The Report begins by detailing the technical elements of training and generative AI, particularly how developers utilize weights to set parameters for models, the different training phases, and the implications of different tactics of model deployment. The Report explains potential copyright implications at various phases of generative AI development, explicitly stating that the infringement analysis is not impacted by the fact that the data may be discarded after training. The bulk of the Report focuses on how the fair use doctrine applies to developers utilizing third-party copyrighted materials to train AI models, first discussing a few relevant recent cases.
The Report noted Kadrey v. Meta Platforms, Inc., in which the complaint alleged that the models themselves were infringing derivative works but failed to allege that the models were actually replicating the copyrighted materials or producing derivative works. No. 23-cv-3417, 2023 WL 8039640, at *1 (N.D. Cal. Nov. 20, 2023). The court described allegations that the models themselves were infringing derivative works as “nonsensical.” Id. In contrast, the Office agreed with the distinction drawn in Andersen v. Stability AI, in which the court denied a motion to dismiss filed by a third party who downloaded and used an already-trained model, because evidence sufficiently demonstrated that copies or protected elements of copyrighted material remained within the model. 744 F.Supp. 3d 956, 982–84 (N.D. Cal. 2024). The Report explains “[w]hether a model’s weights implicate the reproduction or derivative work rights turns on whether the model has retained or memorized substantial protectable expression from the work(s) at issue.”
Fair Use of Copyrighted Material in Generative AI Training
The Report defines the fair use doctrine as an affirmative defense which requires courts to balance authors’ exclusive rights to their works with the importance of enabling others to build on those works. Looking at the equitable factors of the fair use doctrine in the context of generative AI training, the Report provides the following guidance:
Purpose and Character:
• Training generative AI on third-party copyrighted materials will likely be transformative, unless the output is substantially similar to the original work in question such that the generated work could compete with the original.
• The commerciality analysis should turn on whether the specific use in question serves a commercial or nonprofit purpose – not on the status of the entity.
• Known use of a dataset unlawfully accessed weighs against fair use.
Nature of the Work:
• Use of more expressive original works, or previously unpublished works weighs against fair use.
Amount and Substantiality:
• Wholesale taking of copyrightable material (downloading and curating) also weighs against fair use.
• Use of a entire work may be practically necessary for some forms of training and may be reasonable to the extent a transformative purpose exists.
Market Effect:
• Generative AI may result in lost sales, market dilution, and lost licensing opportunities for individuals holding copyrights of training materials.
• The public benefit weighs in favor of fair use but does not support the unlicensed use of copyrighted materials where licensing options exist or are likely to be feasible.
The Office provided an overview of international approaches to generative AI and copyright, noting that other countries generally provide some type of exception to the protections of copyright law for text and data mining. Without taking a position, the Report noted that the Office will continue to monitor developments in international policy on generative AI and copyright.
Turning to the practical solutions to these concerns, the Report recognized the potential merits of some level of government intervention in the licensing space but concluded that the best current solution is to encourage the market to develop creative, unique licensing solutions. The Office will publish the final version of Part 3 in the near future.
The day after the Office issued this pre-publication report, the Trump administration terminated the Register of Copyright’s employment. The former Register of Copyrights, Ms. Shira Perlmutter, filed suit in the United States District Court for the District of Columbia, against the Executive Office of the President and several individuals in their official capacities.
The Register of Copyrights is appointed by the Librarian of Congress. Prior to Ms. Perlmutter’s termination, President Trump terminated the Librarian of Congress and appointed the deputy attorney general of the United States, Mr. Todd Blanche, as the temporary Librarian of Congress, the office responsible for appointing the Register of Copyrights. Ms. Perlmutter argued that because Congress has only authorized the President to fill high-level vacancies in executive agencies, the President does not have the power to appoint the head of the Library of Congress, which is an agency designated as part of the legislative branch under 2 U.S.C. § 171(1).
Ms. Perlmutter sought a Temporary Restraining Order enjoining defendants from removing her from her position, and District Judge Timothy J. Kelly denied that motion on May 28, 2025. Ms. Perlmutter is now seeking a preliminary injunction.
