xAI sues California over AB 2013: trade secrets, compelled speech, and a fast-approaching deadline
Elon Musk's xAI filed a pre-enforcement challenge on December 29, 2025 against California's AB 2013, set to take effect January 1, 2026. The company argues the law forces disclosure of proprietary training datasets, violating the Takings Clause, compelling speech in breach of the First Amendment, and failing due process due to vagueness.
The case targets California Attorney General Rob Bonta in the Central District of California. The ask: declaratory relief plus preliminary and permanent injunctions that would stop enforcement before disclosures become mandatory.
What AB 2013 compels
AB 2013 requires any developer of a generative AI system made available to Californians since January 1, 2022 to post public documentation about the data used to train the system. "High-level" summaries are required, plus specific details across a dozen categories.
Those include dataset sources or owners, number of data points, data types, time periods collected, first-use dates, licensing or purchase status, use of synthetic data, existence of personal or aggregate consumer information, IP status, and any cleaning or processing. The statute carves out three narrow exemptions: models for security and integrity, aircraft operation in national airspace, and national security models provided only to a federal entity.
xAI says the law never defines "datasets" or "data point," and leaves "high-level" undefined. That ambiguity matters: can a developer say "the public internet," or must it list "state and federal court websites" or "Library of Congress" specifically?
xAI's theories of harm
Trade secret takings. xAI argues the value of its training data strategy comes from secrecy-what sources were used, how much, what types, and how they were integrated. Forcing disclosure destroys that value and appropriates the right to exclude, triggering a per se taking.
The Supreme Court has recognized trade secrets as property for Takings purposes. See Ruckelshaus v. Monsanto Co. 467 U.S. 986 (1984). xAI leans on that line, plus recent discovery orders in AI cases treating training data as "Highly Confidential - Attorneys' Eyes Only."
Compelled speech. xAI frames AB 2013 as a content-based mandate that should face strict scrutiny because it forces disclosure about core aspects of its product. California will likely respond that this is a commercial disclosure governed by the more deferential Zauderer standard for "purely factual and uncontroversial" information, while xAI will point to limits reinforced in NIFLA v. Becerra 585 U.S. ___ (2018).
Viewpoint or purpose-based discrimination. The law exempts models with favored purposes (security, aviation, national security). xAI argues the state cannot compel speech from some speakers while shielding others based on the state's view of which ideas or uses are "important enough" to keep secret.
Vagueness. Undefined core terms, uncertain scope, and "high-level" ambiguity create fair notice problems. xAI says the statute leaves developers guessing at what satisfies compliance, inviting arbitrary enforcement.
Why retroactivity is a pressure point
AB 2013 reaches back to AI systems released since January 1, 2022. xAI claims it invested heavily in datasets starting March-April 2023 under settled trade secret protections in federal and California law. The retroactive sweep forces disclosure of information developed before any notice of this regime.
For Takings analysis, that timing supports investment-backed expectations. For First Amendment and vagueness, it raises practical concerns about reconstructing historical dataset disclosures with precision under uncertain standards.
Procedural posture and injunction factors
This is a classic pre-enforcement, facial/as-applied constitutional challenge against a state officer. Expect a motion for preliminary injunction on a compressed timeline given the January 1, 2026 effective date.
- Likelihood of success: turns on whether training data strategy is protected property under Ruckelshaus, whether Zauderer or strict scrutiny applies, and whether the statute's terms can be reasonably construed.
- Irreparable harm: once trade secrets are disclosed, there's no realistic way to claw them back.
- Balance of equities and public interest: the state will argue transparency and bias mitigation; xAI will argue consumer benefit is speculative while competitive harm is concrete.
What AB 2013 means for AI developers right now
Regardless of where you operate, if your system is available to Californians, you're in scope. The law's breadth means even older models may require disclosures, whether or not they're still widely used.
- Map your training data lineage. Identify sources, licensing status, collection windows, processing steps, and synthetic data use. Document what you can prove, under privilege where appropriate.
- Classify trade secrets. Separate what might qualify as a "high-level" summary from details that would reveal unique curation strategies.
- Prepare dual-track plans. One for compliance if no injunction issues; another held in reserve if a court narrows the law (e.g., allows truly high-level summaries or excludes legacy models).
- Tighten internal controls. Update access gating, role-based permissions, time-limited access, logging, and employee confidentiality acknowledgments to reinforce trade secret status.
- Coordinate public messaging. If you publish model cards or output performance tests, align those disclosures with your risk posture on inputs.
Key questions for counsel
- Does required dataset disclosure create a per se taking, or is it a regulatory taking under Penn Central?
- Will the court apply Zauderer's "factual and uncontroversial" standard or strict scrutiny to these disclosures?
- Are the exemptions severable, and could a narrowing construction cure viewpoint or purpose-based concerns?
- Can "high-level" be satisfied without revealing competitive strategy (e.g., by category rather than named sources)?
- How will retroactivity weigh on investment-backed expectations and fair notice?
Market and regulatory context
xAI highlights the cost and speed of model releases (Grok-1 through Grok-4 between 2023-2025) to argue that disclosure hands rivals a playbook. Parallel pressure is coming from privacy and consumer protection regulators in the U.S. and EU, including the Irish DPC's 2025 inquiry into Grok training.
California is moving on multiple fronts, including CCPA updates effective January 1, 2026, and sector-specific rules for healthcare and chatbot disclosures. Expect copycat bills in other states if AB 2013 survives court challenge.
What happens next
Watch for the preliminary injunction briefing schedule, declarations on trade secret measures, and any attempt by the state to issue guidance that narrows "high-level" to blunt the vagueness claim. Also watch severability arguments-courts could preserve a core disclosure duty while striking exemptions or specific categories.
If an injunction issues, companies may get breathing room but should not assume a final win. If it's denied, prioritize a compliant, least-revealing summary that satisfies each statutory category without naming unique, non-obvious sources where a categorical description could plausibly suffice.
Timeline
- January 2024: AB 2013 introduced in the California Legislature.
- September 28, 2024: Governor signs AB 2013.
- March-April 2023: xAI begins dataset acquisition for Grok.
- November 2023-July 2025: Grok-1 through Grok-4 released.
- April 11, 2025: Irish DPC opens inquiry into Grok LLM training.
- August 2025: xAI files antitrust suit against Apple and OpenAI; sues former engineer for trade secret theft.
- December 29, 2025: xAI sues AG Bonta in C.D. Cal.
- January 1, 2026: AB 2013 slated to take effect.
Bottom line
AB 2013 is the sharpest test yet of how far a state can push AI training transparency without crossing constitutional lines. The court's handling of trade secrets as property and the standard for compelled disclosures will set the tone for 2026.
If you advise AI developers, build your disclosure matrix now, draft a high-level summary that avoids tipping strategy, and be ready to pivot based on the first injunction order.
Your membership also unlocks:
