Generative artificial intelligence: training data: copyrighted materials.

Assembly Member Bauer-Kahan's legislation establishes new requirements for generative AI developers to document and disclose their use of copyrighted materials in model training. The bill requires developers to maintain records of copyrighted content used to train their models and create mechanisms for rights holders to request information about the use of their materials.

Under the proposed law, developers must implement website-based systems allowing copyright owners to submit fingerprints of their content and receive confirmation within seven days about whether their materials were used in model training. The fingerprinting system must follow industry standards while preventing reconstruction of the original content. Developers must retain documentation for the duration of the model's use in California plus ten years.

Rights holders can pursue civil actions against non-compliant developers, with penalties of $1,000 per violation or actual damages, whichever is greater, plus potential injunctive relief and attorney fees. The requirements do not apply to developers who make their full training datasets publicly available at no cost. The bill defines covered materials as those registered, preregistered, or indexed with the U.S. Copyright Office, including sound recordings fixed before February 1972.