ROSS Intelligence never showed its users a single Westlaw headnote. The legal research start-up used those short case summaries only to train its AI, through roughly 25,000 memos written by a contractor, and its product answered questions with passages from court opinions. On 29 September the US Court of Appeals for the Third Circuit ruled that this was still copyright infringement and not fair use, the first time a federal appeals court has decided whether training an AI system on copyrighted material can be fair use.
The court stressed how narrow its holding is: ROSS's tool retrieved existing text rather than generating anything new. But for small teams building AI products on other people's data, the reasoning is a clear map of where the legal risk sits.
What ROSS Built and Where the Headnotes Came In
Headnotes are the short summaries that Thomson Reuters editors write to identify points of law in judicial opinions on Westlaw, its legal research platform. ROSS set out to build a search engine that answered plain-language legal questions with relevant passages from existing opinions, and it intended the product to compete directly with Westlaw, according to an analysis by the law firm Ballard Spahr.
To train it, ROSS engaged a third party, LegalEase, which wrote roughly 25,000 memos. Each paired a legal question with four to six opinion passages graded from "great" to "irrelevant", and the drafters used Westlaw headnotes to frame many of the questions, as Authors Alliance summarises the record. The memos were converted into machine-readable data. The people who used ROSS never saw a headnote.
Thomson Reuters sued in May 2020. ROSS Intelligence, which had raised about $13 million, went out of business in January 2021 under the weight of its legal costs, Tech Times reports, but the case outlived the company. In February 2025 Judge Stephanos Bibas, sitting by designation in Delaware, reversed an earlier ruling in the case and held that 2,243 headnotes were original, infringed and not fairly used, then sent both questions to the Third Circuit.
How the Court Weighed Fair Use
The panel of Judges Restrepo, Montgomery-Reeves and Bove, with Judge Montgomery-Reeves writing, first held that the headnotes clear copyright's low bar for originality because they show "some creative spark" in what to include and how to say it. The opinion describes the dispute as "no more than an ordinary copyright case".
US law then weighs four fair use factors. Three went against ROSS:
- Purpose and character. The use was highly commercial and only minimally transformative. Westlaw uses headnotes to help lawyers identify relevant law, and ROSS used them to teach its AI to identify relevant judicial passages. The training step did not change that purpose.
- Nature of the work. This factor slightly favoured ROSS, because headnotes summarise judicial opinions and are more factual.
- Amount used. ROSS copied entire headnotes, and each one counts as a work in its own right, so this factor favoured Thomson Reuters.
- Market effect. The court found harm in two markets: legal research, where ROSS was built as a substitute for Westlaw, and the licensing of headnotes as AI training data, a market it called "rapidly developing".
The court also rejected ROSS's comparison with older cases that allowed intermediate copying of software code to make products work together. Copying the headnotes was the easy route to the training memos, Ballard Spahr notes, but it was not necessary to reach the unprotected material underneath: the court opinions themselves.
The Line It Drew Around Generative AI
The most quoted passage is a footnote. Answering the generative AI cases, the panel wrote that, "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." Ballard Spahr reads this as leaving the central question untouched: whether training a generative model on books or other expressive works is fair use has not been decided.
Authors Alliance, which filed a brief supporting ROSS, thinks the carve-out may be weaker than it looks. Much of the footnote's work, it argues, is done by substitution rather than by the generative distinction: "ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw." It expects plaintiffs in the generative cases to argue that any model competing in their market belongs on ROSS's side of that line.
There are plenty of those cases. The Association of American Publishers counts approximately 150 in the United States. Courts are not yet aligned either. Joe Cahill of Baker Botts points out that two June 2025 rulings in the Northern District of California treated licensing markets very differently: Judge Alsup held that a market for licensing books as training data "is not one the Copyright Act entitles Authors to exploit", and Judge Chhabria held that whether such a market "exists or is likely to develop is irrelevant".
What Small Builders Can Take From the Ruling
This is not legal advice, and the generative questions remain open. But the court's reasoning gives teams that train or fine-tune models on third-party data six points worth checking.
- Know which layer you are copying. The opinions ROSS needed were public; the headnotes were Thomson Reuters's editorial layer on top. "There is nothing that Thomson Reuters created that Ross could not have created for itself or hired LegalEase to create for it without infringing Thomson Reuters's copyrights," Judge Bibas wrote, as quoted by IPWatchdog. On the court's reasoning, primary sources and your own annotations are safer ground than a competitor's summaries.
- Training-only use is not a shield. The headnotes never reached ROSS's users, and the court still found that the use served the same purpose as Westlaw's.
- Building a substitute for your source is the riskiest position. Both the first and the fourth factors turned on ROSS making a direct competitor to the product it learned from.
- A contractor does not move the risk. ROSS outsourced the memos to LegalEase, and the ruling still went against ROSS.
- Licences are now evidence. The court treated Thomson Reuters's own use of headnotes to train its AI search products as evidence of a licensing market, and said the absence of licences to others "does not disprove that a market exists to do so". Baker Botts' takeaway is that licence terms may become evidence in fair use litigation.
- Litigation can outlast a start-up. ROSS Intelligence closed in January 2021. The appeal was decided more than five years later.
Who Welcomed It and Who Did Not
Publishers welcomed it. "Rightsholders are entitled to equitable participation in the AI market," said Maria A. Pallante, president and chief executive of the Association of American Publishers, which called the decision a "powerful precedent".
Fair use advocates were more critical. Brandon Butler, who runs Re:Create, said the reasoning in Judge Bibas's opinion relied on the broad concept of market competition rather than actual market substitution, which limits how far it reaches: "As it stands, however, this first judicial walk through the AI fair use labyrinth has hit a dead end." Dave Hansen of Authors Alliance wrote that the court "gets two important things wrong": it hollows out the merger doctrine, which stops copyright from protecting an idea that can only be expressed in a few ways, and it reads fair use against the non-expressive, non-public copying the doctrine has long protected. The court dismissed the merger argument with an analogy: "just as a banana costume can take myriad designs, so too can headnotes have differing expressions".
What to Watch Next
Outside the Third Circuit, other courts can choose whether to follow the ruling. Baker Botts suggests watching three things: whether further review takes up the circularity question raised in California, whether the next district court follows the Third Circuit, and whether courts start testing training-data licences the way patent courts test comparable licences.
For builders, the immediate work is provenance: where each dataset came from, whose editorial effort it contains and whether the product built on it competes with that source. The question applies just as much to teams fine-tuning open-weight models, such as the one MW3.biz covered when it reported that Mistral Large 4 will put a trillion parameters into open weights. The weights may be open; the data you add to them still needs a clean history.
This article is for information only and is not legal advice.
