisfy two inquiries: first, proof that the law will indeed further legitimate ends; and, second, that those resulting benefits are, for lack of a better phrase, worth it.

Courts look to evidence that the selected policy will indeed further those ends and do so to such an extent that justifies any resulting effects on interstate commerce. The alternative--where the local benefits shown to derive from the policy pale in comparison to the national costs--amounts to one state effectively sending the costs of regulation to other states. The policy in question need not be the most effective policy but courts have looked for affirmative evidence that it will “likely contribute directly” to mitigating the harm in question, for instance. On the whole, “courts are to look at the benefits assumed to exist or to have existed, or the benefits commonly accepted or supposed." Once those benefits are identified, the court can then do a sort of cost-benefit analysis.

This is no easy task. Pike left a lot of details out such as whether moral interests should carry the same weight as economic interests. It’s also unclear how courts should weigh the underlying purpose of the law--though protectionism may not have been explicitly worked into the text of the bill, it may be the case that states had the intent of reaching beyond their borders and favoring the interests of their residents over those in sister states. Courts have also varied in the amount of evidence they want the state to provide to demonstrate the benefits in question. They have also weighed the extent to which less burdensome alternatives were evaluated.

A recent decision by the Supreme Court, National Pork Producers Council v. Ross, brought these ambiguities to the fore, but only resolved some of them. At issue was a California law requiring specific procedures for raising pigs whose meat was to be sold in California. Representative trade organizations sued, alleging that the law imposed undue burdens on interstate commerce. The decision is “fractured” in many dimensions. Four concurrences exposed significant fault linesin how the justices perceive the future of the dormant Commerce Clause, Pike balancing, and challenges arising from state laws with apparent extraterritorial effects. Notwithstanding the wide range of interpretations of the dormant Commerce Clause, several relevant doctrinal principles emerged in the majority opinion. First, Pike balancing remains the test for non-discriminatory state laws that nevertheless burden interstate commerce. Second, the discriminatory effects of state laws--intended or realized in application--are of significant value in the balancing exercise. Third, Pike balancing is an imperfect and arguably improper vehicle for evaluating whether putative local benefits justify costs on interstate commerce--the public and their representatives are the proper actors for conducting such analysis. Fourth, when forced to conduct Pike balancing, courts will dive deep into the record for evidence--details on how economic actors have responded to a law, specifics on any changes to production processes, analysis by expert academics, and much more.

It’s impossible not to read the Ross opinions without wondering just how long Pike will still be the law of the land. Still, it remains in place, and state legislators should consider how it potentially applies to their proposals.

Pike Applied to AI

Though states have passed or are actively considering many AI-related bills, three stand out as particularly worthy of analysis under Pike. SB 53 in California, the Responsible AI Safety and Education (RAISE) Act in New York, and the Colorado AI Act in various ways implicate how AI developers should train or evaluate their models prior to deployment.

SB 53, recently signed by Governor Newsom, mandates that qualifying AI developers create, publish, and follow a "frontier AI framework" that outlines how the lab will specifically mitigate "catastrophic risks.” Additionally, it obligates such labs to produce a transparency report any time they deploy or substantially modify a frontier model.

The RAISE Act subjects frontier AI developers to penalties if they deploy a model that the state attorney general subsequently finds “create[d] an unreasonable risk of critical harm.” It also directs developers to “implement appropriate safeguards to prevent unreasonable risk of critical harm.”

Finally, the Colorado AI Act compels labs to “use reasonable care to protect consumers from any known or reasonably foreseeable risks of algorithmic discrimination in the high-risk system.”

Impact on interstate commerce

These laws will substantially burden interstate commerce in AI.

AI training is an incredibly resource intensive process. Training runs occur over the course of months, require ever-greater amounts of compute (which may be sourced from multiple states), include data sourced from myriad sources, and call on files stored across the country. What’s more, once deployed, these models may alter labor markets, accelerate economic productivity, and otherwise cause substantial macroeconomic effects. These facts demonstrate that each of the acts discussed above has potential massive ramifications on interstate commerce.

These and other unique attributes of AI training distinguish a potential Pike challenge to these bills from the challenge in Ross. Whereas a pig farmer can fairly easily (albeit at some cost, which they may be able to pass on to the regulating state’s residents) alter production processes to comply with one state’s regulations, the same is not true here. AI training is not a modular, segmentable process. It is a unitary process that cannot be done on a state-by-state basis; neither OpenAI nor any other lab has the resources to train 50 models tailored to the specifications of each state. Rep. Ted Lieu (D-CA) raised this very point during a recent hearing on AI. He asked a witness to assess if a lab could comply with even two different training regimes imposed by different states. He and the witness acknowledged the infeasibility.

It’s also the case that the law in Ross and the AI training laws here have a very different impact on the nature of the underlying product. Bacon from pig A and bacon from pig B is not likely to taste much different regardless of the pig’s background. In stark contrast, a model trained to meet New York’s mandates under the RAISE Act, for example, may differ wildly from one meant to comport with Colorado law. Even where the legislative text is similar, such as the training requirements imposed by New York and Colorado, the obligations on labs may shift as each state interprets their respective tests through enforcement actions. SB 53, which is described as a mere “transparency bill,” and lacks the substantive requirements of the other bills, still is obviously intended to shape the culture of the businesses covered and the types of models they develop. The resulting models may perform in noticeably distinct ways. This means that an AI model trained to the satisfaction of one state’s regulators will affect the model’s quality, not just its price, for users in other states.

This alteration to quality combined with the infeasibility of customising AI models for each state eliminates user options, unlike the law in Ross. California’s pork law may have restricted choices in California, but it did not hinder other Americans from finding pig products of vast diversity, reflecting a wide range of farming practices. In contrast, AI users across the U.S. will be limited to models trained to the liking of the New York or Colorado state legislators, and under the business practices required by California.

What’s more, these effects would penetrate beyond the AI models to much of the rest of the AI market. The handful of leading models–ChatGPT, Gemini, Claude, and Grok–directly support some of the most popular apps ever created. But many other software tools also rely on these powerful models to drive their AI functions. Any changes made to core models due to AI training bills will affect the quality and functionality of many downstream AI tools.

A further distinguishing consideration is that these laws, by virtue of targeting frontier AI models, implicate the nation’s economic interests as well as its geopolitical ambitions. The capacities of these models are closely monitored by the White House and the entire defense sector as China and other nations aim to develop and deploy the most sophisticated AI and to do so sooner than everyone else. Of course, pig production is economically important, but the facts of the Ross case lacked this clear nexus with national security concerns. When it comes to enforcing SB 53, the RAISE Act, and the Colorado AI Act, however, a single state could have a substantial effect on vital and sensitive national security questions. While the effect on national security is not a separate prong of Pike balancing, it warrants inclusion here given that Pikeand the Commerce Clause are intended to protect federalist values, including but not limited to congressional authority over interstate commerce and the equal sovereignty of each state.

Balance against purported local benefits

This leaves the question of whether the purported local benefits outweigh the effects on interstate commerce. The aforementioned AI training bills aim to shield their residents from speculative, yet substantial harms. SB 53 warns of “catastrophic risks from both malicious uses and malfunctions, including artificial intelligence-enabled hacking, biological attacks, and loss of control.” The co-author of the RAISE Act cited the bill as a means to “keep everyone in the city, state, country, and world safe from some pretty extreme risks.” The Colorado AI Act reflects the legislature’s determination that “this act is necessary for the immediate preservation of the public peace, health, or safety.” Providing the means to protect the lives of their residents surely seems to qualify as a legitimate use of state police power. That said, as revealed by the RAISE Act author, the motives of at least some supporters seem to extend beyond state borders. This express extraterritorial intent may factor into this balancing question, as explored further below.

The putative benefits of the selected policies, however, are in question. It is not clear that there are any “benefits commonly accepted or supposed.” The International Scientific Report on AI acknowledged as much when it noted, “Policymakers will often have to weigh potential benefits and risks of imminent AI advancements without having a large body of scientific evidence available.” So far, many legislators have balked at that suggestion (in line with the requirements of Pike). Just one year ago, eight members of California’s federal delegation, including Rep. Zoe Lofgren (D-CA) and Ro Khanna (D-CA) had the following to say about legislation, SB 1047, which was similarly intended to prevent worst-case scenarios from AI:

• “There is little scientific evidence of harm of mass casualties or harmful weapons created from advanced models.”
• “[P]remature requirements based on underdeveloped science call into question from the outset the efficacy of the bill in achieving its goals of protecting public safety.”
• “The current state of the technical solutions that would underpin implementation of SB 1047, including standards, benchmarks, and evaluations, is significantly underdeveloped.”

Notably, they urged Governor Newsom to veto that bill--a suggestion he followed. The passage of time has not facilitated any consensus. Earlier this year, the Joint California Policy Working Group on AI Frontier Models, assembled by Governor Newsom, made similar observations:

• “Evaluations are currently constrained by a limited set of techniques for eliciting capabilities.”
• “Overall, current evaluation techniques are nascent and ad hoc, and they have not yet achieved the scientific rigor expected in other domains where long experience with widespread societal use generates greater data and systematic evaluations that more directly guide policy.”
• “Standard measures of capabilities . . . may serve as a poor proxy for model capabilities in the contexts where models are used.”
• “Experts significantly disagree on whether capabilities will advance slowly, rapidly, or extremely rapidly in the coming years, with partial evidence substantiating all three views.”

Supporters of these measures may counter that the resulting information collected from these various requirements will meaningfully improve transparency around AI model development. While that may be true and beneficial, such improvements in information are far short of actually achieving the underlying purpose of reduced exposure to significant risks. As the First Circuit recently concluded in Penobscot Bay & River Pilots Association v. Town of Bar Harbor, Pikerequires evidence as to the actual magnitude of the intended putative local benefits. From the prior statements, it is unclear just how much, if at all, these untested policies will reduce the likelihood of catastrophic risks and related harms.

Courts will also query the absence of states considering less restrictive alternatives to counter the alleged harms. The Supreme Court made this clear in Great Atlantic & Pacific Tea Company v. Cottrell, when it announced:

Inquiry whether adequate and less burdensome alternatives exist is, of course, important in discharge of the Court's task of accommodation of conflicting local and national interests, since any realistic judgment whether a given state action unreasonably trespasses upon national interests must, of course, consider the consequences to the state if its action were disallowed.”

This aspect of Pike survived Ross and appeared in Penobscot Bay & River Pilots Association. There the First Circuit flagged the importance of reviewing the viability of alternatives--specifically whether they would be “more burdensome on commerce” or “less likely to be effective as compared to the challenged law.”

In this context, myriad different approaches to intervening in the AI training process exist that would not raise the same concerns about projecting one state’s laws into another. Rather than interfere with the AI training process, regulators concerned about risks such as the creation of chemical, biological, radiological, or nuclear (CBRN) weapons could impose new restrictions on access to the physical components of those destructive tools. Notably, Rep. Lofgren and others pointed out this alternative in their 2024 letter. Legislators concerned about discriminatory decisionmaking in important local industries could strengthen discrimination standards, add substantive requirements, and increase enforcement resources in those industries. Concerns about models being leaked or stolen via a hack could be mitigated by increased cybersecurity specifications for labs. The list could go on. Yet, alternatives seem to have not been rigorously assessed. If they had been, the widely available research on the merits of different CBRN policies, for instance, may have revealed that these less burdensome alternatives are just as effective, if not more so, than the largely untested policies proposed in the AI training bills.

The motives of the legislature also may tip the scales as the courts balance the putative local burdens against the impact on interstate commerce. Protectionist purposes are antithetical to federalist principles and the desire of the founders to ensure that national matters were addressed through harmonized policy. Here, supporters of the AI training bills have not been shy about their desire for states to fill the perceived gap left by ongoing congressional deliberation on this topic. California State Senator Scott Wiener, for example, has implied that his AI bills--last year SB 1047 and this year SB 53--are necessary so long as there is a federal vacuum on mitigating frontier AI risks. As noted above, the backers of the RAISE Act have made similar statements. Yet, there is no “state savior” exception for California, New York, or any other state with grave concerns about AI worst-case scenarios. Deliberative Congressional consideration is not an invitation for states to jump in and try their hand at addressing national issues. To think otherwise is to contradict the profound desire of the founders to ditch the Articles of Confederation, a system under which single states could derail national goals and impose their policy preferences on their neighbors.

The majority opinion in Ross echoes long standing concern about “judges using the dormant Commerce Clause as a roving license for federal courts to decide what activities are appropriate for state and local governments” (quotations omitted) and explains the difficulty of courts balancing noneconomic concerns against economic concerns in a cost-benefit. But courts surely retain the capability and responsibility to rein in states that would use their own police powers as a “roving license” to regulate what AI products and services moving in interstate commerce will be available to other states’ residents. Courts may struggle to balance non-economic concerns against economic concerns, but perhaps the job is easier when a court weighs whether Colorado's non-economic concerns about algorithmic discrimination should overwhelm Ohio’s or South Carolina’s non-economic concerns about politically biased AI. The answer should be “no.”

Conclusion

When, not if, an AI training bill such as SB 53, the RAISE Act, or the Colorado AI Act gets challenged by an out-of-state resident because of undue interference, Pike balancing factors appear to tip in favor of the challengers. The current state of AI research from leading experts as well as statements from attentive and knowledgeable lawmakers suggests substantial disagreement as to the putative local benefits of these policies. The benefits from AI frameworks and vague standards of care are not “commonly accepted or supposed.” Yet there is strong evidence that such laws will impose a substantial burden to interstate commerce. Proponents of AI training bills also seem to have overlooked less burdensome alternatives, despite an expansive menu of different approaches. Finally, many of these individuals have also not hidden their desire to bypass the constitutional safeguards of horizontal and vertical federalism.

State legislators should carefully consider the effect of their AI regulatory proposals on interstate commerce. They certainly should not view such effects as a feature. Harming interstate commerce in AI is a bug in such proposals, and a potential legal vulnerability.