California now requires every company shipping a generative AI model to the public to admit, in writing, roughly what it trained on. xAI is suing to make that requirement disappear — and the case could decide whether AI transparency laws survive anywhere in the United States.
What AB 2013 Actually Demands
AB 2013, formally the Generative Artificial Intelligence: Training Data Transparency Act, was signed in September 2024 and took effect on January 1, 2026. It is codified at California Civil Code Section 3111, and its core demand is deceptively simple: any developer that makes a generative AI system publicly available to Californians must post a "high-level summary" of the datasets used to train it.
That summary isn't a single sentence. The statute spells out 12 categories developers must address, including:
- The sources or owners of the training datasets
- Whether the data is protected by copyright, trademark, or patent — or sits in the public domain
- Whether it was purchased or licensed, and whether it contains personal information
- Whether synthetic data was used in development
- When the data was collected and first used to build the system
The retroactive reach is the part that stings. The disclosure obligation covers any system "made available to the public since January 2022" — sweeping in models trained long before the law existed.
The law doesn't ask which copyrighted works you used. It asks you to admit, on the record, that you used copyrighted work at all.
Why xAI Went to Court
Elon Musk's xAI — the company behind the chatbot Grok and the platform formerly known as Twitter — filed suit against California Attorney General Rob Bonta in the Central District of California at the end of 2025, seeking to block enforcement entirely.
Its argument rests on three constitutional pillars:
- Fifth Amendment (Takings Clause): xAI says its curated datasets are trade secrets. In the complaint's words, "these datasets are valuable precisely because they are not public." Forcing disclosure, it argues, is an uncompensated taking of private property.
- First Amendment: The law allegedly compels speech — forcing xAI to say things it would rather not.
- Fourteenth Amendment: Due-process and vagueness claims round out the filing.
xAI's framing is blunt. It calls AB 2013 "a trade-secrets-destroying disclosure regime" that hands competitors a peek at its proprietary recipe.
There's a real point buried in the rhetoric. The selection, cleaning, and weighting of training data genuinely is much of the competitive moat in modern AI. The statute never defines how detailed a "high-level summary" must be, and no official guidance distinguishes a compliant disclosure from one that leaks the secret sauce.
The First Round Went to California
In a notable early signal, U.S. District Judge Jesus Bernal denied xAI's motion for a preliminary injunction in March 2026, finding the company hadn't shown its case was likely to succeed — and rejecting both the trade-secret and free-speech arguments outright. The law stays in force while the fight continues, and xAI has appealed to the Ninth Circuit.
The stakes ripple well past one company. Transparency mandates are currently the backbone of state-level AI regulation, precisely because the political appetite for harder rules has cooled. If the Ninth Circuit accepts the trade-secrets-as-property theory, it could defang transparency laws nationwide — not just in California.
How the Big Labs Are Quietly Complying
Here's the twist that undercuts xAI's "impossible to comply" narrative: OpenAI and Anthropic already did it.
Both posted training-data disclosures referencing California Civil Code Section 3111. But read them and you'll find almost nothing proprietary. They describe generalized categories — publicly available web content, licensed third-party data, user data with opt-outs, synthetic data — without naming a single specific dataset.
- OpenAI disclosed only that its data includes "data that may be protected by copyright" alongside public-domain material.
- Anthropic went further on format, using a structured per-item breakdown, and noted it gathers public content via a general-purpose web crawler — calling the practice "standard industry practice."
The lesson: the law appears satisfiable with vague, lawyer-vetted summaries that protect trade secrets. That makes xAI's absolutist stance look less like a compliance impossibility and more like a strategic test case.
A Coordinated Assault on State AI Rules
xAI isn't fighting alone, and California isn't its only target. In April 2026, xAI filed a parallel suit against Colorado's attorney general over that state's AI Act. And in January 2026, a federal AI Litigation Task Force was established to identify and challenge state AI laws seen as conflicting with a lighter-touch national framework — explicitly including disclosure rules argued to violate the First Amendment.
The pattern is unmistakable. The battle over AB 2013 is one front in a broader campaign to clear away the patchwork of state AI regulation before it hardens into a national norm.
The Bottom Line
AB 2013 is a modest law — it asks for a summary, not a confession — but xAI has turned it into a constitutional referendum on whether AI companies can be made to explain themselves at all. The fact that OpenAI and Anthropic complied without bleeding trade secrets suggests the requirement is workable; the fact that xAI is fighting anyway suggests the real target is the principle of disclosure. Watch the Ninth Circuit. Its ruling will tell us whether "show your data" survives as a regulatory tool, or becomes the first casualty in the rollback of AI transparency.
