Generative artificial intelligence: training data: copyrighted materials.

Assembly Members Bauer-Kahan and Kalra propose new requirements for developers of generative artificial intelligence models to document and disclose their use of copyrighted materials in training datasets. The legislation establishes a framework for copyright holders to verify whether their registered works have been used to train AI systems available to California residents.

Under the proposed rules, AI developers must maintain records of copyrighted materials used in training and provide an online mechanism for rights holders to submit information requests. When rights holders submit documentation of their copyrighted works, developers must assess within 30 days whether those materials are present in their training datasets and provide a detailed response. The bill requires developers to retain these records for five years beyond the model's period of commercial availability or use in California.

The legislation creates a private right of action allowing copyright holders to seek damages of $1,000 per violation or actual damages, whichever is greater, along with potential injunctive relief and attorney's fees if developers fail to comply with disclosure requirements. Several categories of AI models are exempt, including those trained exclusively on publicly available data, those used solely for noncommercial research, and those trained only on materials owned by the developer. The requirements apply to models released from January 2022 onward, with compliance required by January 2026.