TCAI, SAG-AFTRA back AI Copyright Transparency Act in Sacramento

Actress and SAG-AFTRA Secretary-Treasurer Joely Fisher testifies in favor of AB 412 before the Assembly’s Privacy and Consumer Protection Committee on March 18.

March 18, 2025 — California’s groundbreaking AI Copyright Transparency Act received its first hearing in Sacramento this afternoon, with Transparency Coalition Co-founder Jai Jaisimha testifying as an expert witness in favor of the bill. Joely Fisher, the actress and SAG-AFTRA secretary-treasurer, joined him in speaking to the need for copyright transparency in AI before the Assembly Privacy and Consumer Protection Committee.

AB 412, the AI Copyright Transparency Act, is aimed at increasing transparency around the use of copyrighted materials to train generative artificial intelligence (GenAI) systems and models. The bill was authored by Assemblymember Rebecca Bauer-Kahan (D-Orinda), one of the leading AI policymakers in Sacramento.

Assemblymember Rebecca Bauer-Kahan testifies on behalf of her AI copyright protection bill on Tuesday.

disclosing the use of copyrighted material

The proposed legislation requires GenAI developers to disclose the presence of copyrighted material to its owner when that material is included in GenAI training datasets.

“This is an incredibly simple bill that allows copyright owners to know when their copyrighted materials are used to train generative AI,” Assemblymember Bauer-Kahan said while introducing the bill before the committee on Tuesday afternoon.

building on last year’s ab 2013

Last year the California legislature passed AB 2013, the AI Training Data Transparency Act, which will require AI developers to disclose the presence of copyrighted material in their training datasets starting on Jan. 1, 2026. AB 412, Bauer-Kahan said on Tuesday, “takes it to the next level” by allowing copyright holders to ask: “If an AI system’s training data contains copyrighted material, is that material mine?”

The bill includes one of the first requirements for developers to install a Training Data Verification Request (TDVR) tool, which would allow individuals to query the developer to find out if specific intellectual property was used to train a certain AI model. TCAI has called for the adoption of TDVRs and accompanying Training Data Deletion Request (TDDR) tools as a required industry standard.

SAG-AFTRA backing the bill

SAG-AFTRA Secretary-Treasurer Joely Fisher spoke out in favor of the bill.

Actress Joely Fisher, Secretary-Treasurer of SAG-AFTRA, appeared on behalf of the actors union to urge adoption of the bill.

“We want to make sure AI improves human lives rather than harming or exploiting us,” she told the committee. “Everything generated by AI originates from a human creative source. AI can’t do anything on its own. No AI algorithm is able to make something out of nothing. So if intellectual property and copyrighted materials are being used to train AI models, the copyright owners need to know.”

“Nothing I’m saying here is particularly radical or controversial,” Fisher added. “Under our current laws, people who illegally distribute copyrighted materials for their own financial gain can face up to five years in prison and a $250,000 fine. Why, then, would we allow these big AI companies to mine others’ copyrighted works without at least giving copyright owners a right to know if their works were mined?”

Jai Jaisimha: This is technically feasible

Image of Jai Jaisimha testifying before a legislative committee in Sacramento.

Transparency Coalition Co-founder Jai Jaisimha testifies in favor of the California AI Copyright Transparency Act before the Assembly’s Privacy and Consumer Protection Committee.

Transparency Coalition Co-founder Jai Jaisimha followed Fisher’s testimony by offering his technical experience regarding the workability of the bill’s requirements for disclosure.

Opponents of the bill have claimed that it may be technically infeasible to implement, “but as someone who has extensive experience with managing the development of similar systems in the past,” Jaisimha said, “I can lay these concerns to rest. It is rudimentary data engineering practice to keep track of content elements and where they came from.”

Generative AI models, Jaisimha said, are largely built using the same large datasets. The difference in the models lies in the ingenuity employed in architecting and training them.

“One such common dataset is called C4,” Jaisimha said. “The Allen Institute of Artificial Intelligence has published research showing that a significant percentage of this dataset includes copyright protected content from reputed publishers. The New York Times and the News Media Alliance have both found evidence that Generative AI models can reproduce or regurgitate copyrighted content verbatim.”

“Rather than taking the responsible step of purging their training data of unauthorized content, model developers have been playing a game of whack-a-mole or hide-the-evidence by using techniques such as adversarial AI or prompt filtering to eliminate AI outputs that are copies of copyrighted content. Meanwhile they’re still using this high-quality content to form the commercial backbone of their products.”

committee approves bill

At the close of testimony, the Privacy and Consumer Protection Committee voted to approve the bill, 9-2.

learn more about copyright disclosure

Previous
Previous

TCAI Guide: How to stop your images and data from being used to train AI

Next
Next

The revised guide to TRAIGA 2.0, the Texas Responsible AI Governance Act