California's Generative Artificial Intelligence Training Data Transparency Act, known as AB 2013, has applied since January 1, 2026. It requires developers of generative AI systems that are available to Californians to post on their websites a high-level summary of the data used to train them, covering systems released or substantially modified since January 1, 2022. A federal court declined on March 4, 2026 to block the law at xAI's request, and the Ninth Circuit appeal, with argument scheduled for November 18, 2026, had no ruling in the sources checked on October 27, 2026.
Key takeaways
- The duty sits in California Civil Code sections 3110 and 3111 and takes the form of a public web posting, not a filing with a regulator.
- Twelve categories of information are required, from data sources and volume to personal information and synthetic data.
- Retraining or fine-tuning that materially changes a system counts as a substantial modification and triggers a new posting.
- Three exemptions cover security systems, aircraft operation, and some national security systems.
- xAI sued the California attorney general on December 29, 2025, and the first-round request for an injunction was denied.
What the statute requires
AB 2013 defines a developer as a person, company, or government agency that designs, codes, produces, or substantially modifies an AI system or service for use by the public. Generative AI means AI that produces synthetic content such as text, images, video, and audio. Training includes testing, validating, or fine-tuning by the developer. Under section 3111, on or before January 1, 2026, and before each later release or substantial modification, the developer must post documentation about the data used to train the system on its website. The duty applies whether or not the public pays to use the system.
The bill was authored by Assemblymember Jacqui Irwin and signed by Governor Gavin Newsom in September 2024. Its stated purpose, in the author's words as reported at signing, was to let consumers judge whether they trust a system and compare competing ones.
The twelve items a posting must address
| Topic | What the developer must state |
|---|---|
| Sources | Sources or owners of the datasets |
| Purpose | How the datasets further the system's intended purpose |
| Volume | Number of data points, which may be given as general ranges |
| Type | Label types for labeled data, or general characteristics for unlabeled data |
| Intellectual property | Whether data is protected by copyright, trademark, or patent, or is entirely public domain |
| Acquisition | Whether the datasets were purchased or licensed |
| Personal information | Whether the data includes personal information, and whether it includes aggregate consumer information |
| Processing | Any cleaning or modification, and its purpose |
| Collection period | Time span of collection, with notice if it continues |
| First use | Dates the datasets were first used in development |
| Synthetic data | Whether synthetic data generation was or is used |
Personal information and aggregate consumer information are two separate items in the statute, which brings the count to twelve. The list does not require publishing the datasets themselves, model weights, or source code.
Who is exempt
Documentation is not required for a system whose sole purpose is security and integrity, one whose sole purpose is operating aircraft in the national airspace, or one developed for national security, military, or defense purposes and made available only to a federal entity. Because "developer" includes anyone who fine-tunes a system for public use, a law firm alert from January 2026 reads the statute as reaching beyond the makers of the largest foundation models.
How companies responded
According to that law firm alert, OpenAI posted a short summary touching each required category, and Anthropic posted a longer, itemized document. Neither named specific datasets; both described data in broad classes such as publicly available information, third-party partner data, user data, human evaluator data, and synthetic data. The statute does not define how detailed a "high-level summary" must be, which is the main open interpretive question. The alert also says that how California will enforce the law remained to be seen.
The xAI lawsuit
xAI sued California Attorney General Rob Bonta on December 29, 2025, in the U.S. District Court for the Central District of California (case 2:25-cv-12295), and on January 16, 2026 it moved for a preliminary injunction. The motion argues three points: that the law takes trade secrets in violation of the Takings Clause, that it violates the First Amendment, and that it is unconstitutionally vague. Judge Jesus Bernal denied the motion on March 4, 2026, finding that xAI had not shown a likelihood of success on the merits. That ruling concerns interim relief only and does not decide the case. xAI appealed to the Ninth Circuit (No. 26-1591). According to the Center for AI and Digital Policy, which tracks the case, California filed its answering brief on July 15, 2026, xAI's reply followed on September 4, 2026, and oral argument is set for November 18, 2026 in Pasadena. No appellate ruling had been reported by October 27, 2026. The Ninth Circuit docket itself was not checked, so these dates come from a secondary source.
Questions a developer can ask
- Is the system available to people in California, and was it released or modified on or after January 1, 2022?
- Has fine-tuning or retraining since then materially changed its functionality or performance?
- Does any exemption apply to the system's sole purpose?
- Who maintains the posting, and what triggers an update?
This article is general information about a statute and a lawsuit. It is not legal advice, and a developer with a specific system should consult a qualified lawyer and the statute text.
FAQ
Does AB 2013 force companies to reveal their training data?
No. It requires a high-level summary in twelve categories. The statute's list does not include the datasets, model weights, or code.
Is the law currently in effect?
Yes, since January 1, 2026. The federal district court refused to suspend it on March 4, 2026, and no order from the Ninth Circuit blocking it was found in the sources checked.
Does it apply to companies outside California?
It applies to systems made available for use by Californians, wherever the developer is based.
Does fine-tuning an existing model trigger a new posting?
It can. A substantial modification is an update that materially changes functionality or performance, and the definition includes the results of retraining or fine-tuning.
Related reading
- Who owns AI systems and their outputs
- Human-centered approaches to AI
- State education codes and AI
- Large investors and AI infrastructure
More in AI Tools and Industry.
Last verified: October 2026
Sources
- California Civil Code sections 3110 and 3111 (AB 2013, Statutes of 2024), California Legislature, effective January 1, 2026
- xAI's motion for preliminary injunction, xAI LLC v. Bonta, No. 2:25-cv-12295, U.S. District Court, Central District of California, filed January 16, 2026
- California's AB 2013 Takes Effect, Goodwin Procter, January 2026
- New California Law Requires AI Training Data Transparency, September 28, 2024
- Training Data Transparency (AB 2013), Center for AI and Digital Policy, case timeline, 2026