First Federal Appellate Fair Use Ruling on AI Training Rejects Non-Generative AI’s Use of Copyrighted Headnotes

In Thomson Reuters v. ROSS, the Third Circuit recently affirmed the district court’s partial summary judgment for Thomson Reuters, holding that ROSS did not make fair use of Thomson Reuters’ Westlaw headnotes when ROSS used headnote-derived training memoranda to build a competing non-generative AI legal-research platform. 

ROSS created a non-generative AI legal search engine that would respond to plain-language legal questions with relevant passages of text from judicial opinions. ROSS trained this AI tool using legal memos prepared using thousands of Westlaw headnotes, although ROSS never showed the headnotes to its users. ROSS marketed this AI platform as a Westlaw alternative. Thomson Reuters sued ROSS for copyright infringement and tortious interference with contract. The interlocutory appeal followed the U.S. District Court for the District of Delaware’s partial summary judgment for Thomson Reuters on copyright infringement and ROSS’s alleged fair use defense.

“Unlike Necessity, Ease is not a Justification for Copying.”

The Third Circuit held that Thomson Reuters’ headnotes are not mere citations to law only.  Instead, they have enough original and creative content to attain copyright protection. The court then rejected ROSS’s fair use defense. Fair use is an equitable defense to copyright infringement codified at 17 U.S.C. § 107. It permits certain uses for purposes such as criticism, comment, news reporting, teaching, scholarship, or research. Courts weigh four nonexclusive factors when considering this defense:

1. Purpose and character of the use. This factor weighs whether a use is commercial and how transformative the use is. This factor weighed against ROSS. The court found that the use was commercial and minimally transformative.

In response to ROSS’s argument that copying entire works to build a search tool is transformative, the court explained that such copying qualifies only when the tool serves a function different from the original work and may even steer users toward it. ROSS’s platform, by contrast, did what Westlaw already does with its headnotes and, as ROSS admitted, aimed to replace Westlaw. The use was therefore “minimally transformative, at best.” 

In response to ROSS’s argument that intermediate copying is permissible, the court explained that such copying is justified only when it is necessary to reach unprotected elements of a work, such as functional computer code needed to make software compatible with an existing system. ROSS had no such need, because it could have copied the freely available opinions but chose the headnotes because doing so was easier. “Unlike necessity, ease is not a justification for copying.”

2. Nature of the copyrighted work. This factor favored ROSS. The headnotes were published and were more factual than fictional, though they had originality and remained copyrightable.

3. Amount and substantiality of the portion used. This factor weighed against ROSS. The court inquired whether “no more was taken than necessary.” Because ROSS copied the entire text of the 25,000 Westlaw-written headnotes into its memos for a purpose that was considered to be “minimally transformative, at best” and where the underlying judicial opinions were freely available, the court held that ROSS “took more than necessary.” 

4.Effect on the potential market or value. This factor also weighed against ROSS. The copying threatened Westlaw’s market and made the headnotes less valuable as a draw for subscribers. It also usurped a fast-growing market for licensing headnotes as AI training data.

Because the first, third, and fourth factors outweighed the second, the court held that ROSS's use was not fair.

The case now returns to the District of Delaware for trial on the remaining issues and damages. 

Distinguishing Generative AI

The court distinguished generative AI in footnote 7 of its opinion. The court acknowledged the concerns raised by the U.S. Department of Justice in its statement of interest in In re OpenAI, pending in the Southern District of New York, and stated that those concerns did not apply here. Relying on Bartz v. Anthropic (N.D. Cal. 2025), the DOJ had argued that training a large language model, which can “generate original responses,” is a transformative use. The court noted that, unlike the AI models in Bartz and In re OpenAI, ROSS's AI platform cannot generate original expression, and therefore its use is not transformative. 

Although the court did not rule on LLM training per se, the footnote suggests that generative AI developers may fare better on the first factor in analyzing a fair use defense. However, the developers would still face the fourth factor, in which the Third Circuit counted the licensing market against ROSS. That said, this finding was in the context of a use that was at most minimally transformative. Whether a lost license market, which would not exist if there were a strong fair use defense based on the other factors, counts against a highly transformative generative model remains an open question.

This Third Circuit Court of Appeals decision arrives while the generative AI cases mostly remain in the trial courts. Current research does not identify a merits appeal pending in a federal court of appeals on whether generative AI training or AI-generated code infringes copyright:

• Bartz v. Anthropic has settled with final approval.
• In re OpenAI remains in the Southern District of New York MDL, where the generative AI question is at summary judgment before Judge Stein, with the DOJ’s statement of interest on the docket. 
• Andersen v. Stability AI remains in district court, with trial scheduled for April 2027.

Key Takeaways

AI developers. The decision raises risk when protected content is copied to build a competing product, even if users never see it. Developers should take care with new development and consider using unprotected sources where available. Because ROSS’s contractors’ copying was attributed to ROSS, developers should require provenance, licenses, warranties, and indemnities from data vendors and other contractors. 

Content owners. Owners should document both the editorial choices that make their materials original and their own use of that content to train AI. Monitoring uses and building licensing programs will help establish a licensing market. 

Users of AI tools. A tool built on unlicensed content may face litigation that disrupts the service, so users should ask vendors about training-data provenance and seek indemnities. Users also should avoid lending their own subscriptions to AI developers, as ROSS’s law firm investor did in this instance, where the terms of service prohibited it. 

If you have any questions about the issues raised in this alert, please contact the authors or the Womble Bond Dickinson attorneys with whom you normally work.