California AB 2013 (2024), codified at California Civil Code Section 3111, requires developers of generative AI systems to publicly disclose information about their training data. The raw source for this page combines the text of the law with Anthropic's AB 2013 compliance disclosure for the Claude model family, a public document describing a frontier lab's training data practices.
Disclosure requirements
The law requires developers of generative AI systems to publicly disclose information about training data. The disclosure must cover eight categories:
- Dataset sources and ownership
- Purpose of training data
- Dataset scale
- Types of data points
- Intellectual property status
- Data acquisition methods
- Personal information handling
- Aggregate consumer information
Anthropic's Claude compliance disclosure
Anthropic's AB 2013 disclosure for the Claude model family describes the data sources, scale, intellectual property status, and personal information handling behind that family.
Data source categories. The disclosure groups training data into five types:
- Publicly available internet data — publicly accessible repositories and general-purpose web crawling, which follows robots.txt and does not bypass password protection or CAPTCHA.
- Third-party acquired data — non-public data obtained via commercial agreements.
- Data labeling services and contractors — human preference selection, safety evaluation, and adversarial testing.
- User data — data from Claude users who have not opted out.
- Internally generated data — including synthetic data.
Scale and types. The disclosure characterizes the corpus as "billions of tokens of text data, images, and multimedia content," noting that exact counts vary by model version and training phase. Data types include text (code, prose, math, synthetic), images, and human preference annotations (RLHF), along with safety and adversarial data.
Intellectual property status. The data includes third-party intellectual property, which the disclosure describes as "consistent with standard industry practice," as well as some public domain content. The disclosure states that Anthropic respects robots.txt but otherwise uses publicly accessible material.
Personal information. The disclosure states that training data may incidentally include personal information present on the internet, that Anthropic uses privacy-preserving analysis tools to filter or obfuscate sensitive data, and that post-training techniques minimize personal information in model outputs.
Analysis and connections
AB 2013 takes a transparency-through-disclosure approach to training data governance, addressing data lineage where SB 53 addresses safety and the California CCPA Regulations address privacy. State-level transparency requirements are being layered across these three measures: SB 53 for safety, AB 2013 for data, and CCPA for privacy.
On copyright, the law surfaces but does not resolve whether using third-party intellectual property in training data is lawful. Anthropic's disclosure characterizes the practice as "consistent with standard industry practice," which can be read as treating it as legal while acknowledging legal uncertainty.
On user data, the disclosure states that Anthropic uses data from users who have not opted out, which functions as an opt-out arrangement for future training data. This interacts with the consumer rights provisions of the California CCPA Regulations, which go further with automated decisionmaking technology (ADMT) opt-out rights.
On synthetic data, the disclosure identifies "internally generated" synthetic data as a training source. Synthetic data is a driver of AI Software Progress and typically lacks public disclosure requirements.
The disclosure also bears on open versus closed models: open-weight labs often do not disclose training data, while AB 2013 creates partial parity by requiring disclosure from developers subject to the law. Training data governance is the upstream input to alignment training under approaches such as Constitutional AI and Anthropic's RSP.
Provenance
Source: California Legislature, AB 2013, codified at California Civil Code Section 3111. The raw source also includes Anthropic's compliance disclosure for the Claude model family.