AI Regulation and the Looming Problem of the Takings Clause
In December 2025, Elon Musk’s artificial intelligence (AI) company, xAI, filed a lawsuit against the attorney general of California to prevent the enforcement of AB 2013, which was scheduled to take effect on Jan. 1. AB 2013 requires generative AI developers to publicize “high-level” information about the training data used to develop their models, including the sources and owners of datasets used, whether that data is copyrighted or contains personal identifying information, and the manner in which the data was collected. xAI alleged that both state and federal trade secret law protects this type of information. Given its purported property interest in keeping that information secret, the company claimed that the required disclosures were an uncompensated “taking” in violation of the Fifth Amendment Takings Clause, which prevents the government from taking private property for public use without compensation.
On March 4, a federal court in the Central District of California dismissed xAI’s request for a preliminary injunction because it had failed to show that the AB 2013’s list of high-level disclosures actually threatened any trade secrets. (That decision is now on appeal before the U.S. Court of Appeals for the Ninth Circuit.) But this is likely not the end of this particular legal theory. Regulators often favor disclosure mandates to address the risks posed by AI, and eventually some of those disclosure mandates will become granular enough to threaten trade secrets, even if that was not true of the very high-level information mandated under AB 2013. The question then becomes: Does the Takings Clause prohibit transparency-focused regulations that publicize proprietary information? An interplay of legal, technological, and political factors has made a collision between AI regulation and the Takings Clause increasingly likely, with the result that the Takings Clause could pose a significant barrier to many current AI regulatory proposals. This article walks through these legal, technological, and political considerations and how they relate before reflecting on what this means for AI policy and AI regulation.
The Law
Whether intellectual property counts as “private property” within the meaning of the Takings Clause is not exactly clear. Patents, for instance, may not be: In 2006, the U.S. Court of Appeals for the Federal Circuit held that victims of government patent infringement cannot assert a constitutional claim against the government, and in 2018, the Supreme Court held that patents are “public franchises” that the executive branch may cancel in at least some circumstances. Back in 1984, however, the Supreme Court held in Ruckelshaus v. Monsanto that trade secrets count as “private property” for Takings Clause purposes. Among scholars, this decision was controversial, with some scholars critiquing the treatment of information as a form of constitutional property while others embraced it wholeheartedly. But while the Supreme Court has not taken another case involving an alleged taking of a trade secret since 1984, it continues to cite Monsanto favorably.
If the Takings Clause protects an asset, evaluating whether government regulation “takes” the asset proceeds in two stages. First, a regulation may be deemed a “per se” taking if it eliminates “all economically beneficial use” of the property—what’s known more colloquially as the total wipeout rule. (There are other per se rules, but they are largely irrelevant to trade secrets.) If the regulation doesn’t eliminate the complete economic value of an asset, it’s evaluated under a three-factor test established by Penn Central Transportation Co. v. City of New York, which asks whether the regulation substantially diminished the economic value of the asset, interfered with reasonable investment-backed expectations, or had the “character” of a classic, eminent-domain-style taking. (What that last one means is often mysterious.)
There is not a great deal of case law on how these standards apply to regulations affecting trade secrets. Monsanto itself concerned a complicated statutory scheme requiring applicants for pesticide registration to submit proprietary data to the Environmental Protection Agency (EPA). The Supreme Court held—at least for a period where the statute promised secrecy for that data—that the EPA could not repurpose the data to approve generic competitors without compensating the original applicant. Only once, in Philip Morris v. Reilly, a case decided by the U.S. Court of Appeals for the First Circuit, has a court invalidated an entire statute on the basis that it took trade secrets without compensation. And even there, the two-judge majority disagreed on why. One judge thought that forcing tobacco companies to disclose their formulas was a total wipeout of a property right, as disclosing a trade secret to the public fully eliminates trade secret protection. The other judge thought that the total wipeout rule should apply only to regulations of land but that the Massachusetts law at issue nevertheless failed the Penn Central test. Philip Morris and Monsanto show that courts might protect trade secrets from government interference in at least some circumstances, but little guidance exists on what specific regulations might result in invalidation—a feature that makes this area of the law unpredictable.
Although courts have rarely weighed in on disputes concerning alleged takings of intellectual property, academics have debated the issue for a few decades. Some have worried that as trade secret protection laws become more robust over time, the Takings Clause stands as a barrier to government interventions that attempt to incentivize transparency or that directly publicize corporate secrets. Others have discussed whether international law creates additional constraints on the ability of governments to disclose trade secrets. When it comes to evaluating whether specific types of reforms likely amount to takings, a more robust literature exists regarding patent reform. But while courts might draw from that academic literature in resolving disputes one day, the decades-long debate has not led to a consensus even on whether patents are protected by the Takings Clause, much less how to apply takings doctrine to patent regulations. So although a robust scholarly debate has emerged on the topic, this debate primarily underscores how uncertain the law is.
The Technology
Not many businesses have tried to challenge government regulation by alleging that it “takes” a trade secret. The few challenges that have been raised involve a miscellany of alleged trade secrets: tobacco formulas, historical prescription pricing information, and FDA-submitted drug formulas. But technology- and industry-specific factors suggest that AI companies may be particularly likely to raise such challenges.
First, there has been an overall trend for companies within the technology industry to increasingly rely on trade secret law—as opposed to other forms of intellectual property (IP)—as the primary legal regime to protect their innovations. For AI specifically, trade secrets are often the only available form of legal protection for the IP owned by AI companies. Not only has the U.S. Copyright Office taken the position that AI outputs are uncopyrightable, but AI weights—unlike human-written code—are the result of a mathematical training process that lacks the copyright requirement of human authorship. Patents are theoretically available for some types of AI innovations, but only if those innovations are (a) sufficiently specific to be defined in a patent application and (b) not simply an “abstract” idea (such as a description of a new algorithm or model architecture).
In practice, AI progress is driven not by new, definable inventions, but by a combination of improvements in data curation, large-scale computing expenditures, management techniques, and technical know-how. These types of improvements are individually (and collectively, in the form of the resulting model weights) protectable under trade secret law, but not under any other form of IP law. Because an owner of trade secrets loses their property right if the secret is disclosed to the public—unlike with copyrights or patents—this industry-wide reliance on trade secrets places stress on employment contracts, nondisclosure agreements, and other tools designed to keep the secrets, well, secret.
The other relevant feature of AI is the fact that it is highly susceptible to reverse engineering, even through limited forms of disclosure. With only the ability to observe the inputs and outputs of Model A, “model extraction” or “model distillation” techniques allow a third party to train a smaller Model B that can recreate Model A’s behavior for far less money. Similarly, “membership inference” attacks allow a third party to use targeted inputs to evaluate whether particular data was included in a model’s training data. And disclosing even high-level information about how data is collected can allow for targeted “data poisoning” attacks, in which a small number of manipulated data points are seeded online to induce undesirable behavior in AI models trained on that data. Companies that want to protect their model weights and prevent manipulation of their model’s performance thus have strong reasons to control what information they disclose about their training and deployment pipelines. (Consider just how carefully Anthropic has controlled both information about and access to its new Mythos model.)
The Politics
On the other side, “transparency” has become a preferred proposal for regulating AI models and addressing the potential harms of AI systems. There are a number of distinct reasons for this, but the proposals have become so commonplace as to be a default premise for government intervention. (Notably, AI companies themselves have sometimes invited this reaction by refusing to cooperate in even minimal investigations of their products’ quality—for instance, by relying on trade secrecy to prevent scrutiny of algorithms used in criminal proceedings, where their products are used to influence decisions affecting defendants’ rights.)
Consider the issue of AI fairness. In 2016, ProPublica’s reporting on a widely used recidivism-prediction algorithm revealed that it was more likely to erroneously predict that a Black defendant would reoffend than a white defendant. At first, it seemed a way to address this problem might be to mandate that certain AI programs satisfy some quantitative “fairness constraint.” But computer scientists subsequently demonstrated that whenever background rates of reoffense or rearrest are unequal, an algorithm that satisfies multiple intuitive ways of measuring fairness is mathematically impossible to design. Legal scholars questioned whether it would be constitutional to require algorithms to satisfy any type of group fairness constraint. And although a number of federal agencies have warned companies that the use of an opaque, complex AI system does not shield them from liability under anti-discrimination law, these agencies have made little progress in providing substantive metrics for identifying algorithmic discrimination.
As a result, political attention shifted significantly toward attempts to mandate disclosures that would reveal whether a model was trained on biased data. In Congress, the Algorithmic Fairness Act of 2020 would have required AI companies to maintain detailed reports when their models were used to make decisions about individuals, and to provide affected consumers with data compiled about them upon request. The Justice in Forensic Algorithms Act of 2021 would have eliminated any trade secrecy privilege when AI algorithms were used in connection with criminal proceedings. And the Algorithmic Accountability Act of 2025 would have required AI developers to produce detailed “impact assessments” of their models and share summaries with the Federal Trade Commission, which in turn would summarize those summaries for the public.
Transparency has been a prominent theme in AI-related legislation coming out of California (and at the state level more generally, as regulators grapple with unfamiliar technologies). AB 2013 itself was passed to promote consumer choice by requiring AI companies to disclose high-level but relatively wide-ranging information about training datasets, theoretically enabling more informed consumer choice among models. SB 53, known as the Transparency in Frontier Artificial Intelligence Act, addressed risks posed by frontier algorithms by requiring companies to disclose information about their safety testing procedures. SB 53 was passed after Gov. Gavin Newsom vetoed SB 1047, which would have imposed more substantive legal obligations, such as “kill-switch” shutdown tools and developer liability for catastrophic harm. It seems that while jurisdictions such as California are interested in regulating AI, they remain unsure how to design appropriate and adaptable substantive liability or behavioral rules, causing policymakers to often fall back on transparency-related measures as a default means of government intervention. And as more substantive regulation at the federal level appears to have stalled, state-level experimentation becomes even more likely to involve detailed disclosure mandates.
Paths Forward
Most current proposals focusing on transparency tend to demand very high-level disclosures that don’t appear to implicate trade secrets in the first place. But the lurking issue of the Takings Clause means that AI policy advocates need to be aware of the risk that more granular disclosure mandates could face significant legal scrutiny. Although no silver bullet exists for the problem, policy advocates should consider the following four proposals when debating or drafting AI regulation. These proposals are offered in the spirit of risk management and not as a firm prediction about how courts will actually rule in concrete cases; as discussed above, the law on this topic is uncertain and evolving. But by being aware of litigation risks and the ways to adjust regulatory proposals, policymakers can develop proposals that face a lower risk of being invalidated by courts.
Avoid making “transparency” the preferred solution to all AI problems. The AI fairness example highlights that regulators sometimes gravitate toward transparency-based interventions because it seems too hard to design substantive performance standards that can satisfy everyone. But this can risk satisfying no one: Merely disclosing the existence of “biased data” doesn’t solve discrimination, and AI companies themselves might prefer to be regulated by clear, numerical “fairness constraints” that don’t require any disclosures about their training data. Rather than treating disclosure as an unobjectionable, first-step intervention toward AI regulation, regulators should consider that substantive rules can be promulgated even under conditions of opacity. This can sometimes be done by requiring numerical performance standards or by relying on existing liability frameworks to address potential AI harms.
Design disclosure obligations to avoid constitutional issues. Where disclosure is promising as a regulatory intervention, legislators and regulators need to be more careful to avoid triggering the Takings Clause. Fortunately, the original Monsanto case provides a bit of a road map to accomplish this. Where disclosure to the government is a precondition for some benefit (such as a license to operate a model), simply refusing to promise continued secrecy can provide cover to disclose trade secrets to the public later if necessary to serve some public purpose. It’s not entirely clear just how far this adjustment, by itself, will go. A number of intellectual property scholars have argued that the absence of a statutory promise of secrecy is sufficient to defeat any takings claim with respect to proprietary information submitted to the government and subsequently disclosed to the public. But others reject this view and argue that requiring a company to forfeit its property rights in information as a precondition for a government license violates the Supreme Court’s doctrines regarding “unconstitutional conditions.” Even so, if AI regulation involves required submissions to the government, clarity about whether that information can be disclosed to the public, and when, will at least reduce the likelihood of subsequent takings-based challenges.
Provide mechanisms for raising compensation claims. Perhaps most importantly, regulators should incorporate mechanisms for AI companies to raise challenges other than via a lawsuit in federal court. In Phillip Morris, the First Circuit invalidated an entire statute because it provided no mechanism for compensation to tobacco companies. But in Monsanto, the Supreme Court blessed a system that required dissatisfied pesticide companies to first proceed to arbitration to value their trade secrets (in a forum overseen by the EPA itself, no less) and to use court only as a last resort to challenge the arbitrator’s decision. The existence of such arbitration procedures (a) makes it unlikely that a court would strike down an entire statute as “taking” trade secrets, while also (b) making it less likely that a court would second-guess the value of compensation identified by an arbitrator even in individual cases.
Be prepared to litigate. Finally, if all else fails, regulators who pursue transparency-focused regulation that implicates trade secrets should be prepared to litigate the constitutionality of those regulations. As discussed above, sparse case law exists for applying takings doctrine to intellectual property. But takings doctrine contains a number of “outs” for upholding uncompensated government action, at least some of which might apply when the government compels the disclosure of information about AI models. For instance, takings of property in response to emergencies can be uncompensated. And states have traditionally enjoyed more leeway to regulate property interests when the regulation aims to protect public health. At least some AI disclosure proposals—such as requiring frontier model developers to share information about their safety testing procedures—are arguably justified as an extension of those principles. Whether courts would accept those extensions can be determined only through litigation.
