California AB 2013 Compliance Guide 2026: AI Training Data Transparency Act
As of January 1, 2026, generative AI developers doing business in California must publicly post what their models were trained on. The disclosure requirement is narrow on paper — but it hands the copyright plaintiffs' bar a roadmap, and the definition of "developer" is broader than most fine-tuning teams assume.
What AB 2013 Actually Requires
California's AI Training Data Transparency Act targets a specific gap: businesses and the public have had no reliable way to learn what data trained a generative AI system they were about to adopt. AB 2013 closes that gap by requiring the developer of a covered generative AI system to post documentation on its website, in a location accessible to the public, describing the datasets used to train the system.
The obligation attaches at release: it covers systems or substantial modifications made available to Californians on or after January 1, 2022, which means developers cannot treat the law as forward-looking only — models already in wide commercial use needed retroactive documentation posted by the January 1, 2026 effective date.
What the Disclosure Has to Cover
Dataset sources and general description
REQUIREDA high-level summary of where the training data came from — scraped web content, licensed third-party datasets, user-submitted data, synthetic data, or a combination — sufficient for a reader to understand the data's origin.
Copyright, licensing, and public-domain status
HIGH RISKWhether the datasets included copyrighted material, and if so, whether that material was licensed, and whether any datasets consisted of public domain or otherwise unprotected content.
Presence of personal information
REQUIREDWhether the training datasets included personal information, which creates overlap with the developer's separate CCPA obligations around personal data used in AI training.
Collection time period and acquisition method
REQUIREDThe general date range during which the data was collected, and whether it was purchased, licensed, or independently collected — details that narrow down exactly which corpora and vendors were involved.
Why This Is a Copyright Litigation Problem, Not Just a Transparency One
AI developers are currently defending training-data copyright suits from publishers, authors, image libraries, and music labels — cases that typically stall on discovery, because plaintiffs cannot prove their specific work was used without first getting the defendant to disclose its training corpus. AB 2013 removes that bottleneck for any California-facing system: the same disclosure meant to inform consumers also tells a rights holder's counsel, in the developer's own words, whether copyrighted material was in scope and roughly when it was collected.
That reframes the compliance question. A developer drafting the AB 2013 disclosure is not just satisfying a transparency statute — it is creating a document that may be quoted directly in the next training-data copyright complaint filed against it, which is why the disclosure language increasingly runs through the same legal review as the litigation defense strategy.
Compliance Checklist Before the Disclosure Goes Live
Confirm whether your fine-tuned model counts as a covered system
If your team fine-tuned or substantially modified a base model for a California-facing product, you may have become a 'developer' under AB 2013 for that modified system, independent of the base model provider's own disclosure.
Route the disclosure draft through legal, not just marketing or PR
Because the disclosure can be used as evidence in copyright litigation, treat the drafting process as a legal filing, not a transparency blog post — precise, defensible language matters more than completeness for its own sake.
Audit retroactively covered systems released since 2022
AB 2013 reaches back to systems released or substantially modified since January 1, 2022 — confirm every currently-marketed generative AI product traces back to a system with a compliant public disclosure, not just newly launched ones.
Keep the disclosure current as training data changes
Retraining or incorporating new datasets into an existing system is itself a substantial modification that can trigger a fresh disclosure obligation — build disclosure updates into the model-release process rather than treating it as a one-time filing.
Frequently Asked Questions
Are there exemptions to AB 2013?
Yes. The law exempts generative AI systems developed solely for security or integrity purposes, such as detecting malicious content or preventing fraud, and systems not made available to Californians. It does not exempt a system simply because it is offered for free or embedded inside another product.
Who enforces AB 2013?
Enforcement runs through the California Attorney General rather than a private right of action, which means individual consumers generally cannot sue directly for a missing disclosure — but the documentation a developer posts to comply can still be cited as evidence in separate copyright or consumer-protection litigation brought by others.
Does AB 2013 require disclosing the specific works or sources used in training?
No — the statute requires a high-level summary of dataset categories, sources, and characteristics, not an itemized list of every document or work included. That said, plaintiffs' counsel routinely uses the high-level disclosure to justify discovery requests seeking the underlying itemized data.
How does AB 2013 relate to the EU AI Act's training data transparency requirements?
Both require some level of training data disclosure, but the EU AI Act's summary obligations for general-purpose AI models are more detailed and apply to a different regulatory track. A developer serving both markets typically needs separate disclosure documents tailored to each statute's specific required content, not a single shared summary.
Find AI Compliance and Legal Documentation Tools on RatedWithAI
RatedWithAI reviews AI governance and compliance platforms — including tools built to track training-data provenance, generate AB 2013 and EU AI Act disclosure documentation, and manage the legal review workflow around model releases.
Explore AI Legal & Compliance Guides