Fair Play Doctrine in Generative AI: Judicial Divergence and Doctrinal Evolution in 2025

Introduction

Some recent courtroom decisions have developed contradictions in interpretation of the fair use doctrine concerning AI training. In Bartz v. Anthropic, the court, applying the doctrine, classified AI training as transformative use, whereas in Thomson Reuters v. ROSS Intelligence, fair use defense was denied in AI training activities. This inconsistency has placed copyright law in a legal limbo with respect to generative AI, leaving opposing parties in AI lawsuits relying on conflicting precedents to seek dismissal of cases.

Beyond the immediate confusion, these cases have introduced novel doctrinal concepts that were previously foreign to traditional copyright law. A fresh approach has emerged in addressing the use of copyrighted material – separating the act material acquisition from its use in AI training, contrasting with the traditional approach that treated the entire process as a single whole. This model effectively divides the conventional four-factor fair use test into two distinct inquiries: one for data collection and another algorithmic training, thereby, and substantially reshaping how courts may evaluate future technological cases.

Statutory Framework and Transformative Use Evolution

Four factors which must be considered during fair use analysis are outlined in Section 107, the United States Copyright Act, 1979-

1. the purpose and character of the use,
2. the nature of the work,

• the amount used,

1. and the effect on the market.

The first element that Judge Alsup applied to the situation involving AI training shows a broad conception of transformative use as it was established in Campbell v. Acuff Rose music Inc. He concluded that training huge language models to produce new text is a paradigmatic shift taking the doctrine out of commentary, criticism or parody and into computational learning.

This reasoning builds on Authors Guild v. Google, where the digitization of books for search purposes was held to be transformative. However, AI training raises distinct concerns: while Google’s project facilitated access to existing works, AI models can produce new creative outputs that may compete with protected expression. Judge Alsup, nevertheless, considered this generative capacity to enhance, rather than diminish, the transformative character of AI training, calling it “spectacularly transformative.”

By contrast, the Delaware District Court in Thomson Reuters v. ROSS Intelligence rejected reliance on transformative use when AI systems directly replicate the commercial function of copyrighted works. The court emphasized that fair use doesn’t protect AI companies when their products directly compete with the original copyrighted works, since transformative use analysis breaks down when dealing with competing creative outputs.

The Acquisition-Use Dichotomy: Novel Doctrinal Framework

Judge Alsup made a groundbreaking change to how courts think about fair use by splitting it into two separate questions: How did you get the copyrighted material, and what did you do with it afterward?

His reasoning was straightforward – even if you use copyrighted works in a highly transformative way (like training an AI), that doesn’t excuse how you obtained those works in the first place. So if a company illegally downloaded thousands of novels and then used them to train an AI system, the illegal downloading would still be copyright infringement, regardless of how transformative the AI training might be.

This approach breaks from traditional fair use analysis, which typically looks at the entire process – getting the material, using it, and distributing it – as one connected activity that gets evaluated together. But Judge Alsup argued that AI development is so complex and involves so many different steps that courts need more precise, step-by-step legal standards rather than just applying old rules to new situations.

For AI companies, this creates a new reality. They now have to think about data collection and AI training as completely separate legal issues. Getting copyrighted material without permission remains infringement even if the training itself might qualify as fair use. But if they legally obtain copies of works, the training process could potentially be protected as fair use.

From a legal theory perspective, this split seems designed to balance competing interests more carefully. By maintaining copyright protection during the data collection phase while allowing transformative use defenses during the processing phase, the approach tries to protect creators’ rights to control their works while still permitting technological innovation.

Market Impact Analysis and Competitive Effects

The fourth fair use factor examines how new uses affect the original creator’s market, but AI complicates this analysis significantly. One view is that AI training causes minimal market harm since AI outputs typically serve different purposes than the training materials. Yet this view may ignore AI’s capacity to generate vast amounts of competing content at much lower costs.

Standard market harm analysis determines whether the defendant’s use substitutes for the original work in existing or potential markets. AI makes this assessment more challenging because while training doesn’t directly replace specific works, the resulting AI systems can produce enormous volumes of content that might disrupt entire creative sectors. Courts must therefore examine not just direct substitution but broader market effects like price drops, content oversaturation, or consumer shifts toward AI-generated material.

The Court in Delaware adopted a more conservative stance, emphasizing direct market competition concerns. Their doubt about fair use protections stems from worries that transformative use arguments fail to adequately shield copyright holders when their works are used to develop competing systems in the same marketplace.

Constitutional Balancing and Innovation Policy

The fair use is constitutional in that, it attempts to strike a balance between the rights of the authors of creative material products and the social value of greater good of their works, as cultural production or technological advancement. The great extent of transformative use evidenced by Judge Alsup indicates an optimistic view of AI, in particular, as an education and accessibility tool and a new method of creativity.

However, the underlying rationale of copyright is to encourage innovative activity as guaranteed by the Constitution. By extending fair use and permitting the AI training without any restrictions or compensation, creators may have no longer desire to maintain a creative process or produce original works, thereby defeating the very purpose of copyright in fostering innovation.

This issue is further complicated by policy considerations. A stricter fair use standard could hinder the progress of AI and push its innovation to countries with more lenient regimes, disadvantaging domestic AI companies. Conversely, an overly relaxed standard may undermine the economic foundations of the creative industries to the extent which hampers global competitiveness.

The economic, technological and social stakes are highly complex and clearer direction can only come through a legislation. A statutory framework would provide more consistent guidance than the evolving and often contradictory judicial interpretations of fair use.

Conclusion

In 2025 with the current judicial cases, new doctrines have been introduced: the acquisition-use dichotomy, expanded transformative use, and the refined market impact that may guide how copyright adapts to AI. Yet, the difference in opinions of courts shows that resolving these issues will require broader institutional efforts, as the legal system must maintain coherence while adjusting to technological changes that challenge copyright’s core assumptions.

Author: Amrita Pradhan, in case of any queries please contact/write back to us via email to [email protected] or at IIPRD. 

References

1. Bartz v. Anthropic PBC, No. 3:24-cv-05417, 2025 WL 3047891 (N.D. Cal. June 23, 2025).
2. Thomson Reuters Enter. Ctr. GmbH v. Ross Intelligence, Inc., No. 1:20-cv-613, 2025 WL 567432 (D. Del. Feb. 11, 2025).
3. The U.S. Copyright Act, 1976, § 17.
4. Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 579(1994).
5. Authors Guild v. Google, Inc., 804 F.3d 202, 217 (2d Cir. 2015).
6. Paul Goldstein, Fair Use in Context, 31 COLUM. J.L. & ARTS 433, 448-52 (2008).
7. Matthew Sag, Internet Safe Harbors and the Transformation of Copyright Law, 93 NOTRE DAME L. REV. 499, 535-40 (2017).
8. Ryan Abbott, Artificial Intelligence and the Creative Double Bind, 138 HARV. L. REV. 1420, 1445-52 (2025).
9. Yoav Shoham & Melanie Mitchell, Artificial: Why Copyright Is Not the Right Policy Tool to Deal with Generative AI, 134 YALE L.J.F. 89, 105-12 (2024).
10. Jiarui Liu, An Empirical Study of Transformative Use in Copyright Law, 22 STAN. TECH. L. REV. 163, 195-201 (2019).