Arnold Peter
Date: October 3, 2026
Originally published on LinkedIn

Summary: The Third Circuit’s decision in Thomson Reuters v. ROSS Intelligence illustrates how established copyright principles can apply to AI development. The article examines protection for Westlaw’s editorial headnotes, the purpose of ROSS’s competing research platform, and the relevance of training-data licensing markets. Its central distinction is that ROSS was not a generative-AI case: the ruling applies conventional copyright analysis without deciding the broader fair-use questions surrounding generative models.

Sometimes the most consequential thing a court can say about a revolutionary technology is that there is nothing particularly revolutionary about the law needed to govern it. That is essentially what the Third Circuit has now said in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., affirming the district court’s landmark copyright decision in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025). The appellate court’s opinion, initially filed under seal and now publicly released, describes what has been characterized as the first appellate battle over artificial-intelligence training and fair use as, ultimately, an “ordinary copyright case.”

That characterization may prove almost as significant as the result itself. AI may be revolutionary. The technology is developing at extraordinary speed and presents questions of scale, capability and economic disruption that copyright law’s architects could scarcely have imagined. But the Third Circuit’s opinion suggests something I believe we may see repeatedly as AI litigation develops: revolutionary technology does not necessarily require revolutionary legal principles. Sometimes established law evolves perfectly well to meet a new technology.

An “Ordinary Copyright Case” Involving Extraordinary Technology

The underlying facts are important. Thomson Reuters owns Westlaw, which includes judicial opinions that are public law and freely available for anyone to use. But Westlaw also contains substantial original editorial content, including the familiar headnotes created by its lawyer-editors. Those editors exercise judgment in deciding which propositions within an opinion deserve emphasis and how those propositions should be expressed.

ROSS Intelligence was developing an AI-powered legal research platform that would compete with Westlaw. It used thousands of Westlaw headnotes in connection with approximately 25,000 training memos used to develop its system. The district court ultimately held that ROSS infringed 2,243 Westlaw headnotes and rejected its fair-use defense.

The Third Circuit has now affirmed. It concluded that the headnotes possessed sufficient originality for copyright protection because Westlaw’s editors exercised judgment in deciding which legal propositions to highlight and how to phrase them. At the same time, the court rejected the argument that copyright protection would somehow give Thomson Reuters ownership over the law itself. Judicial opinions remain available for everyone to use. What Thomson Reuters could protect was the original editorial work layered on top of those opinions. That distinction seems both straightforward and enormously important.

AI Training Does Not Automatically Transform Copying Into Fair Use

The most consequential portion of the decision may be the court’s fair-use analysis. Relying on the Supreme Court’s 2023 decision in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023), the Third Circuit focused on whether ROSS used the copyrighted material for a meaningfully different purpose.

It concluded that it did not.

Westlaw used its headnotes to help legal researchers locate relevant judicial opinions. ROSS used material derived from those headnotes to develop a competing platform that also helped legal researchers locate relevant judicial opinions. The court therefore characterized ROSS’s use as commercial and “minimally transformative, at best.”

ROSS invoked Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015), where Google’s scanning of books to create a searchable database was held to be highly transformative. But the Third Circuit saw an important distinction. Google’s search tool helped users locate books they might subsequently read or purchase; it was not intended to substitute for the books themselves. ROSS, by contrast, was developing a product intended to compete with Westlaw in performing substantially the same research function.

That distinction should command the attention of anyone following AI copyright litigation. The insertion of an artificial-intelligence system somewhere between copyrighted input and commercial output does not automatically transform the underlying use into fair use. Courts will still ask what the copyrighted material was used for, what the resulting product does and whether the new use serves a genuinely different purpose.

“Ease Is Not a Justification for Copying”

ROSS also relied on another important line of technology cases, including the Supreme Court’s decision in Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021), and Ninth Circuit decisions involving Sega and Sony. Those cases involved intermediate copying of computer code where copying was necessary to obtain access to unprotected functional elements or achieve interoperability.

The Third Circuit found those cases distinguishable for a simple reason: ROSS already had access to the underlying judicial opinions. The law was available. ROSS could have used those opinions to develop its own training materials. It instead used Westlaw’s editorial headnotes because doing so made the process easier. The Third Circuit summarized the distinction in a memorable sentence: “Unlike necessity, ease is not a justification for copying.”

That principle may eventually assume significance far beyond legal research.

AI systems can ingest and process enormous amounts of human-created material at a speed and scale that would have been unimaginable only a few years ago. But technological convenience does not itself answer the copyright question. The ability to acquire, copy, process or train on millions of works efficiently does not necessarily establish a legal right to do so.

The Emerging Market for AI Training Data

The Third Circuit’s treatment of market harm may have even broader implications.

The court concluded that ROSS threatened Westlaw’s existing market because it sought to compete with Westlaw. But importantly, the panel also accepted Thomson Reuters’ argument concerning a developing market for licensing copyrighted content as AI training data.

That is significant because such a market does not necessarily have to be fully mature before copyright law recognizes its potential relevance. Thomson Reuters was already using its headnotes to train its own AI search products, and the court found evidence that a market for licensing headnotes as AI training material was rapidly developing. We are therefore beginning to see recognition of what might be called the training value of human creativity.

Books have training value. Journalism has training value. Music, photography, film, television, legal scholarship, medical literature, software and professional databases may all possess training value because they contain human-created information from which artificial-intelligence systems can derive enormous commercial benefit.

That raises an increasingly important question. Copyright analysis cannot necessarily end by asking whether an AI system ultimately reproduced a copyrighted work verbatim for an end user. Courts may also have to ask whether copyrighted material had independent economic value as training material and whether unauthorized use impaired an actual or reasonably developing licensing market.

What This Decision Does—and Does Not—Say About Generative AI

Here the Third Circuit was particularly careful, and anyone discussing this case should be equally careful. This was not a generative-AI case.

The court expressly distinguished the ROSS technology from generative systems capable of producing new expression. Indeed, it specifically observed that concerns raised by the Department of Justice in the consolidated OpenAI copyright litigation in New York involving generative-AI training “do not apply here.”

That limitation matters enormously. The Third Circuit has not held that training ChatGPT, Claude, Gemini or another generative model on copyrighted works necessarily constitutes infringement. Nor has it resolved the complex fair-use questions presented when an AI system learns from enormous datasets and generates outputs that may bear little or no resemblance to any individual training work. Those cases remain to be decided.

But the Third Circuit has now rejected what could become the opposite extreme: the proposition that putting copyrighted material into an AI development or training process somehow takes the conduct outside conventional copyright analysis. It does not.

Courts can still ask familiar questions: Was the work copyrightable? Was it copied? Why was it copied? Was the use genuinely transformative? How much was taken? Was copying necessary? Did the resulting product compete with the copyright owner? And did the use interfere with an existing or reasonably developing market for the copyrighted material?

Those are not particularly futuristic questions. They are copyright questions.

Revolutionary Technology, Evolutionary Law

This is why I find the Third Circuit’s description of the case as an “ordinary copyright case” so compelling. Artificial intelligence can make conduct appear legally unprecedented simply because the technology accomplishing it is unprecedented. But the underlying legal problem may sometimes be surprisingly familiar.

Imagine a company taking thousands—or millions—of proprietary photographs, songs, articles, books, legal summaries or other copyrighted works and using them to develop a competing commercial product. Long before generative AI, lawyers would have asked who owned the material, what was taken, whether it was protected, whether permission was obtained, why it was copied, whether the use was transformative and whether the resulting product harmed the copyright owner’s market.

The arrival of a neural network between the original work and the finished product complicates those questions enormously. It does not necessarily eliminate them.

The same observation has relevance to the broader concept of misappropriation, although copyright infringement and state-law misappropriation remain distinct legal doctrines. The fundamental concern is recognizable: when does learning from another person’s work become taking another person’s protected work? AI makes the factual and legal analysis vastly more complicated because machines can ingest, analyze and derive value from human creativity on an unprecedented scale. But scale and sophistication do not themselves resolve questions of ownership, permission, transformation or economic harm.

The Human Labor Behind the “Data”

There is also a human dimension that should not disappear behind the technology. We frequently speak about AI “data” as though data simply exists in nature waiting for machines to discover it. Much of the information that gives AI its extraordinary capabilities exists because human beings created it.

Someone wrote the article, composed the song, photographed the image, developed the software, researched the treatise, wrote the screenplay or assembled the database. In this case, lawyers and editors read judicial opinions, identified significant legal propositions and exercised professional judgment in creating the headnotes that made Westlaw more useful.

AI can process that human work at a scale and speed no person could possibly duplicate. That computational power is part of what makes AI revolutionary. But computational efficiency does not necessarily extinguish the legal rights associated with the human creativity that made the information valuable in the first place.

The challenge is not to choose between protecting creators and developing artificial intelligence. AI can expand access to knowledge, accelerate scientific discovery, transform legal and medical research and democratize capabilities previously available only to enormous institutions. Copyright protection and technological innovation therefore should not be treated as mutually exclusive objectives. A functioning licensing market may, in fact, become one of the mechanisms through which human creators and AI developers share the economic value created by this extraordinary technology.

ROSS has announced that it intends to seek review in the U.S. Supreme Court, so the litigation may not be over. Because the Third Circuit decided the copyright issues on interlocutory appeal, proceedings also remain in the district court once appellate review concludes.

Whatever happens next, however, the Third Circuit has given us a remarkably useful framework for thinking about the intersection of copyright and artificial intelligence. We sometimes become so mesmerized by AI’s unprecedented capabilities that we assume every legal question surrounding it must also be unprecedented.

The Third Circuit offers a useful corrective. AI may force courts to refine copyright doctrine and apply it to circumstances its creators could never have envisioned. Generative AI may eventually present harder questions than those presented by ROSS, and courts may draw very different lines in those cases.

But perhaps the larger lesson is simpler: AI is revolutionary. Copyright law is evolutionary. And sometimes the law does not need to be reinvented simply because the technology does.

Frequently Asked Questions About AI Training and Copyright

What was the Westlaw copyright dispute about?

The dispute concerned ROSS Intelligence’s use of material derived from Westlaw’s editorial headnotes to develop a competing AI-powered legal research platform. The article distinguishes those headnotes from the underlying judicial opinions.

Why were Westlaw’s headnotes treated differently from judicial opinions?

The article explains that Westlaw’s editors exercised judgment in selecting legal propositions and expressing them in headnotes. Protection for that original editorial work does not give Thomson Reuters ownership of the underlying law.

Does using copyrighted material for AI training automatically make it fair use?

No. As the article explains, adding AI to a development process does not itself establish fair use. The purpose of the use, its relationship to the original work, and its market effects remain relevant to the analysis.

Why did competition with Westlaw matter?

The article describes ROSS’s platform as serving substantially the same legal research function as Westlaw. That competitive purpose was relevant to the court’s assessment of transformation and market harm.

What does the ruling mean for generative AI?

The ruling addressed ROSS’s non-generative legal research technology. It does not establish that training every generative model on copyrighted works is infringement or resolve the broader fair-use questions involving those systems.

Why are AI training-data licensing markets relevant?

The article explains that copyrighted works may have economic value as training material. An actual or reasonably developing licensing market may therefore matter when assessing whether unauthorized use harms a copyright owner’s market.

Source for the FAQ’s case discussion: Third Circuit opinion in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153.