Trade Secret Law May Block AI Transparency Regulations

Constitutional challenge threatens AI disclosure mandates

When xAI sued California in December 2025 to block enforcement of AB 2013—a law requiring AI developers to disclose high-level information about their training data—the company raised a novel constitutional argument: forcing disclosure of trade secrets amounts to an uncompensated "taking" under the Fifth Amendment. Though a federal court dismissed xAI's preliminary injunction request in March, finding the law's disclosures too general to threaten actual trade secrets, the legal theory remains very much alive as the case proceeds to the Ninth Circuit.

The collision between AI regulation and property rights law represents more than a single lawsuit. It signals a fundamental tension between the technology industry's growing reliance on trade secret protection and policymakers' preference for transparency-based regulation.

Why it matters

As federal AI regulation stalls, states are experimenting with disclosure mandates as their primary regulatory tool. If courts ultimately rule that the Takings Clause protects AI trade secrets from forced disclosure, entire categories of proposed transparency regulations could face invalidation—forcing a wholesale rethinking of how governments can oversee AI systems without either compensating companies or abandoning oversight entirely.

The uncertain legal landscape

The constitutional status of intellectual property as "private property" remains murky. While a 1984 Supreme Court decision in Ruckelshaus v. Monsanto held that trade secrets qualify for Takings Clause protection, courts have rarely applied this principle since. Only one appellate case—Philip Morris v. Reilly—has invalidated a statute for taking trade secrets without compensation, and even there the two-judge majority disagreed on the reasoning.

If a regulation does "take" protected property, courts apply either a "total wipeout" test (whether the regulation eliminates all economic value) or a three-factor balancing test from Penn Central Transportation Co. v. City of New York. The problem: virtually no case law explains how these standards apply to trade secret disclosures, leaving the legal terrain highly unpredictable.

Why AI companies are uniquely vulnerable

AI developers rely on trade secrets far more than traditional intellectual property protections. The Copyright Office has ruled that AI outputs lack human authorship required for copyright protection, while AI model weights—the core asset of most AI companies—result from mathematical training processes that similarly lack copyrightability. Patents require specific, non-abstract inventions, but AI progress stems from incremental improvements in data curation, computing scale, and technical know-how that don't fit patent requirements.

This makes trade secrecy the only viable legal protection for most AI innovations. Unlike patents or copyrights, trade secrets vanish entirely once disclosed publicly—creating an all-or-nothing property interest that could trigger the "total wipeout" standard.

Meanwhile, AI systems are particularly vulnerable to reverse engineering. Model extraction techniques can recreate a model's behavior from limited input-output observations, while membership inference attacks can determine whether specific data appeared in training sets. Even high-level disclosures about data collection methods enable targeted data poisoning attacks.

The transparency default

Faced with uncertainty about how to regulate AI substantively, policymakers have defaulted to transparency mandates. California's AB 2013 and SB 53, congressional proposals like the Algorithmic Accountability Act of 2025, and numerous state-level bills all emphasize disclosure over performance standards or liability rules.

This trend accelerated after attempts to impose algorithmic fairness constraints ran into mathematical impossibility theorems and constitutional concerns. Unable to define what "fair" AI looks like numerically, regulators pivoted to requiring disclosure of training data and testing procedures.

Navigating the collision

The analysis, detailed by Lawfare Media, suggests policymakers should reconsider reflexive transparency mandates. Where substantive rules are feasible—numerical performance standards or traditional liability frameworks—they may prove more durable than disclosure requirements. Where disclosure is necessary, regulators should study the Monsanto precedent: when disclosure to government is a precondition for operating, explicitly avoiding promises of secrecy may provide legal cover for later public disclosure.

The fundamental challenge remains unresolved. As AI regulation moves forward, the Takings Clause looms as a potential barrier that could force difficult choices between effective oversight and constitutional constraints.

This analysis was originally published by Lawfare Media.