- AB 2013 requires developers of generative AI systems offered to Californians to post documentation about their training data.
- It applies to systems released or substantially modified on or after January 1, 2022, with documentation due by January 1, 2026.
- Developers must say whether datasets were purchased or licensed and whether they include personal information or copyrighted material.
- Sellers should expect buyers to ask for source descriptions, date ranges and rights records they can disclose.
What the law requires
AB 2013 adds a requirement to California's Civil Code: a developer of a generative AI system or service made available to Californians must post documentation on its website describing the data used to train it. This covers systems released on or after January 1, 2022, including substantial modifications, and the documentation was due by January 1, 2026 (LegiScan).
What the summary must include
The high-level summary covers, among other things:
- The sources or owners of the datasets
- How the datasets further the system's intended purpose
- The number of data points, which may be given as ranges
- The types of data points
- Whether the datasets include data protected by copyright, trademark or patent, or are entirely in the public domain
- Whether the datasets were purchased or licensed by the developer
- Whether they include personal information or aggregate consumer information
- Whether the developer cleaned, processed or modified them
- The time period during which data was collected
- The dates the datasets were first used
- Whether synthetic data was used
There are exemptions, including for systems used solely for security and integrity, for aircraft operation, and for national security or defense available only to federal entities.
What it means for data sellers
Buyers now publish a description of where their data came from. Expect them to ask you for:
- A short description of the data and its source that they can disclose
- Date ranges for collection
- Whether the data contains personal information, and what was removed
- Confirmation of the rights you hold
Agree in the license how you will be named, if at all. Some sellers prefer a generic description ("licensed customer support conversations from a US software company") to being named. A data manifest gives the buyer what it needs without exposing the data itself.
What it means for AI teams
Keep records per dataset from day one: source, license, date range, personal information handling and processing steps. A dataset card and a chain of custody log make the AB 2013 summary a matter of compiling, not reconstructing.
Read next
Questions
Does AB 2013 require naming each data seller?
It requires the sources or owners of the datasets in a high-level summary. How specifically each source is described is a judgment for the developer, so sellers should agree the description in the license.
Does AB 2013 apply to companies outside California?
It applies to developers whose generative AI systems are made available to Californians, wherever the developer is based.