The U.S. Court of Appeals for the Third Circuit recently affirmed summary judgment in favor of Thomson Reuters in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, holding that (1) Thomson Reuters’ Westlaw headnotes are sufficiently original to receive copyright protection and (2) ROSS Intelligence’s use of those headnotes to train its AI-powered legal research platform was not fair use.
The Sept. 29, 2026, decision is significant because it is the first federal appellate decision to address copyright fair use in the AI-training context. However, the court repeatedly emphasized the case should not be viewed as a broad ruling on generative AI training. Instead, the court treated the dispute as a conventional copyright infringement case involving the copying of editorial content to build a directly competing product.
Background
ROSS developed an AI-powered legal research tool that returned relevant judicial opinion passages in response to natural-language legal questions. To train its system, ROSS used copyrighted Westlaw headnotes as the basis for framing legal questions to create research memoranda. The AI ultimately learned relationships between legal questions and responsive judicial opinions.
The court emphasized that ROSS’s system was not generative AI. It did not generate new text or synthesize answers. Rather, it returned passages from existing judicial opinions.
Originality Analysis
The central copyrightability issue was whether Westlaw headnotes possess the minimal creativity necessary for protection under Feist Publications v. Rural Telephone Service. The Third Circuit held they do because the headnote editors exercise creative judgment to determine which legal propositions from an opinion merit a headnote and to decide how to express those propositions while complying with editorial guidelines.
Although the underlying judicial opinions remain public domain and uncopyrightable, the headnotes represent editorial descriptions of those opinions. The court concluded that each headnote exhibited the requisite “creative spark” required by Feist.
Fair Use Defense
The court found three of the four statutory fair-use factors (17 U.S.C. § 107) strongly favored Thomson Reuters.
- Purpose and Character of the Use (Minimally Transformative Commercial Use)
The court found ROSS was building a commercial product intended to compete directly with Westlaw, making its use commercial in nature.
Further, the court found the ultimate purpose of both parties’ uses was substantially the same:
- Thomson Reuters uses headnotes to facilitate legal research.
- ROSS used the headnotes to build a legal research system that would facilitate legal research.
While AI training introduced an intermediate technical step, the court concluded ROSS used the headnotes to create a substitute legal research product serving the same function. Consequently, the use was only “minimally transformative, at best.”
- Nature of the Copyrighted Work
This was the only factor favoring ROSS.
The court acknowledged headnotes are largely factual and functional in nature because they accurately describe legal holdings. As a result, they fall farther from the core of copyright protection than creative literary works.
- Amount and Substantiality
The court found this factor favored Thomson Reuters.
Although ROSS copied only a tiny fraction of the approximately 28 million Westlaw headnotes, it copied entire headnotes and did so because they provided an “easy” method of creating training data.
Importantly, the court stressed that ROSS could have used the underlying opinions themselves, which were freely available and uncopyrighted. Because copying the headnotes was not necessary to access underlying legal information, the court found the extent of copying unjustified.
- Market Harm
The court found substantial market harm. ROSS used the headnotes to build a competing legal research platform, threatening Westlaw’s commercial market. Moreover, the court noted a developing derivative market for licensing copyrighted materials as AI-training data.
Thomson Reuters presented evidence that it was already using headnotes to train its own AI products. The court concluded that unauthorized training uses could impair Thomson Reuters’ ability to exploit or license that market.
Why This Decision Does Not Resolve Generative AI-Training Cases
The most important takeaway for AI developers is what the court did not decide. The Third Circuit expressly distinguished pending generative AI cases involving large language models (“LLMs”), including cases against OpenAI, Anthropic and Meta.
The court highlighted several distinctions:
- ROSS Was Not Generative AI
The opinion repeatedly notes ROSS’s technology did not create new expression.
Instead, it retrieved and returned existing judicial opinions.
By contrast, LLMs generate novel responses and synthesized outputs. The court cited decisions such as Bartz v. Anthropic and Kadrey v. Meta in discussing whether LLM training may be transformative because the resulting systems generate new expression.
- Direct Competition Was Central
The court viewed ROSS as a commercial substitute for Westlaw.
Its analysis repeatedly focused on the fact ROSS used the copied material to compete in the same market and serve the same function.
Many generative AI defendants argue their systems serve different purposes than the books, news articles, photographs or other works used during training. Whether courts accept that argument remains unresolved.
- Availability of Non-Copyrighted Source Material Mattered
Another recurring theme was that ROSS had unrestricted access to the underlying judicial opinions and therefore did not need to copy Westlaw’s copyrighted editorial content.
That factual circumstance differs from many generative AI disputes, where the copyrighted works themselves are the source material from which the model learns patterns of language and expression.
Key Takeaways
- Editorial summaries, annotations, classifications and similar value-added content may satisfy the low originality threshold for copyright protection even when based on public-domain materials.
- Using copyrighted content as AI training data is not automatically transformative simply because machine learning is involved.
- Courts are likely to scrutinize whether the resulting AI output competes with the copyright owner’s product and serves the same commercial purpose.
- The existence of a potential market for licensing training data may weigh heavily against fair use.
- The court did not issue a broad ruling that training AI on copyrighted works is always infringing.
For copyright owners, the decision strengthens arguments that curated databases, annotations and other editorial enhancements remain protectable notwithstanding the public-domain status of the underlying materials. For AI developers, the opinion suggests fair-use defenses are weaker where copyrighted material is used to create a directly competing substitute product, particularly when the copied content itself embodies meaningful human editorial judgment.
Harris Beach Murtha’s Intellectual Property Practice Group closely watches emerging intellectual property trends. If you have questions or need assistance regarding enforcement of intellectual property rights, please contact attorney Andy Corea at (203) 772-7739 and acorea@harrisbeachmurtha.com or the Harris Beach Murtha attorney with whom you most frequently work.
This alert is not a substitute for advice of counsel on specific legal issues.
Harris Beach Murtha’s lawyers and consultants practice from offices throughout Connecticut in Bantam, Hartford, New Haven and Stamford; New York State in Albany, Binghamton, Buffalo, Ithaca, New York City, Niagara Falls, Rochester, Saratoga Springs, Syracuse, Long Island and White Plains; as well as in Boston, Massachusetts; Providence, Rhode Island; and Newark, New Jersey.