The California AI training data law has moved from a 2024 bill to a disclosure requirement that is now in effect. Developers of covered generative AI systems must publish a high-level account of the data used to train them. The rule does not ask companies to post every training file or release a downloadable copy of a dataset.
The change matters because it gives users, creators and researchers a common set of questions to ask about a model’s inputs: where the data came from, what kinds of material it contains, how it was processed and when it was first used. The answers may still be broad, but they can make different companies’ disclosures easier to compare.
What the California AI training data law asks companies to disclose
California Civil Code Section 3111 says that covered developers must post documentation on their own websites. The first deadline was Jan. 1, 2026; after that, the documentation must be posted before each covered system or substantial modification is made publicly available to Californians. The law applies to generative AI systems or services released on or after Jan. 1, 2022, and it applies whether access is free or paid.
The California AI training data law calls for what the statute describes as a “high-level summary of the datasets used.” According to the current text of Civil Code Section 3111, that summary must cover dataset sources or owners, how datasets support the system’s intended purpose, the number and types of data points, whether protected intellectual property or public-domain content is included, and whether data was purchased or licensed.
The list also reaches beyond copyright. Developers must say whether datasets include personal information or aggregate consumer information, describe cleaning or other processing, give the period when data was collected and whether collection continues, state when datasets were first used in development, and disclose whether synthetic data was used. Counts can be given in ranges, and estimates are allowed for datasets that change over time.
That structure can reveal the shape of a training pipeline without exposing its full contents. A reader might learn that a developer used public and partner-provided data, that a dataset contains several media types, or that collection is ongoing. The disclosure is not necessarily specific enough to tell an individual which exact page or image appeared in a model’s training material.
One company’s disclosure shows the level of detail
OpenAI’s public training-data summary offers a current example. The company says it uses publicly available data, material accessed through third-party partnerships, information provided or generated by users, human trainers and researchers, and synthetic data. It describes datasets containing text, images, audio and audiovisual material.
OpenAI also says its data may include copyrighted material and personal or aggregate consumer information, while noting steps it takes to reduce personal information. Its summary says data collection began around 2018 and continues, and that the datasets were first used for system development in 2021. Those are the company’s own statements under the law, not an independent audit of every dataset or a separate finding about whether each source was lawfully used.
The example makes a useful distinction: a disclosure can be informative while leaving important details unresolved. A creator can see broad source categories, collection dates and the company’s description of processing, but not necessarily trace a particular work into a particular model. The California AI training data law is designed to make documentation visible; it does not turn the disclosure page into a record-level search tool.
It is also a different kind of transparency from content provenance. GenZNewz’s C2PA explainer describes credentials attached to a media file to show its origin or editing history. California’s rule instead asks for information about the datasets used during model development. One addresses a particular piece of content; the other describes the material behind a system.
What the rule does not settle
A training-data summary is not the same as releasing raw data, and the statute does not say that a developer must publish each training example. It requires a high-level account with specified categories. That limit can preserve some confidentiality, but it also means a short or general disclosure may leave readers with follow-up questions about individual works or collection practices.
The law also does not decide copyright disputes simply by requiring a company to say whether protected material appears in a dataset. A disclosure that material may be copyrighted does not, by itself, prove that a particular use was permitted or prohibited. Those are separate questions; this article explains the disclosure rule and is not legal advice.
California’s current code lists exceptions for generative systems whose sole purpose is security and integrity, systems whose sole purpose is operating aircraft in the national airspace, and certain national-security, military or defense systems made available only to a federal entity. The law is therefore not a blanket filing rule for every AI feature, chatbot or developer.
What readers can watch for next
The original measure, Assembly Bill 2013, was approved in 2024; the current code records a later amendment that took effect in 2026. The timeline is important: the first publication date has passed, and future disclosures are tied to public releases and substantial modifications rather than a single one-time website notice.
For readers comparing products, the most useful questions are practical: Does the disclosure name source categories? Does it give a collection period and first-use dates? Does it explain whether data was licensed, processed or synthetically generated? Are personal information and intellectual-property status addressed? A page that answers these clearly can be more useful than a longer statement that leaves the categories vague.
The California AI training data law creates a public starting point, not a complete map of how a model learned. Its value will depend on how clearly developers answer the required questions and whether those pages are updated as systems change. Readers can follow related coverage in GenZNewz’s AI News section as disclosure pages and systems change. That makes the first wave of disclosures worth reading closely, while keeping in view what the documents can—and cannot—prove.
Comments 0
No comments yet. Be the first to share your thoughts!
Leave a comment
Share your thoughts. Your email will not be published.