California’s new law requiring generative AI developers to publish high-level summaries of their training data remains in force upon surviving its first challenge in federal district court. This decision comes at a time when the federal government is also in the process of beginning to weigh in on the development of state AI laws. Notably, on March 20, 2026, the White House recommended a framework for AI legislative priorities, and the US Department of Commerce’s target date to publish an evaluation of existing state AI laws under Executive Order 14365 (Ensuring a National Policy Framework for Artificial Intelligence) (EO 14365) is pending. In this context, maintaining ongoing compliance efforts and proactively preparing for evolving state-level AI requirements remain high priorities for companies.

On March 4, 2026, the US District Court for the Central District of California denied X.AI LLC’s (X.AI) motion for a preliminary injunction seeking to block the enforcement of California Assembly Bill 2013 (AB 2013), also known as the Generative Artificial Intelligence Training Data Transparency Act (TDTA) (X.AI LLC v. Rob Bonta, Docket No. 2:25-cv-12295 (C.D. Cal. Dec 29, 2025)). As background, the TDTA was signed into law in 2024 and took effect at the start of this year on January 1, 2026 (Cal. Civ. Code § 3111). This new law requires developers of generative AI systems or services that are publicly available to California residents to post on their websites information about the data used to train those generative AI systems or services. The law also retroactively imposes disclosure requirements and applies to generative AI systems or services that were released or substantially modified on or after January 1, 2022.

Below, we provide a discussion of (1) the disclosure rules imposed by the TDTA, (2) the key claims and allegations from X.AI’s challenge to the TDTA’s disclosure requirements and a summary of the court’s reasoning for denying X.AI’s motion and (3) considerations for companies and developers of generative AI systems as transparency requirements from the TDTA remain in force.

The TDTA’s disclosure requirements

The information disclosure requirements of the TDTA apply to an AI system “developer,” defined as an entity who “designs, codes, produces, or substantially modifies an artificial intelligence system or service for use by members of the public.” Notably, the requirements also apply more broadly to entities who “substantially modif[y]” the AI system in a manner that “materially changes its functionality or performance, including the results of retraining or fine tuning.” Accordingly, entities that specialize in adapting general-purpose AI models to specific industries or applications are potentially subject to the disclosure requirements of the TDTA. As previously discussed, developers of AI systems subject to the TDTA have some latitude to describe the data used for training. The TDTA requires a “high-level summary of the dataset used in the development of the generative artificial intelligence system or service” including, but not limited to, the following topics:

  • Sources or owners of the datasets, and whether the datasets were licensed or purchased

  • A description of how the datasets further the intended purpose of the AI system or service

  • The time frame for collection and use of the data, and whether there was any cleaning, processing, or modification of the data

  • Types and numbers of data points in the dataset

  • Whether the datasets are in the public domain, and if they include any of the following types of information:

    • Data protected by copyright, trademark or patent
    • Personal information
    • Aggregate consumer information
    • Synthetically generated data

The TDTA exempts certain AI developers from the disclosure requirements, including developers of AI systems developed exclusively for a federal entity for national security, military or defense purposes, as well as AI systems whose sole purpose is the operation of aircraft in the national airspace, or to help ensure physical safety and network security and integrity.

The TDTA does not specify a particular enforcement mechanism or penalties for noncompliance. Nevertheless, an analysis of the bill provided in an assembly hearing on privacy and consumer protection suggests that enforcement would likely occur pursuant to California’s Unfair Competition Law, which allows the state Attorney General and other public prosecutors to bring civil actions against offending parties.

X.AI’s challenge in federal court

On December 29, 2025, shortly before the TDTA went into force, X.AI filed a complaint in the Central District Court of California against the California Attorney General, seeking declaratory and injunctive relief from the TDTA. X.AI alleged that the required disclosures amounted to a taking of trade secrets that would “gut the AI industry, violating numerous constitutional provisions along the way.” X.AI followed the complaint with a motion for a preliminary injunction of the TDTA that argued the same points on January 16, 2026.

In summary, X.AI argued that the TDTA was unlawful because:

  • The TDTA constitutes an unconstitutional taking of valuable trade secrets because information about training datasets is highly proprietary and essential for maintaining an edge over competitors even where sourced from publicly available materials as other AI companies may not have thought to use those same materials for their own models.
  • The TDTA’s disclosure requirements constitute a violation of the First Amendment by compelling X.AI to speak a particular message prescribed by the state of California, and the exemptions for security, aviation and military-related AI systems prove a clear viewpoint by the state that certain ideas are more valuable and worthy of protection than others.
  • The TDTA is unconstitutionally vague, as it fails to define key terms such as “dataset” or “data point” and does not specify how much information must be disclosed.

On March 4, 2026, the District Court denied X.AI’s motion, stating that, although training datasets and details about them could be trade secrets, X.AI’s own lack of specificity in its arguments made it difficult to make a determination that trade secrets were at play: “[X.AI] has made general arguments about datasets, dataset sizes, cleaning methods, and configurations of data, but has not identified any dataset or approach to cleaning and using datasets that is distinct from its competitors in a manner warranting trade secret protection.”

On the First Amendment issue, the District Court found that the TDTA could be a content-based speech regulation that regulates commercial speech and is therefore subject to lesser scrutiny than regulations of non-commercial speech. To find that the TDTA likely regulates commercial speech, the District Court reasoned that the TDTA did not include any regulation of AI models’ outputs based on their training data, and there was “nothing political” about consumers considering whether certain kinds of information, such as medical data, were used to train a model when choosing between options on the market. The District Court expressed skepticism towards X.AI’s argument that consumers cannot derive value from training-data disclosures, noting it “strains credulity to essentially suggest that no consumer is capable of  making a useful evaluation of [X.AI’s] models by reviewing information about the datasets used to train them.” At the stage of the preliminary injunction, the District Court found X.AI had shown only a “distinct possibility” of prevailing on the merits, but not the necessary likelihood of success to prevail.

The District Court also did not find X.AI’s arguments about vagueness convincing, stating that X.AI “takes issue with ‘dataset’ and ‘data point’ being undefined in the statute, yet [X.AI] seems to understand and use with ease ‘dataset’ throughout its Complaint.” As with the trade secrets issue, the District Court found that X.AI presented hypothetical issues with ambiguity over what systems would fall within the scope of the statute, but failed to provide an actual example that would allow the District Court to make a proper determination on the issue.

As such, the TDTA remains in force at a time when federal efforts to curb the development of state AI laws is still in progress. March 11, 2026, saw the expiration of the 90-day target imposed by EO 14365 for the Secretary of Commerce to publish its evaluation of existing state AI laws. The executive order established a process for the US Department of Commerce to identify state AI laws that conflict with a “minimally burdensome national policy framework for AI,” including laws that would require AI developers to disclose information in a manner violating the First Amendment or other provisions of the Constitution. Pursuant to EO 14365, these laws would then be challenged by a new AI Litigation Task Force headed by the US Attorney General, which was established on January 9, 2026. As the Department of Commerce’s first report has not yet been published, the scope and targets of any future challenges to state AI laws remain uncertain. However, the White House did publish a framework of legislative recommendations on March 20, 2026, which included a recommendation to preempt any state laws “that impose undue burden” on the development and deployment of AI. It is unclear whether the TDTA would fall in that category, and at this stage, the White Houses’ framework is only a recommendation and no federal legislative framework is pending. For the time being, states are not prohibited from passing or enforcing AI regulations.

Our take

The TDTA demonstrates the growing regulatory trend for more transparency in generative AI models. For example, New York’s pending Assembly Bill A6578B would establish another Artificial Intelligence Training Data Transparency Act with disclosure requirements, almost identical to the TDTA, for AI systems made publicly available to New York residents. Outside the US, Article 53(1)(d) of the EU AI Actrequire developers to publish a “sufficiently detailed summary about the content used for training” of the general purpose AI models marketed in the EU (Regulation (EU) 2024/1689). Although the TDTA appears to allow developers to satisfy their obligations through high‑level summaries of training data, companies may find it efficient to leverage disclosures developed under the EU AI Act when preparing disclosures under the TDTA.

Though new in its discussion of generative AI systems, the X.AI decision is consistent with recent court decisions upholding and enforcing AI regulations, as seen in the recent dismissal of a challenge to New York’s algorithmic pricing disclosure law by the Southern District of New York (National Retail Federation v. James, No. case 1 25-cv-05500-JSR (S.D.N.Y. Oct. 8, 2025)). Although some organizations may wait for further guidance and other potential legal actions, many are taking steps now to reassess and strengthen their AI policies and procedures to ensure compliance with applicable state AI laws. Accordingly, we are observing that most companies are electing to move ahead with TDTA compliance efforts, reflecting a broader market practice of preparing now rather than waiting for further regulatory or judicial clarification.

Compliance typically includes understanding and documenting training data practices during fine tuning and at each stage of the development cycle–including data acquisition, collection, processing, modification and publication. Obtaining this data can be challenging, especially where datasets are not clearly sourced or licensed, or where historical documentation is lacking for older systems within the scope of the January 2022 cutoff under the TDTA. At the outset, the TDTA may require audits of both internal and vendor documentation to bring older datasets into compliance, with an eye toward maintaining best practices for data management and hygiene moving forward.

Contact