‘This story begins with our country’s founding’ writes Third Circuit Judge Montgomery-Reeves in Thomson Reuters v. ROSS Intelligence. Referring to the much older history of American commercial legal publishing, Montgomery-Reeves frames this dispute as ‘no more than an ordinary copyright case’ that involves issues surrounding originality and fair use. Technology might have changed, but it would seem that an AI-powered legal search engine (ROSS AI Search), trained on material derived from Westlaw headnotes, confronts a similar challenge: balancing access to information against a publisher’s right to control future uses of reporting.
Facts
Thomson Reuters commenced copyright infringement proceedings against ROSS Intelligence in May 2020. It argued that ROSS AI Search could retrieve relevant case law when asked plain-language legal questions because it was trained on material derived from Westlaw headnotes owned by Thomson Reuters. After Thomson Reuters denied ROSS a Westlaw license, ROSS turned to a third-party legal-research company, LegalEase, to create about 25,000 question-and-answer-sets that involved legal questions ‘a lawyer would ask’. ROSS subsequently converted them into usable machine-learning training data to train ROSS AI Search.
While Thomson Reuters argued that these sets were essentially Westlaw headnotes, ROSS maintained that the lawyers drafted, not copied, them. In February 2025, the District Court granted partial summary judgment for Thomson Reuters and held that ROSS’ use of 2,243 headnotes infringed Thomson Reuters’ copyright and was not fair use. A few months later, District Circuit Judge Bibas granted ROSS’ motion for interlocutory review on originality and fair use.
Originality
Judge Montgomery-Reeves found that the Westlaw headnotes had ‘some creative spark’ pursuant to Feist as Thomson Reuters’ editors ‘made creative decisions’ (e.g. significant points of law to include, compliance with Westlaw’s drafting criteria). Although written opinions are excluded from protectable copyright subject matter (Wheaton v. Peters), case reporting that is the result of ‘intellectual labour’ such as headnotes can be protected (e.g. Callaghan v. Myers and much later, Georgia v. Public.Resources.Org). In response, ROSS unsuccessfully argued that if the headnotes were protected by copyright it would result in a monopoly over legal reporting, contravene the merger doctrine, and was more akin to the arrangement of information evident in Matthew Bender v. West Publishing.
Fair use
The court found that ROSS’ use of the headnotes was commercial and minimally transformative. Distinguishing Authors Guild v. Google, it explained that ROSS AI Search sought to replace Westlaw rather than supplementing it. ROSS’ use of the headnotes had a similar function to the original work: both were used ‘to create and optimize a legal-research platform that helps users find responsive legal material’. The court also distinguished Oracle (IPKat here) and Sega as ROSS already had access to the judicial opinions, meaning that the first factor weighed against fair use.
Although the second factor favoured fair use, the third turned on how to characterise the amount and substantiality of the material taken. ROSS argued that it only used 0.08% of Thomson Reuters’ 28 million headnotes. The court, however, held that ROSS copied the entire text of 25,000 headnotes to create a commercial substitute for Westlaw. Not only did ROSS’ lacking transformative purpose and character fail to support such extensive copying but, in the court’s view, it was completely unnecessary.
Finally, the court found that the fourth factor weighed against fair use too. While there is no standalone market for headnotes, Thomson Reuters uses headnotes to attract subscribers to Westlaw. ROSS’ use appropriated and diminished that value and caused harm in the legal-research platform market. It also foreclosed opportunities for Thomson Reuters to license its headnotes for training AI products. Again, given the freely accessible nature of judicial opinions and the comparable subscription prices, the court found no clear public benefit to support fair use.
Comment
The framing of ROSS as an ‘ordinary’ copyright case is significant. For this Kat, the most interesting reasoning appears, surprisingly, in the footnotes. In particular, the court emphasized that the Department of Justice (DOJ) filed a statement of interest supporting model training as fair use in the pending OpenAI consolidated multidistrict litigation, but did not intervene in ROSS. This statement outlined that, following Bartz v. Anthropic (IPKat here and here) and Kadrey v. Meta Platforms (IPKat here), training LLMs using written work is ‘exceedingly transformative’.
Referring to Oracle, the DOJ explained that ‘teaching an LLM to recognise relationships between data and adapt to information’ is akin to programmers learning new code to use these ‘basic skills’ to build a new system. By contrast, in ROSS, the lacking generative nature of ROSS AI Search is critical; it reduces ROSS’ use to a competing legal-research platform that uses the headnotes as a ‘shortcut’ for understanding the relationships between legal concepts encoded as tokens. The court repeatedly explains that ROSS could simply have drafted the headnotes itself using the freely accessible judicial opinions.
This characterization renders ROSS AI Search a ‘lazy’ competitor that diminishes the value of the headnotes and usurps future licensing opportunities. However, it is difficult to reconcile this reasoning with the approach in Bartz and Kadrey regarding transformativeness. Although the decisions differ in their treatment of market harm, both regard the use of copyright works to train generative AI models as highly transformative. In Bartz, however, only the lawfully acquired copies used for training (e.g. physical books scanned and destroyed) were fair use, not the pirated copies. Yet from the DOJ’s perspective, the highly transformative character of training offsets its commercial nature. Here, the DOJ opined that generative AI ‘expression’ is unable to act as ‘significant substitutive competition’ for the original work. Sidelining Kadrey’s market dilution theory, the statement endorsed commentary which argues that such an approach which accounts for indirect market competition is potentially unconstitutional (e.g. Lee).
It is possible that, if ROSS AI Search generated summaries of case law rather than merely identifying the relevant passages, its fair use argument might have been stronger (though users would have found it less useful…) ROSS would still need to explain how that function changed the purpose of its use of the headnotes, despite remaining a commercial competitor to Westlaw. More broadly, however, this comparison reveals the potentially significant weight attached to generative AI ‘expression’ when assessing whether model training is transformative. Following this line of argument, it would seem that the expressive qualities that make a work highly desirable as training material for generative AI features, in turn, support a strong transformative argument for the models trained on them.
Returning to the fact pattern in ROSS, much of the fair use analysis hinges on the court’s finding that ROSS’ use and the function of Westlaw headnotes are rough equivalents. Instead of considering the internal function of the headnotes within the training process, the court broadened its lens to consider the finished platform. In this Kat’s view, this is somewhat of an oversimplification of the subsequent use as it merges the function of headnotes within the training process with the resulting platform service. It also connects to a debate at the hearing of the pending Like Company CJEU referral (IPKat here) on whether acts involved in training and producing outputs should be assessed separately or together in the context of territoriality (Technollama here). This Kat has paws crossed for some clarity when the Advocate General’s opinion is delivered on 3 November.
In the meantime, ROSS has indicated that it will appeal the decision by petitioning the US Supreme Court, so it would seem that the headache surrounding the training of LLMs using copyright works may be far from over.
Thomson Reuters commenced copyright infringement proceedings against ROSS Intelligence in May 2020. It argued that ROSS AI Search could retrieve relevant case law when asked plain-language legal questions because it was trained on material derived from Westlaw headnotes owned by Thomson Reuters. After Thomson Reuters denied ROSS a Westlaw license, ROSS turned to a third-party legal-research company, LegalEase, to create about 25,000 question-and-answer-sets that involved legal questions ‘a lawyer would ask’. ROSS subsequently converted them into usable machine-learning training data to train ROSS AI Search.
While Thomson Reuters argued that these sets were essentially Westlaw headnotes, ROSS maintained that the lawyers drafted, not copied, them. In February 2025, the District Court granted partial summary judgment for Thomson Reuters and held that ROSS’ use of 2,243 headnotes infringed Thomson Reuters’ copyright and was not fair use. A few months later, District Circuit Judge Bibas granted ROSS’ motion for interlocutory review on originality and fair use.
 |
| A paws-on approach to model training Photo by G U on Unsplash |
Judge Montgomery-Reeves found that the Westlaw headnotes had ‘some creative spark’ pursuant to Feist as Thomson Reuters’ editors ‘made creative decisions’ (e.g. significant points of law to include, compliance with Westlaw’s drafting criteria). Although written opinions are excluded from protectable copyright subject matter (Wheaton v. Peters), case reporting that is the result of ‘intellectual labour’ such as headnotes can be protected (e.g. Callaghan v. Myers and much later, Georgia v. Public.Resources.Org). In response, ROSS unsuccessfully argued that if the headnotes were protected by copyright it would result in a monopoly over legal reporting, contravene the merger doctrine, and was more akin to the arrangement of information evident in Matthew Bender v. West Publishing.
Fair use
The court found that ROSS’ use of the headnotes was commercial and minimally transformative. Distinguishing Authors Guild v. Google, it explained that ROSS AI Search sought to replace Westlaw rather than supplementing it. ROSS’ use of the headnotes had a similar function to the original work: both were used ‘to create and optimize a legal-research platform that helps users find responsive legal material’. The court also distinguished Oracle (IPKat here) and Sega as ROSS already had access to the judicial opinions, meaning that the first factor weighed against fair use.
Although the second factor favoured fair use, the third turned on how to characterise the amount and substantiality of the material taken. ROSS argued that it only used 0.08% of Thomson Reuters’ 28 million headnotes. The court, however, held that ROSS copied the entire text of 25,000 headnotes to create a commercial substitute for Westlaw. Not only did ROSS’ lacking transformative purpose and character fail to support such extensive copying but, in the court’s view, it was completely unnecessary.
Finally, the court found that the fourth factor weighed against fair use too. While there is no standalone market for headnotes, Thomson Reuters uses headnotes to attract subscribers to Westlaw. ROSS’ use appropriated and diminished that value and caused harm in the legal-research platform market. It also foreclosed opportunities for Thomson Reuters to license its headnotes for training AI products. Again, given the freely accessible nature of judicial opinions and the comparable subscription prices, the court found no clear public benefit to support fair use.
Comment
The framing of ROSS as an ‘ordinary’ copyright case is significant. For this Kat, the most interesting reasoning appears, surprisingly, in the footnotes. In particular, the court emphasized that the Department of Justice (DOJ) filed a statement of interest supporting model training as fair use in the pending OpenAI consolidated multidistrict litigation, but did not intervene in ROSS. This statement outlined that, following Bartz v. Anthropic (IPKat here and here) and Kadrey v. Meta Platforms (IPKat here), training LLMs using written work is ‘exceedingly transformative’.
Referring to Oracle, the DOJ explained that ‘teaching an LLM to recognise relationships between data and adapt to information’ is akin to programmers learning new code to use these ‘basic skills’ to build a new system. By contrast, in ROSS, the lacking generative nature of ROSS AI Search is critical; it reduces ROSS’ use to a competing legal-research platform that uses the headnotes as a ‘shortcut’ for understanding the relationships between legal concepts encoded as tokens. The court repeatedly explains that ROSS could simply have drafted the headnotes itself using the freely accessible judicial opinions.
This characterization renders ROSS AI Search a ‘lazy’ competitor that diminishes the value of the headnotes and usurps future licensing opportunities. However, it is difficult to reconcile this reasoning with the approach in Bartz and Kadrey regarding transformativeness. Although the decisions differ in their treatment of market harm, both regard the use of copyright works to train generative AI models as highly transformative. In Bartz, however, only the lawfully acquired copies used for training (e.g. physical books scanned and destroyed) were fair use, not the pirated copies. Yet from the DOJ’s perspective, the highly transformative character of training offsets its commercial nature. Here, the DOJ opined that generative AI ‘expression’ is unable to act as ‘significant substitutive competition’ for the original work. Sidelining Kadrey’s market dilution theory, the statement endorsed commentary which argues that such an approach which accounts for indirect market competition is potentially unconstitutional (e.g. Lee).
It is possible that, if ROSS AI Search generated summaries of case law rather than merely identifying the relevant passages, its fair use argument might have been stronger (though users would have found it less useful…) ROSS would still need to explain how that function changed the purpose of its use of the headnotes, despite remaining a commercial competitor to Westlaw. More broadly, however, this comparison reveals the potentially significant weight attached to generative AI ‘expression’ when assessing whether model training is transformative. Following this line of argument, it would seem that the expressive qualities that make a work highly desirable as training material for generative AI features, in turn, support a strong transformative argument for the models trained on them.
Returning to the fact pattern in ROSS, much of the fair use analysis hinges on the court’s finding that ROSS’ use and the function of Westlaw headnotes are rough equivalents. Instead of considering the internal function of the headnotes within the training process, the court broadened its lens to consider the finished platform. In this Kat’s view, this is somewhat of an oversimplification of the subsequent use as it merges the function of headnotes within the training process with the resulting platform service. It also connects to a debate at the hearing of the pending Like Company CJEU referral (IPKat here) on whether acts involved in training and producing outputs should be assessed separately or together in the context of territoriality (Technollama here). This Kat has paws crossed for some clarity when the Advocate General’s opinion is delivered on 3 November.
In the meantime, ROSS has indicated that it will appeal the decision by petitioning the US Supreme Court, so it would seem that the headache surrounding the training of LLMs using copyright works may be far from over.
Do you want to reuse the IPKat content? Please refer to our
Policies section.
If you have any queries or requests for permission, please get in touch with the IPKat team.
Comments
Post a Comment
All comments must be moderated by a member of the IPKat team before they appear on the blog. Comments will not be allowed if the contravene the IPKat policy that readers' comments should not be obscene or defamatory; they should not consist of ad hominem attacks on members of the blog team or other comment-posters and they should make a constructive contribution to the discussion of the post on which they purport to comment.
It is also the IPKat policy that comments should not be made completely anonymously, and users should use a consistent name or pseudonym (which should not itself be defamatory or obscene, or that of another real person), either in the "identity" field, or at the beginning of the comment. Current practice is to, however, allow a limited number of comments that contravene this policy, provided that the comment has a high degree of relevance and the comment chain does not become too difficult to follow.
Learn more here: http://ipkitten.blogspot.com/p/want-to-complain.html