For years, the question hung over every AI copyright case in the United States: when a company trains a model on someone else’s content, is that fair use? On September 29, 2026, a Federal Appeals Court gave the first answer. In Thomson Reuters v. ROSS Intelligence, the Third Circuit said no.
There is an important caveat, and we will come back to it: the case did not involve generative AI. Even so, the decision is worth reading closely, because the court explained its reasoning in terms that go well beyond legal research tools. The opinion, written by Judge Tamika Montgomery-Reeves, was first filed under seal and released in full the next day. It is precedential, so it binds the federal courts within the Third Circuit, and judges handling other AI cases will certainly look at it.
So, what exactly did ROSS do?
ROSS was a startup founded by three computer science students who wanted people to ask legal questions in plain language and get relevant passages from court decisions in return. The tool did not write anything new. It simply pointed users to excerpts from about ten million judicial opinions, which anyone can copy freely.
To train the system, ROSS hired LegalEase Solutions, which prepared around 25,000 memos, each with a legal question and possible answers. And here is the problem: the questions were drafted from Westlaw headnotes, the short summaries that Thomson Reuters’ editors write for each point of law in a decision. Why? According to the record, because it was an easy way to frame the questions.
ROSS did not hide where it wanted to go. It advertised itself as an alternative to Westlaw, at similar prices, and some law firms made the switch. Thomson Reuters sued in 2020. ROSS shut down in 2021, citing the cost of the litigation. In February 2025, Judge Stephanos Bibas ruled that 2,243 headnotes had been infringed and that fair use did not apply, and the case went up on appeal.
But can a summary of a court decision really be protected?
That was ROSS’s first line of defense, and it did not work. The court recalled that the bar for originality is low: a creative spark is enough. Choosing which points of law deserves a headnote, and how to write them so they make sense on their own, involves real editorial choices. And there is more than one way to summarize the same holding.
ROSS also argued that protecting headnotes would give Thomson Reuters a monopoly over the law. The court’s answer was simple: headnotes are not the law. Judicial decisions are, and they remain free for anyone to use.
Why didn’t fair use save ROSS?
Under U.S. law, fair use is decided by weighing four factors. Three of them went against ROSS.
The first, and most important, was the purpose of the use. ROSS wanted money, and it wanted to do exactly what Westlaw does: help people find the law. The court accepted that using the headnotes to train an AI added a small step in between, but that was not enough to make the use truly transformative.
This is also where the court drew a line that will be quoted many times. ROSS relied on precedents in which copying was allowed as an intermediate step, such as Google v. Oracle. The court noted that, in those cases, copying was the only way to reach elements that were not protected. ROSS was in a very different position: it had free access to the decisions themselves and chose the headnotes because they were convenient. In the court’s words, “unlike necessity, ease is not a justification for copying.” The opinion also mentioned that ROSS employees had tried to access Westlaw using other people’s credentials, against its terms of use, which did not help.
The second factor favored ROSS, since headnotes are mostly factual. The third did not: each headnote is a complete work, and ROSS copied them whole. The argument that 2,243 headnotes were a tiny fraction of Westlaw’s 28 million went nowhere.
The fourth factor, the effect on the market, is the one AI companies should read twice. The court recognized that there is a market for licensing content as AI training data, which it called rapidly developing, and noted that Thomson Reuters already uses its headnotes to train its own AI tools. The fact that it had never licensed them to anyone else did not mean the market did not exist. By copying, ROSS took that opportunity away.
Does this settle the generative AI cases?
No, and the court made a point of saying so. In a footnote, it mentioned that the U.S. Department of Justice argued, on September 1, 2026, in the OpenAI litigation in New York, that training large language models capable of generating new content is transformative. The Third Circuit said that concern simply did not apply here, because ROSS could not create anything new and was built to replace Westlaw.
So the big disputes, such as The New York Times against OpenAI and the record labels against Suno and Udio, are still open. But it is no coincidence that the music industry and major film studios filed briefs supporting Thomson Reuters. The idea that convenience is not necessity, and the recognition of a market for training data, will certainly appear in those cases.
And how would this case look in Brazil?
Probably worse for ROSS. Brazil does not have fair use. Under the Brazilian Copyright Law (Lei 9.610/98), any use of a protected work requires the author’s prior and express authorization (article 29), and the law specifically mentions storing works in a computer and including them in a database. The exceptions are listed in articles 46 to 48, and none of them cover AI training.
It is true that the Superior Court of Justice (STJ) has, on occasion, accepted uses outside that list by applying the three-step test of the Berne Convention, as in REsp 964.404/ES. But one of those steps asks whether the use conflicts with the normal exploitation of the work. That is essentially the same question the Third Circuit answered when it looked at the market. A tool designed to replace the original would hardly pass.
The logic about what is protected would also be very similar. In Brazil, court decisions are not protected by copyright (article 8, IV), but databases and compilations are, when the selection and organization of their content involve intellectual creation (article 7, XIII). In practice: training a system on court decisions is fine; copying the editorial work a publisher built on top of them probably is not. And ROSS’s use of other people’s logins would open a second front, for breach of the license terms, which the STJ has treated as contractual liability, with a ten-year limitation period (REsp 1.907.034).
There is also an AI bill. The version of PL 2338/2023 approved by the Senate in December 2024 requires developers to disclose the protected content used in training (article 62), limits the data mining exception to research, educational and cultural institutions such as museums, archives and libraries (article 63), allows rights holders to prohibit the use of their works (article 64) and, if they do not, gives them the right to be paid (article 65). A commercial competitor like ROSS would be outside the exception. The problem is that the bill is still waiting for the rapporteur’s opinion in the Chamber of Deputies, and copyright remains one of its most disputed topics.
What should companies take from this?
For those who own content, the message is encouraging: the licensing market for AI counts, even when it is still taking shape. Keeping a record of licenses, negotiations and even your own internal use of content for AI can make a real difference if a dispute comes up, in the United States or here.
For those developing or buying AI tools, the lesson is less comfortable. Where the data comes from matters as much as how good the model is. If the underlying information is freely available, copying someone else’s curated version of it because it saves time is exactly what the court refused to accept.
No Brazilian court has ruled on AI training yet. But our law starts from authorization, not from exceptions, and that already makes it stricter than the American system. If even under fair use a company could not build a substitute out of someone else’s work, it is hard to imagine a Brazilian court seeing it differently.
—
Advogado(a) autor(a) do comentário: Natalia Eleutério Garcia Gazote, Lígia Ferreira Marcondes Rocha e Cesar Peduti Filho, Peduti Advogados
“US appeals court upholds Thomson Reuters’ landmark win in AI training lawsuit (Reuters)”:
“Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026), full opinion”:
“US appeals court rejects ‘fair use’ defense over AI training – as Thomson Reuters wins copyright case backed by RIAA and NMPA”:
“Signed, Sealed, Affirmed: Third Circuit Backs Thomson Reuters over ROSS”:
https://copyrightlately.com/thomson-reuters-v-ross-third-circuit/
“3rd Circ. Affirms Thomson Reuters’ Westlaw AI Copyright Win”:
https://www.law360.com/articles/2531563/3rd-circ-affirms-thomson-reuters-westlaw-ai-copyright-win
“Lei nº 9.610, de 19 de fevereiro de 1998 (Lei de Direitos Autorais)”:
https://www.planalto.gov.br/ccivil_03/leis/l9610.htm
“Em evento religioso, música não rende direito autoral (STJ, REsp 964.404/ES)”:
https://www.conjur.com.br/2011-abr-01/eventos-religiosos-musica-nao-rende-direitos-autorais-ecad/
“PL 2338/2023, tramitação completa”:
https://www25.senado.leg.br/web/atividade/materias/-/materia/157233
“Comissão Especial sobre Inteligência Artificial (PL 2338/23)”:
—
Se quiser saber mais sobre este tema, contate o autor ou o Dr. Cesar Peduti Filho.
If you want to learn more about this topic, contact the author or the managing partner, Dr. Cesar Peduti Filho.