California Bill Will Require Disclosures About Data Used in Training AI Models
On January 1, 2026, California Assembly Bill 2013, the "Generative Artificial Intelligence Training Data Transparency Act" ("AB 2013"), will come into effect, requiring generative artificial intelligence ("GenAI") developers to publicly disclose on their websites information about the data they use to train their artificial intelligence ("AI") models.
Stated Goal of AB 2013
According to the Assembly Floor Analysis, AB 2013 aims to increase the transparency of AI systems and services so consumers can compare competing systems and evaluate their confidence in the system.
Application of AB 2013
The disclosure requirements apply to GenAI systems (not AI systems more broadly) released on or after January 1, 2022, and made available to Californians, whether for free or for a fee. It applies to "Developers" of such systems, defined as those that design, code, produce, or "substantially modify" a GenAI system. AB 2013 outlines a few narrow exemptions: for example, GenAI systems with the sole purpose of ensuring security and integrity, such as systems to detect security incidents. Otherwise, AB 2013 applies to all GenAI systems, not just high-risk systems, unlike Utah's Senate Bill 226, Colorado's AI Act, and the European Union's AI acts.
Summary of Requirements
GenAI developers will be required to disclose a "high-level summary" of the datasets used to develop their GenAI systems. Despite referring to a "high-level summary," AB 2013 lists 12 required disclosures, some of which are more detailed. Some examples include:
- The sources and owners of the data;
- The number of data points;
- The types of data;
- Whether the datasets include data protected by copyrights, trademarks, or patent;
- Whether the datasets include personal information;
- The time period during which data was collected, and if data collection is ongoing; and
- The dates the dataset was first used in development.
Form of Disclosure
Impacted developers must post this disclosure on their website. AB 2013 does not prescribe a format for disclosures, raising concerns in the Assembly Committee on Privacy And Consumer Protection that the disclosures may not achieve the goal of transparency if they are drafted in a highly technical manner.
Next Steps
Impacted developers will need to prepare their disclosures for publication on their websites by January 1, 2026. This will require performing a review of the developer's training datasets, including datasets provided by third parties. Developers should ensure they do not inadvertently reveal trade secrets or other competitive or proprietary information in their disclosures.
