In a ruling dated September 29, the U.S. Court of Appeals for the Third Circuit upheld a lower court's decision that ROSS Intelligence infringed copyright by using the legal research service Westlaw's editorial summaries to train its AI, and rejected ROSS's fair use defense.
The material at issue was not the publicly available court opinions themselves. It was the "headnotes": short summaries in which Thomson Reuters editors select the key legal points from a ruling and condense them.
ROSS's service did not display the headnotes it was trained on to users. Even so, the court gave great weight to the fact that ROSS used the headnotes to build a legal search service that competes directly with Westlaw.
On appeal, the court also reversed the lower court on the third fair use factor, "the amount and substantiality of the portion used," which the district court had found favored ROSS. The appeals court found that it weighed against ROSS.
The case turns on more than whether an AI reproduces its training text in its output. It asks what was copied at the training stage, how much was copied, and what kind of service was ultimately built.
Public court opinions are one thing; an editor's summary is another
Westlaw headnotes condense the key legal determinations in an opinion and point readers to the corresponding passages.
Editors decide which points of an opinion to include, how to word them, and how much context to add so the meaning is clear without reading the original opinion.
In other words, an opinion written by a judge and a summary written by a private company's editors to organize its content are not the same thing under copyright law.
The appeals court recognized the 2,243 headnotes at issue in this case as protectable because it found they showed the minimal creativity required in these choices and in their expression.
The bar for the creativity needed for copyright protection is not high. Even text that conveys legal information accurately can be protected if it was independently created and involves some creative choices in expression.
ROSS argued that granting copyright protection to headnotes would amount to giving someone a monopoly over "the law itself."
The court distinguished between the two, however, saying that what is protected is the headnotes created by Thomson Reuters editors, not the court opinions themselves. The opinions remain free to use.
ROSS had access to roughly 10 million court opinions that are not protected by copyright.
Its service let users enter a legal question in natural language and returned the existing passages from court opinions relevant to it.
It was not a generative AI that writes new text like today's large language models. It was a search AI trained to learn which passages of opinions answer a given question.
The training data was about 25,000 training memos created by LegalEase Solutions and its subcontractors.
Each memo contained a legal question and 4 to 6 corresponding passages from opinions. The passages were rated for relevance, from those that fully answered the question to those that were unrelated, and were then converted into a format the AI could use.
This is where Westlaw's headnotes came in.
The headnotes were used as source material for the questions in the memos, and the passages rated as the best answers were often the ones that Westlaw linked from that headnote.
Pages 5–6 of the appeals court opinion show that the relationship between the "legal questions" created by Westlaw editors and their corresponding opinion passages was carried into ROSS's AI training data.
Simply collecting large numbers of court opinions does not produce a search system that returns the right passage for a natural-language question.
You need to build the correspondence: which passage answers which question. ROSS used Westlaw's headnotes to make that work more efficient.
Even without headnotes in the output, copying them in full was a problem
Fair use is a doctrine in U.S. copyright law that allows copyrighted works to be used without the owner's permission when certain conditions are met.
The U.S. Copyright Office explains that courts weigh the following four factors together:
- The purpose and character of the use
- The nature of the copyrighted work
- The amount and substantiality of the portion used
- The effect on the market for or value of the work
Commercial use does not automatically defeat fair use, and there is no uniform standard such as "up to X percent is allowed."
The major problem for ROSS was the ultimate purpose for which it used the headnotes.
Westlaw uses headnotes to help users find court opinions relevant to a legal question.
ROSS likewise used Westlaw's headnotes as training material to build a legal search service that finds relevant opinion passages for natural-language questions.
The court concluded that even with the technical step of AI training in between, the ultimate purpose was very close to Westlaw's.
ROSS ran advertising that directly compared it with Westlaw and sought to win customers at a similar price point. Some law firms actually switched from Westlaw to ROSS.
This direct competition was weighed heavily in deciding whether the use was "transformative," meaning a use for a purpose different from that of the original work.
The appeals court held that ROSS's use had, at most, very little transformative character.
Still, upholding the lower court did not mean adopting the same reasoning on every one of the four factors.
In particular, on the third factor, "the amount and substantiality of the portion used," the district court found for ROSS, while the appeals court found against it.
| Fair use factor | District court, Feb. 11, 2025 | Appeals court, Sept. 29, 2026 |
|---|---|---|
| Purpose and character of use | Against ROSS. Commercial use to build a competing service | Against ROSS. Even through AI training, the ultimate purpose is very close to Westlaw's |
| Nature of the work | For ROSS. Largely factual, with limited creativity | Slightly for ROSS. Published, and text that accurately explains the law |
| Amount and substantiality | For ROSS. Headnotes are not shown in answers to users | Against ROSS. Each headnote was copied in full, and training memos could have been made from the opinions alone |
| Effect on market or value | Against ROSS. Serves as a substitute for Westlaw and affects the market for training data | Against ROSS. Undermines the value of the headnotes that give people a reason to subscribe to Westlaw, and the opportunity to license them for training |
The table organizes the district court ruling and the appeals court ruling by the four fair use factors.
The analysis is not a majority vote on how many factors favored each side; it is an overall judgment that includes the purpose of the use and its effect on the market.
The district court treated the fact that ROSS's service did not show the headnotes themselves to users as a circumstance favoring ROSS on the third factor.
The appeals court, by contrast, emphasized that each headnote is an independent work and that ROSS copied each one in its entirety.
ROSS also argued that it used only about 0.08% of the roughly 28 million headnotes in Westlaw.
The court held, however, that the proper focus is not the proportion of Westlaw as a whole but the fact that each individual headnote was copied wholesale.
ROSS also had the option of creating training memos from the original court opinions.
The court said that the fact that using Westlaw's headnotes made it easy to build training data is separate from whether copying them was necessary.
The circumstance that the material was used only at the AI training stage and the original text never appeared in the final answers could not justify the full copying here.
The reason to subscribe to Westlaw, and the market for AI training licenses
That Thomson Reuters does not sell headnotes as a standalone product does not mean their text has no market value.
The appeals court emphasized that headnotes are one of the reasons people subscribe to Westlaw.
Being able to efficiently find relevant opinions from a legal issue is part of what makes Westlaw valuable.
The court concluded that if a competing search service uses Westlaw's headnotes to perform the same function, it shifts that value to the competitor.
The court also considered the potential future market for licensing the material as AI training data.
According to the ruling, Thomson Reuters itself uses headnotes to train AI search features such as WestSearch Plus.
The court added that because a market for licensing content for AI training is developing, the fact that Thomson Reuters has not previously licensed headnotes to third parties for training does not rule out its entering that market in the future.
The assessment is that by using the headnotes as training data without permission, ROSS deprived Thomson Reuters of a future licensing opportunity.
This, however, is the court's evaluation of market effect under the fourth fair use factor; the ruling did not fix a specific amount of damages.
The difference from Google Books: whether it replaces the original product
ROSS also cited the Google Books precedent, in which books were scanned in bulk and made searchable.
The appeals court held, however, that the purpose of use differed from that in Google Books.
In Google Books, the purpose of making books full-text searchable was different from the purpose of having people read the books themselves.
Search results can also lead people to learn about a book and go on to buy it.
ROSS, on the other hand, was not a mechanism that sent users back to Westlaw.
ROSS itself aimed to be a replacement for Westlaw.
For the court, this difference was important in evaluating transformativeness and market effect.
From the standpoint of protecting access to the law, meanwhile, concerns have been raised that copyright protection for Westlaw headnotes could be extended too far.
In a September 2025 statement, the Electronic Frontier Foundation (EFF) explained that it had filed an amicus brief together with the American Library Association and others.
Their position was that the creative portion of headnotes is limited, and that their character as conveyors of legal facts should weigh more heavily in the fair use analysis.
The appeals court also acknowledged that headnotes are more factual than works such as novels, and found the second factor slightly favored ROSS.
Still, it held that using the publicly available law itself is different from building a competing service using expression added by a private company's editors.
The mere fact that text is factual does not eliminate copyright protection or the other fair use factors.
This is not a ruling on generative AI training in general
The AI used in this case was a non-generative AI that searches for and returns existing passages from court opinions.
In a footnote, the appeals court also mentioned a statement of interest on generative AI training filed by the U.S. Department of Justice in separate copyright litigation involving OpenAI.
It made clear, however, that the circumstances differ from ROSS's system in this case.
Large language models can use the data they learn from to generate new expression.
ROSS's AI, by contrast, did not generate new text; it found and presented relevant passages from existing court opinions.
Moreover, ROSS's aim was to build a legal search service that competes directly with Westlaw.
For that reason, the court said, the idea of "producing new expression through training," which is debated in other lawsuits over generative AI, cannot simply be applied to ROSS.
This does not mean generative AI can freely train on copyrighted works.
Conversely, the ruling does not hold that all AI training is copyright infringement.
It is a judgment based on specific circumstances: what was copied, how much was used, what the AI does as a service, and whether it competes with the original work.
The scope of this review is also limited.
The appeal was not an ordinary appeal after the entire case had ended, but an "interlocutory appeal" in which a higher court was asked to rule on particular legal questions during the litigation.
The main issues considered were whether the 2,243 headnotes have the creativity required for copyright protection and whether ROSS's use qualifies as fair use.
The appeals court did not decide whether copyright subsists in headnotes that simply reproduce language from the opinion.
It also did not decide the creativity of Westlaw's classification system, the "Key Number System," because ROSS did not contest it on appeal.
One key takeaway for AI developers is the need to distinguish between public information and the editing and organization that a third party has added to it.
The path of using the publicly available court opinions themselves remains open.
But when the correspondence between questions and the opinion passages that answer them, built by another company's editorial work, is carried directly into training data, developers need to consider the copyright in that expression and its relationship to the market.
For rights clearance, what matters is not only whether the AI's output shows the original text of its training data, but also who created the training data, how much of it was copied, and how the service built from that training competes with the original service.