California AI Training Data Transparency (AB 2013)

California AB 2013 mandates transparency for generative AI systems, requiring developers to disclose details about their training data. This law applies to any person, partnership, or corporation that creates, codes, produces, or substantially modifies a generative AI system or service, particularly those made available to California residents.

The law defines 'generative artificial intelligence' as AI capable of producing synthetic content (like text, images, video, audio) that mimics its training data. 'Training data' includes all datasets used for training, retraining, or fine-tuning, even for reinforcement learning. 'Personal information' adopts the broad definition from the California Consumer Privacy Act (CCPA), covering data that identifies or relates to a consumer or household. The law applies prospectively to generative AI systems released after January 1, 2022, meaning current and future models are covered, with disclosures due by January 1, 2026.

The core of AB 2013 requires developers to publicly disclose comprehensive information about the datasets used to train their generative AI models. This includes detailing the sources from which data was obtained, such as internet scraping, data brokers, or licensed content, to provide transparency on data provenance.

Developers must also explicitly state whether their training datasets contain copyrighted, trademarked, or patented materials, addressing intellectual property concerns. Furthermore, the law mandates disclosure of whether training data includes 'personal information' as defined by the CCPA, supporting individual privacy rights. Other required disclosures include the general volume and characteristics of the data, the methods used for data acquisition (e.g., purchases, licensing, web scraping), and any processing, cleaning, or transformation applied to the data. Developers must also specify the temporal scope of data collection and training, and whether synthetic data was used in the training process.

AB 2013 outlines a straightforward implementation approach focused on public accessibility. Developers are required to publish all mandated training data documentation on their publicly accessible websites. The law does not prescribe a specific format, template, or technical standard for these disclosures, offering flexibility in presentation but potentially leading to varied compliance approaches across the industry.

Crucially, these disclosures are not a one-time requirement. Developers must maintain and update the documentation as their training data practices evolve, especially when 'substantial modifications' to AI systems involve new or different training datasets. While the law takes effect on January 1, 2026, it applies retrospectively to generative AI systems released after January 1, 2022. This means developers must gather and document training data information for existing systems that fall within this timeframe, even if initial development did not anticipate such requirements.

AB 2013 does not specify direct penalties for non-compliance, but enforcement will primarily fall under the existing powers of the California Attorney General's Office. Violations may be pursued under California's Unfair Competition Law (UCL), which prohibits unlawful, unfair, or fraudulent business practices. UCL violations can lead to civil penalties of up to $2,500 per violation, injunctive relief (requiring compliance), and restitution to affected parties.

If inadequate or misleading disclosures are deemed false advertising, the False Advertising Law (Business and Professions Code Section 17500) may also apply, carrying similar penalties. Private plaintiffs, such as copyright holders, may also bring UCL claims. Importantly, AB 2013 disclosures could serve as evidence in separate copyright infringement claims if unauthorized use of protected works is discovered. Furthermore, if training data disclosures intersect with personal information processing, violations could trigger enforcement under the California Consumer Privacy Act (CCPA), which includes penalties up to $7,500 per intentional violation. The Attorney General's office is enhancing its capacity with AI experts to support enforcement.

AB 2013 operates within a broader regulatory landscape, both in California and internationally. It complements California's SB 942 (California AI Transparency Act), which focuses on output transparency by requiring detection tools and provenance labels for AI-generated content. Together, these laws create a comprehensive transparency framework for generative AI in California, covering both input (training data) and output.

The law leverages definitions from the California Consumer Privacy Act (CCPA) and California Privacy Rights Act (CPRA) for 'personal information,' indicating an overlap in privacy obligations. Federal copyright law forms the backdrop for intellectual property disclosures. Internationally, AB 2013 aligns with aspects of the European Union AI Act, which also includes training data documentation requirements for general-purpose AI models, particularly those with systemic risks. This convergence may simplify compliance for multinational AI developers. While other jurisdictions like the UK and China have AI regulations, AB 2013's emphasis on public disclosure sets a high standard, potentially influencing global AI governance and requiring developers serving California residents to adapt their practices regardless of their physical location.