Overview

On June 30, the Supreme Court denied the government’s application to stay the D.C. Circuit order allowing Shira Perlmutter to remain in office as Register of Copyrights at the U.S. Copyright Office.1 Although the underlying litigation concerns who has the authority to remove the Register of Copyrights, the Court’s denial may have broader copyright and AI policy implications. In particular, Perlmutter’s continued tenure could result in the Copyright Office’s formal publication of Part 3 of its AI report, “Copyright and Artificial Intelligence: Generative AI Training” (the Part 3 Report). If similar to the pre-publication version issued last year, this could formalize the Office’s view of fair use in the AI training context and provide support for plaintiffs (including those in active appellate litigation) seeking to upend those decisions from the 9th Circuit finding in favor of a “fair use” defense.2

On May 9, 2025, the Copyright Office issued a pre-publication version of the Part 3 Report. The Pre-Publication Report addressed the use of copyrighted works in the development and training of generative AI systems; promptly following which the Trump Administration sought to remove Perlmutter from her position.3 While even the final Part 3 Report would not represent the binding law of the land, a formally published version could carry persuasive weight with courts and Congress, particularly while the law governing fair use and generative AI training remains unsettled.4

Key Takeaways From the May 9, 2025 Report

The Pre-Publication Report, which had been issued in response to tens of thousands of comments, provides a comprehensive analysis of the technical, legal, and policy issues raised by the use of copyrighted works to train generative AI models. It also reflects the Copyright Office’s view that a fair use defense in this context is highly fact-specific, with certain commercial uses of expressive copyrighted works presenting weaker fair use arguments (particularly where outputs are substitutive of existing works or other market harm is shown). Below is a recap of our main takeaways from the Pre-Publication Report:

Prima Facie Infringement 

Copyrighted works are involved in multiple aspects of the development and deployment of generative AI systems. The process of assembling training datasets, the making of copies for training, and the potential for models to memorize and reproduce protected expression in a “black box” manner all are identified as acts that, absent a license or defense, generally constitute infringement in the Copyright Office’s view. As the data collection, curation, training, and output generation may involve copying, retention, or reproduction of protected expression in copyrighted works and implicate the copyright owner’s exclusive rights, including the rights of reproduction, distribution, and preparation of derivative works.

Fair Use

The Pre-Publication Report provides an extensive analysis of the fair use doctrine as applied to generative AI training, structured around the four statutory factors in Section 107 of the U.S. Copyright Act of 1976.

While concluding that fair use analysis in the context of generative AI is highly fact-specific, the Copyright Office emphasizes several points under each statutory factor:

    1. Purpose and Character of the Use. The Copyright Office finds that training a generative AI model on a large and diverse dataset may be transformative, supporting fair use, where the model is used for research, analysis, or other non-substitutive purposes. But the analysis remains context-dependent, and use cases that generate outputs substantially similar to copyrighted works, or that compete directly in the same market, are less likely to be considered transformative. The Copyright Office also notes that commerciality remains relevant, including where models trained on ostensibly noncommercial datasets are later deployed commercially.
    2. Nature of the Copyrighted Work. The Copyright Office explains that using highly creative or unpublished works weighs against fair use, while using factual or functional works may weigh in favor of it, and that generative AI models are often trained on a mix of both.
    3. Amount and Substantiality of the Portion Used. The Copyright Office notes that generative AI training often involves copying entire works, which generally weighs against fair use, although such copying may be reasonable where necessary to achieve a sufficiently justified purpose.
    4. Effect on the Market. The Copyright Office places particular emphasis on this factor, identifying market harm as especially significant in the generative AI training context. The Pre-Publication Report points to traditional forms of market harm, including lost sales, lost licensing opportunities, and AI-generated outputs that serve as substitutes for the original works. The Copyright Office also introduces more novel concepts under market harm, including market dilution and stylistic imitation, suggesting that even outputs that are not substantially similar to a specific copyrighted work may harm creators by flooding the market with competing AI-generated works or enabling imitation of an author’s style as a market-substitution concern, even though style alone is not independently protected by copyright. These concepts do not appear in the text of the Copyright Act and have not been clearly adopted in controlling fair use case law. If accepted by courts or Congress, they could significantly weaken fair use defenses for generative AI training.

Licensing and Policy Considerations

The Pre-Publication Report surveyed the current landscape of voluntary licensing of copyrighted works for training and development of generative AI models, noting the emergence of both individual and collective licensing agreements in sectors such as music, news, and images. While individual and collective voluntary licensing may be feasible in certain circumstances, large-scale licensing projects for the range of works needed for certain AI training present significant practical challenges. The Pre-Publication Report discusses potential statutory approaches, including compulsory licensing, extended collective licensing, and opt-out regimes, but it cautions that compulsory licenses are a significant derogation of copyright owners’ rights and should be considered only in cases of clear market failure.

Potential Impact on AI Development and Policy

The continued tenure of Perlmutter, and the publication of the final Part 3 Report, could complicate the Trump Administration’s agenda for accelerated AI development. The Trump Administration has pursued that agenda through Executive Order No. 14179, “Removing Barriers to American Leadership in Artificial Intelligence,” and related AI-focused executive actions aimed at accelerating industry development, including by promoting AI education and workforce development, expanding AI infrastructure and data center build-outs, promoting exports of American AI technology, and limiting regulatory barriers to AI innovation.5 Formal publication of the Part 3 Report could create friction with the administration’s pro-development approach by strengthening the case for licensing, market-based compensation, or other limits on the use of copyrighted works to train generative AI models. Although the final Part 3 Report would not bind courts or directly regulate AI developers, it could carry persuasive weight in litigation and before Congress, particularly because the case law remains limited and unsettled. To date, the limited federal district court decisions addressing fair use in AI training have been narrow, fact-specific, and procedurally nuanced. Because those early decisions do not yet provide a robust or uniform body of law, formal publication of a skeptical view of broad fair-use defenses from the Copyright Office could give copyright owners additional persuasive authority; despite 2025 decisions from the Federal Courts in the Northern District of California in Bartz and Kadrey that deemed the use of copyrighted materials in the training of foundational AI models to be transformative, and thus a “fair use” of the materials. It could also increase pressure on AI developers to pursue licensing or other risk-mitigation strategies rather than rely solely on fair use for unlicensed training.

Those with questions about the potential copyright, AI, or policy implications of the Part 3 Report may contact the authors of this alert.

1 See Blanche v. Perlmutter, No. 25A478 (U.S. June 30, 2026).
2 U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training 1 (Pre-Publication Version, May 2025).
3 The Pre-Publication Report was released in response to congressional inquiries in advance of its finalization.
4 See Bryan Sterba, Matt Savare & Mark P. Kesslen, "US Copyright Office Releases Long-Awaited Report on Generative AI Training and Copyright Law," Lowenstein Sandler LLP (May 12, 2025) (https://www.lowenstein.com/news-insights/publications/client-alerts/us-copyright-office-releases-long-awaited-report-on-generative-ai-training-and-copyright-law).
5 See Exec. Order No. 14277, Advancing Artificial Intelligence Education for American Youth, 90 Fed. Reg. 17519 (Apr. 28, 2025); Exec. Order No. 14318, Accelerating Federal Permitting of Data Center Infrastructure 90 Fed. Reg. 35385 (July 23, 2025); Exec. Order No. 14320, Promoting the Export of the American AI Technology Stack 90 Fed. Reg. 353939 (July 23, 2025).