Thomson Reuters v ROSS: What the AI Training Fair Use Ruling Means.

The Third Circuit’s opinion in Thomson Reuters Enterprise Centre GmbH; West Publishing Corp v. ROSS Intelligence Inc., 2026 WL 2932557 (3^rd Cir. Sep. 29, 2026), is an important and first appellate ruling on copyright and AI training. It affirms that ROSS’s copying of Westlaw headnotes to train a competing legal-research service was not fair use. It illustrates that copying a competitor’s copyright works to develop an AI system for a service that competes with the competitor may not be fair use.

But the significance of the ROSS opinion must be understood having regard to the evidentiary record before the court and how the court analyzed the facts in the case. ROSS is not a ruling that all AI training infringes copyright, or that the training of large language models is not fair use It is likely to be referred to in subsequent cases so its actual holdings and the reasoning of the Court is important to understand.

The Third Circuit Court of Appeals rejected the suggestion that because the case involved AI the basic legal inquiry was different.

Under ROSS’s framing, this case appears to concern the future of AI legal technology.  But appearances can be deceiving.  In truth, this is no more than an ordinary copyright case; we must determine whether Thomson Reuters’s materials are sufficiently original for copyright protection and whether ROSS’s use of those materials was fair.

A statement that the ROSS case involves the application of ordinary copyright principles may be true. But considering the 148 or so cases pending in the U.S. related to copyright and AI training the first Circuit Court of Appeals case to examine this issue surely cannot be “an ordinary copyright case”.

Key takeaways from the ROSS case

Before getting into the details of the case, here is an FAQ with key takeaways from the ROSS opinion.

• Did the Court hold that copying copyright works for all AI training is unlawful?

No: it held on the record before it that such copying may be infringing as the fair use defense did not apply.

• Why did the court find that ROSS’s copying was not fair use?

The Court found that ROSS’s copying of Thomson Reuter’s headnotes was for a commercial purpose and was not transformative because the copying was for the purpose of training a machine learning model that would be used to compete with Thomson Reuters in the legal research market. The Court also accepted that ROSS harmed Thomson Reuter’s position in a potential derivative market for licensing headnotes as AI training data.

• Why did the Court not find that ROSS’s intermediate copying of headnotes to train its AI system to be transformative?

The Court did not accept that it was necessary for ROSS to copy the Thomson Reuter’s headnotes. It made this finding even though ROSS’s AI system was able to generate new selections of published opinions based on it’s trained model.

• Will the case have any impact on the generative AI cases now before U.S. District Courts?

Very unlikely. The Court found that copying training data which cannot generate new content as not transformative. It also expressly distinguished the ROSS case from the generative AI fair use cases. The Court’s finding that loss of licensing revenue may be be a cognizable harm may be referred to in future cases, especially in the Third Circuit, but it is not consistent with at least some of the cases in the circuits where the majority of the cases are being tried.

• Is the Opinion in Ross final?

It may not be as ROSS’s lawyers have indicated that ROSS intends to seek cert, or en banc review by the Third Circuit.

Background: what ROSS copied and why

ROSS developed an AI system that answered natural-language legal questions by retrieving existing passages from approximately ten million judicial opinions. To teach the system which passages best answered which questions, ROSS engaged LegalEase Solutions, which, with subcontractor Morae Global, prepared approximately 25,000 training memoranda. Each memo paired a legal question with four to six opinion passages labelled great, good, topical or irrelevant. The questions were derived from Westlaw headnotes; the great answers most frequently came from the passages in legal opinions those headnotes linked to.

This was not simply a project to make judicial opinions searchable in the abstract. ROSS intended to sell a Westlaw substitute, advertised direct comparisons and offered comparable pricing. Some firms switched from Westlaw to ROSS. Those facts became important both to the purpose of the copying and to its market consequences.

The District Court granted partial summary judgment concerning 2,243 headnotes, finding that the Bulk Memo questions were substantially similar to the headnotes because of copying. The Third Circuit affirmed that ruling on an interlocutory appeal.

To fully understand the ruling, it is important to understand what the Third Circuit said about what ROSS did and how it characterized the AI system developed by ROSS.

ROSS…created an AI legal search engine that would respond to plain-language legal questions with relevant passages of text from judicial opinions.  ROSS had access to a bank of approximately ten million uncopyrighted judicial opinions from which it could respond with an excerpt relevant to a user’s query.  ROSS’s AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions.

For ROSS’s AI legal search engine to work, the AI program had to learn what made a judicial opinion responsive to a user’s legal question. To train its AI program to do so, ROSS enlisted a company called LegalEase Solutions (“LegalEase”) to create AI training memoranda.  LegalEase wrote approximately 25,000 legal memos, each of which presented a legal question and four to six judicial opinion passages in response. Memodrafters for LegalEase, and its subcontractor Morae Global, used thousands of Westlaw headnotes to create the memos’ legal questions, as the headnotes provided “an easy way [to] fram[e] questions.”  The memos’ responses to the legal questions needed to include at least one great, good, topical, and irrelevant judicial opinion passage, labeled as   such.  A great passage was completely responsive; a good passage was mostly responsive; a topical passage was somewhat responsive; and an irrelevant passage was not responsive.  The “[g]reat” passages were “[m]ost frequently” the judicial opinion passages that Westlaw linked to the headnote that ROSS used as the memo’s legal question… These memos were then converted into a machine-readable format and used to train the ROSS AI program to identify patterns regarding which judicial opinion passages respond well to legal questions.

The District Court, Thomson Reuters Enterprise Centre GMBH v. Ross Intelligence Inc., 765 F.Supp.3d 382 (D.Del,2025), summarized here, gave an example of a Question, the West Headnote from which it was derived, and the Case Opinion.  The table below is the District Court’s hypothetical illustration based on Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340  (1991), not an actual headnote or question from the record.

The District Court granted partial summary judgment finding substantial similarity between the headnotes and the Bulk Memo questions. The infringement finding concerned the editors’ protected expression in the headnotes, not copyright in the underlying judicial opinions.

The District Court also found that ROSS’s copying was not fair use. In its analysis, it distinguished ROSS’s system from “generative AI” stating:

Ross was using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with Westlaw. It is undisputed that Ross’s AI is not generative AI (AI that writes new content itself). Rather, when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written.

The Third Circuit concluded that the 2,243 nonverbatim headnotes were protected by copyright. That finding was limited to the editors’ expression, not the judicial opinions. It then examined whether ROSS’s use of those headnotes qualified as a fair use and concluded, on the record before it, that it did not.

Fair use rejected by the Court

Section 107 of the U.S. Copyright Act requires consideration of the purpose and character of the use, the nature of the work, the amount and substantiality taken, and market effect. ROSS bore the burden of establishing this affirmative defense. The factors are considered together in light of copyright’s purposes. I summarize the Court’s findings below and then provide comments.

Factor one: the purpose and character of ROSS’s use

In accordance with U.S. Supreme Court precedents, the Court assessed whether the use was commercial and whether it was transformative. It summarized its interpretation of the framework for analyzing the first factor, stating:

In sum, we must balance “the degree of difference [in purpose] . . . against the commercial nature of the use.”  Id. at 532.  And if “an original work and a secondary use share the same or highly similar purposes, and the secondary use is of a commercial nature, the first factor is likely to weigh against fair use, absent some other justification for copying.”

ROSS’s commercial purpose was not contentious. However, the parties disputed whether it was transformative. The Court concluded that the use was minimally transformative, reasoning:

But because commercialism is not “dispositive” of fair use, id. at 531, we must ask whether ROSS’s use of Thomson Reuters’s headnotes is transformative—that is “whether and to what extent” ROSS’s use of the headnotes “has a purpose or character different from” Thomson Reuters’s use.  Id. at 529 (emphasis in original) (quoting Campbell, 510 U.S. at 579).  Thomson Reuters uses its headnotes to help legal researchers find and understand judicial opinions containing points of law related to their searches.  ROSS uses Thomson Reuters’s headnotes to create a legal-research platform to help users find judicial opinion passages responsive to their searches.  True, ROSS did not use the headnotes in precisely the same fashion as Thomson Reuters.  ROSS took an intermediate step of using the headnotes to train an AI program, which arguably presents a slight degree of difference in use.  But the undisputed evidence demonstrates that ROSS used the headnotes to train an AI program for the benefit of its legal-research platform.  So both Thomson Reuters and ROSS use the headnotes to create and optimize a legal-research platform that helps users find responsive legal material. Thus, ROSS’s use of the headnotes shares the same ultimate purpose as Thomson Reuters’s use, making ROSS’s use minimally transformative, at best.

In reaching this conclusion, the Court distinguished the U.S. District Court decisions in Kadrey and Bartz that had found the training of large language models (LLMs) was transformative based on ROSS’s system not generating original expression stating in a footnote:

We note that the United States Department of Justice (“DOJ”) filed a statement of interest relating to generative AI in a pending case in the Southern District of New York.  Statement of Interest of the United States, In re: OpenAI, Inc. Copyright Infringement Litig., No. 1:25-md-3143 (S.D.N.Y. Sept. 1, 2026) (hereinafter “DOJ Statement __”). The concerns raised in that separate case do not apply here.  The DOJ relied on Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007, 1014 (N.D. Cal. 2025) to contend that training an AI large language model, which can “generate original responses,” is a transformative use.  DOJ Statement at 10; see also Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026, 1034 (N.D. Cal. 2025) (exploring whether generative AI programs transformatively use training materials).  Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness.  The DOJ also argued that the large language model training in OpenAI did not result in “substitutive competition.”  DOJ Statement at 13 (citation modified).  Here, ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw.  In any event, the DOJ’s filing in In re: OpenAI demonstrates that the DOJ is tracking these issues and knows how to assert its interests, but the DOJ notably did not do so here.

ROSS had argued that its use was transformative because it had only made intermediate uses of the headnotes. The Court rejected this contention, finding that ROSS did not need to copy Thomson Reuters’s headnotes to access the underlying unprotected information. It distinguished Oracle, Sega and Connectix on necessity:

In each of these cases, computer code was copied to enable users to make their software operable with an existing computer system.  These uses were transformative because, in each case, copying was necessary to access the unprotected functional aspects of computer code.  See Oracle, 593 U.S. at 30– 31 (stating that Google copied Oracle API “only insofar as needed” and explaining that “shared interfaces are necessary for different programs to speak to each other”); Sega, 977 F.2d at 1514 (noting that “no other means of access to the unprotected elements exists”); Connectix, 203 F.3d at 599 (holding that copying was “necessary to permit Connectix to make its non-infringing Virtual Game Station function with PlayStation games”). Here, ROSS does not need to copy Thomson Reuters’s headnotes to access the underlying unprotected information.  ROSS had access to the underlying judicial opinions and could freely copy them to make the memos needed to train its AI.

Given its findings that ROSS’s use was highly commercial and minimally transformative, the Court weighed the first factor against fair use.

Factor two: nature of Thomson Reuters’s headnotes

The second factor slightly favored ROSS. The headnotes were already published and were more factual than fictional: they accurately described uncopyrightable judicial opinions.

Factor three: the amount and substantiality of the portion of the headnotes used

For the third factor, the Court described ROSS as having copied entire headnotes and treated each headnote as an independent copyrighted work.

ROSS argued that it had copied only 0.08% of Thomson Reuters’s 28 million headnotes. The Court reasoned that this database-wide comparison did not answer how much of each protected work ROSS had taken or justify whole-work copying for the minimally transformative purpose it had identified.

Factor four: the effect on the market

The fourth factor in the U.S. fair use analysis concerns “the effect of the use upon the potential market for or value of the copyrighted work.” This analysis generally focuses on the risk of market substitution and whether the defendant’s use of protected expression from a copyrighted work supplants market demand for that work. In ROSS the Court stated the “potential market” includes both the market for the original work and relevant markets for derivative uses of the original work. The Court also included any “public benefits the copying will likely produce.” The Court found against ROSS on each of the above prongs on the record before it.

The parties disputed the relevant original market. ROSS argued that it was the market for headnotes as a standalone product. The Court assumed that narrower market for the sake of argument, but did not accept the inference that the absence of standalone sales meant the absence of harm. It focused instead on the effect of the use on the value of Thomson Reuters’s headnotes, reasoning:

While there may be no standalone market for headnotes, Thomson Reuters touts the headnotes as a feature to attract users to buy a Westlaw subscription.  By copying the headnotes and using them to build its own competing legal-research platform, ROSS appropriated the headnotes’ value for finding and analyzing judicial opinions and diminished their value as a draw for users to Thomson Reuters’s legal-research platform.  So, Video Pipeline supports the conclusion that ROSS’s copying harmed the value of Thomson Reuters’s headnotes.

The Court went on to consider competition in the legal-research platform market. It found that ROSS had copied the headnotes to create a competing platform in the same market as Westlaw.

The Court also accepted Thomson Reuters’s argument that ROSS harmed its position in a potential derivative market for licensing headnotes as AI training data.

Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing.  Additionally, Thomson Reuters presented evidence that it is using its headnotes as training data for its own AI search products. ROSS offers no evidence to disprove this. That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so. Additionally, ROSS’s copying harmed Thomson Reuters in this market.  By copying the headnotes for use as training data without Thomson Reuters’s authorization, ROSS usurped Thomson Reuters’s opportunity to enter that derivative market and license its headnotes for that purpose.  The market harm caused by ROSS’s use weighs against fair use.

The Court weighed the fourth factor in favor of Thomson Reuters, concluding also that ROSS had not established public benefits sufficient to overcome the market harm it had identified.

ROSS claims its copying will benefit the public by increasing access to the law.  But the underlying judicial opinions are freely available, and ROSS’s service was offered at prices comparable to Westlaw.  So it is unclear how ROSS’s copying would greatly increase public access to the law.

Comments on ROSS Opinion

The ROSS opinion is an important U.S. Court of Appeals ruling on how fair use may apply to copying works for training AI systems. Its implications for other cases will depend on the facts of those cases and how other courts will interpret the decision. As ROSS’s lawyers have indicated that ROSS intends to seek cert, or en banc review by the Third Circuit, the ultimate significance of the opinion will also depend on whether it will be reviewed by the Third Circuit or the U.S. Supreme Court.

The ROSS decision does not deal with fair use in the context of training large language models or generative AI in the context of LLMs.  But, it may provide guidance on less transformative machine learning models, particularly where the models are trained on copyright protected data of an intended competitor.

In assessing the impact of the decision, the following matters may be relevant.

The purpose and character of the use by ROSS

The Court described supervised learning from paired inputs and outputs, without describing the detailed architecture of ROSS’s system. The Court only defined ”AI” as referring “to machine learning technology in which a software program analyzes data to learn how to perform a task.  The program is given data which contains examples of inputs paired with outputs.  By analyzing enough data, the program can improve its performance, effectively learning to recognize patterns regarding which outputs are responsive to which inputs.”

As noted above, the District Court distinguished ROSS’s system from generative AI. The Court of Appeals distinguished the LLMs subject to the Bartz and Kadrey cases saying they can generate “original” responses while “ROSS’s AI platform cannot generate original expression”. The nomenclature used by the Court confusingly might be read to suggest that LLMs generate content that is “original” “expression” protected by copyright, something the courts in the U.S. and the Copyright Office do not currently accept. See Barry Sookman, “Understanding the Copyrightability of AI: Insights from the U.S. Copyright Office”.

More importantly, the distinction appears to be that LLMs can generate new text or new strings of text not in its training data whereas the Ross system can only return extracts copied from its (non-Thomson Reuters) training data. The opinion did not consider whether ROSS’s AI generates new selections or arrangements of text from the millions of cases in its data bases. As selection and arrangement of data are also considered as copyrightable expression, could this have been characterized as more than minimally transformative?

To be clear, ROSS’s AI system may have been more transformative than the Court acknowledged. However, unlike generative AI models that are trained to be able to “generalize” their training data to create new text, ROSS’s AI system was still far less capable as it could only generate selections of pre-existing text from its training data. It was trained to generate new selections of pre-existing text, while generative AI systems are trained not to generate verbatim or near verbatim extracts of training data. Based on the Court’s analysis, this aspect of the Court’s opinion will likely have no impact in assessing the first factor in generative AI cases.

But, even if the Court had assessed the ROSS AI system as being somewhat transformative, it is unclear how much this level of “creation” would have moved the dial for ROSS in assessing the transformation consideration, especially here as the purpose of the use was to develop a competing legal research system.

ROSS argued that its use was also transformative because it had only made intermediate uses of the headnotes. This possibility was adverted to by the District Court in the first summary judgment ruling which stated it would be “transformative intermediate copying if Ross’s AI only studied the language patterns in the headnotes to learn how to produce judicial opinion quotes.” Arguably, ROSS was seeking to have its models discover the patterns – the “secret sauce” – used by Thomson Reuters in creating headnotes and links to content in corresponding opinions that would facilitate returning good answers to questions unrelated to the opinions for which the memos were prepared.

The Court rejected this contention finding that ROSS did not need to copy Thomson Reuters’s headnotes to access the underlying unprotected information and otherwise found the copying not to be transformative. The Court did not explain ROSS could have done that.

The finding in the case was highly fact specific. Every fair use case has to examine the facts to determine whether copying is needed or, as stated by the Supreme Court in Warhol, “reasonably necessary to achieve the user’s new purpose”. The opinion on this point is unlikely to have any impact on training generative AI (LLMs) where massive data sets are needed or reasonably required  to train these models.

The effect on the Thomson Reuter’s market

The Court’s discussion of the fourth factor will likely also feature in future cases. In ROSS, the Court identified two distinct market effects. One concerned competition with Westlaw: headnotes helped attract users, and the Court treated ROSS’s service as a substitute for Thomson Reuters’s product. The other concerned a potential derivative market for licensing headnotes as AI-training data.

Bartz made different findings on the effect on the market. Judge Alsup found on the record of that case, no displacement of demand for the authors’ works by the training copies. But, he held that copying books to train LLMs was highly transformative and rejected the claimed loss of training licence revenues, stating that “such a market for that use is not one the Copyright Act entitles Authors to exploit” and referring to Fox News Network, LLC v. TVEyes, Inc., 883 F.3d 169, (2d Cir. 2018), ), Ringgold v. BET, Inc., 126 F.3d 70, (2d Cir. 1997), American Geophysical Union v. Texaco Inc., 802 F.supp.1 (S.D. N.Y 1992), Ringgold v. BET, Inc., 126 F.3d 70, (2d Cir. 1997), Seltzer v. Green Day, Inc., 725 F.3d 1170  (9th Cir. 2013).

Kadrey likewise rejected lost training licence fees as a cognizable market injury for the use with Judge Chhabria referring to Tresóna Multimedia, LLC v. Burbank High School Vocal Music Association, 953 F.3d 638  (9th Cir. 2020) (quoting 4 Melville B. Nimmer & David Nimmer, Nimmer on Copyright § 13.05 (2019)), Bill Graham Archives v. Dorling Kindersley Ltd., 448 F.3d 605, 614-15 (2d Cir. 2006) and stating:

The parties therefore go back and forth at length about whether a market for licensing general trade books exists or is likely to develop…But whether such a market exists or is likely to develop is irrelevant, because this market is not one that the plaintiffs are legally entitled to monopolize.

A later decision, Chicken Soup for the Soul, LLC v. Meta Platforms Inc., 2026 WL 2668437 (N.D. al. 2026) followed the Kadrey case on whether loss licensing opportinities is recoverable. However, in In re Mosaic LLM Litigation 2025 WL 2294910, (N.D.Cal., 2025) anothet District Court expressed the opinion on a disceovery motion that “while the conclusions in Kadrey and Bartz are well-reasoned, the Ninth Circuit has not yet addressed factor four in the context of generative AI. This remains an unsettled area of law.”

Judge Chhabria recognized a different potential harm: market dilution through the production of many competing, non-infringing works. On the facts, this harm was not established on the record before him. Judge Alsup rejected this theory in Bartz holding that “This is not the kind of competitive or creative displacement that concerns the Copyright Act. The Act seeks to advance original works of authorship, not to protect authors against competition.”

The ROSS opinion opens up whether, at least in the Third Circuit, the loss of potential licensing revenues is a cognizable harm. Given the findings of the Court that ROSS’s AI system was intended to compete with Westlaw and was a potential substitute for it, with little public benefit, even if ROSS had won on the licensing revenue point, it may have had little impact on the outcome of the case.