At 3rd Circuit, Judges Press ROSS and Thomson Reuters on Fair Use, AI Training and Market Harm

The closely watched copyright litigation between Thomson Reuters and ROSS Intelligence returned to the spotlight this month as a three-judge panel of the 3rd U.S. Circuit Court of Appeals heard arguments over whether ROSS’s use of Westlaw headnotes to train its legal research system was protected by fair use.

At oral argument June 11 in Philadelphia, the three-judge panel, consisting of  Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves and Emil J. Bove,  focused much of its questioning on the first and fourth fair-use factors of whether ROSS’s use was transformative and whether it harmed an actual or potential market for Thomson Reuters’ copyrighted works.

Earlier this week, the court released the transcript of the hearing.

Representing ROSS, White & Case partner Mark S. Davies urged the court to reverse the district court’s ruling and find that its use of the headnotes was fair use.

Thomson Reuters and West Publishing, represented by Dale M. Cendali, partner at Kirkland & Ellis, argued that ROSS copied protected editorial content to build a competing legal research product and that the district court correctly rejected the fair-use defense.

The argument came in an appeal from the Delaware federal court’s rulings in a case that has drawn attention across the legal tech industry because of its implications for AI training and copyright law. Even though the case predates the current generation of generative AI tools, ROSS argued that its technology was an early example of the same broader AI transformation now reshaping legal research.

ROSS Argues Fair Use

Davies opened by invoking the 3rd Circuit’s recent decision in ASTM v. UpCodes, arguing that under that precedent, the court should summarily reverse based on fair use. He said ROSS had built a legal search platform in 2015 by adapting headnotes into questions and using them to create training materials for a system that could connect natural-language legal questions to judicial opinions.

But Judge Restrepo immediately redirected the argument to the fourth fair-use factor, asking Davies to address the effect of the use on the market. Davies responded that the copyrighted work at issue was the headnote itself and that there was no market for a single headnote or for headnotes as AI training material.

Judge Bove pressed him on whether there was, or could be, a market for AI training. Davies argued that ROSS’s use was too specific to define a market around it. ROSS did not merely take headnotes as raw data, he said, but built training memos that included questions and multiple answers, some right and some wrong, to train the model.

To let Thomson Reuters claim a lost market for that use, he argued, would be circular because any copyright owner could say it would have licensed the very use later claimed to be fair.

The judges repeatedly questioned Davies on how, from the user’s perspective, ROSS’s product differed from Westlaw. Judge Bove asked why ROSS’s system was transformative “as opposed to just a different type of legal search engine.”

When Davies focused on the internal technology, Judge Restrepo noted that ROSS was nevertheless a direct competitor to Westlaw, and Judge Montgomery-Reeves asked whether the product was aimed at serving as a commercial substitute.

Davies acknowledged that ROSS competed with Westlaw, but said copyright law was not designed to protect incumbents from competition. He described ROSS’s product as one in which a user asked a natural-language legal question and received a stream of answers in the form of quotations from judicial opinions. Because the output came from court opinions rather than generated text, he said, there was “no risk of hallucinations.”

Judge Bove continued to press for a factual description of the product, at one point telling Davies there was “a lot of dancing going on” and that he was trying to understand what ROSS planned to do and why it was materially different from Westlaw.

“Could you please describe to me what Ross’s proposed engine would have looked like, and the differences, and why it was so transformative from what happens when I log into Westlaw in the morning?” Judge Bove asked.

Davies responded that, from a user perspective, both products helped users find cases, but that ROSS operated differently internally because it used deep-learning techniques to teach the system the semantic relationships of legal language.

“The question has legal language in it, and what we have taught this machine now is how to think like a lawyer,” Davies said.

The panel also questioned Davies on the copyrightability of the headnotes themselves. Judge Montgomery-Reeves asked why headnotes that condensed and reformulated judicial language were not sufficiently original to qualify for copyright protection.

Davies argued that judicial opinions belong to the public and that the differences between the opinions and the headnotes were too slight to support protection. The judges, however, probed whether the selection and condensation of opinion language reflected editorial judgment.

Judge Bove focused on one example in which West’s editor had excerpted and condensed language from an opinion. Even if the headnote closely tracked the opinion, he suggested, the editor still made judgments about what was salient and what to include. Davies maintained that West’s editorial instructions were designed to remove creativity and track the language of the opinion as closely as possible.

Davies argued that ROSS should prevail on the second and third fair-use factors because the headnotes were factual and because ROSS copied only a small fraction of Westlaw’s total headnotes. He said ROSS used only 0.08% of Westlaw’s 28 million headnotes.

“We took so little,” He said. “We just took 0.08% is all we copied of 28 million headnotes.”

When Judge Montgomery-Reeves asked what result should follow if Thomson Reuters were ahead on factors one and four, while ROSS prevailed on factors two and three, Davies said ROSS should still win. Judge Restrepo pressed him: “If you take 2 and 3 and they take 1 and 4, you win?” Davies answered that fair use is a multi-factor test.

Judge Bove also asked Davies about Thomson Reuters’ allegations of bad faith, including communications involving Dentons and efforts by ROSS to obtain Westlaw material. (The transcript does not elaborate on how Dentons was implicated and I could find no reference to that in the parties’ redacted briefs.)

Davies said ROSS had paid for the original database and for the memos and that any license issues should be addressed in separate contract claims, not the fair-use analysis. He cited cases for the proposition that copyright is not reserved for the “well-behaved” and argued that even if there were bad-faith evidence, it should not defeat fair use.

‘A Classic Case of Substitution’

Arguing for Thomson Reuters, Cendali began by addressing the scope of the appeal. She said the district court had granted summary judgment on 2,834 headnotes, of which Judge Stephanos Bibas had found 2,430 protectable and infringed. She said the court need not go beyond that set of headnotes, although she also argued that headnotes are categorically copyrightable under long-established precedent.

Cendali then turned to fair use, framing the case as one of substitution. Citing Supreme Court authority, she said substitution is “copyright’s bete noire” and “this is a classic case of substitution” in that ROSS built and marketed a product to replace Westlaw.

She pointed to ROSS marketing materials that, according to Thomson Reuters, urged users to “choose ROSS or Westlaw,” and said that was a quintessential example of a substitute product.

Judge Bove asked Cendali to identify the relevant market for the fourth factor, noting that the parties appeared “worlds apart” on that question.

Cendali identified three forms of market harm: harm from ROSS’s product competing as a substitute for Westlaw, harm to Thomson Reuters’ ability to use its own headnotes exclusively to train its own AI products, and harm to a potential market for licensing headnotes for AI training by other AI companies.

Cendali disputed ROSS’s suggestion that Thomson Reuters had not used its headnotes for AI. She said Thomson Reuters had trained on its own headnotes years before ROSS was founded and that Westlaw’s use of AI and natural-language search predated ROSS. She also pointed to amicus briefing by Lexis as evidence that other legal research providers were developing AI products during the same period.

“So as much as they try to cast themselves as tremendous innovators, they really were latecomers,” Cendali argued.

Judge Montgomery-Reeves drew an analogy to the 3rd Circuit’s Video Pipeline case involving Disney trailers, asking whether the absence of a stand-alone market for the copied material defeated a market-harm argument.

It did not, Cendali said. She argued that, as in Video Pipeline, copied content could still harm the value of a broader platform even if it was not separately sold on its own.

On the first fair-use factor, Cendali argued that ROSS’s use was not transformative because it used Westlaw headnotes to build a competing legal research tool. Training was merely the means to an end, she said, and under the Supreme Court’s Warhol decision, the proper inquiry focuses on the purpose of the use. Here, both products served the same purpose: helping users conduct legal research.

Judge Bove asked what the record showed about the nature of ROSS’s training and technology. Cendali responded that the product ultimately returned cases and quotes, which she said was similar to what Westlaw did. She distinguished the case from disputes involving gen AI systems that produce new text, saying ROSS’s product was a legal research system for finding court opinions in response to legal topics.

Cendali also argued that ROSS took the “heart” of Westlaw’s work, even if the number of headnotes copied was small compared with Westlaw’s overall database. She said Westlaw’s editorial enhancements, including headnotes and the Key Number System, are what differentiate it from a bare collection of cases. ROSS, she said, took the question-and-answer pairs created through that editorial work rather than creating its own from the underlying opinions.

The panel also explored the relevance of cases involving intermediate copying, such as Sega, Sony and Google v. Oracle. Cendali argued that those cases did not help ROSS because they involved copying necessary to understand functional computer code or achieve compatibility. Here, she said, ROSS had access to the public judicial opinions and did not need to copy Westlaw’s headnotes to build its system.

Rebuttal: Not Merely ‘AI’ As A Label

In rebuttal, Davies returned to the technology, emphasizing that ROSS was not merely invoking “AI” as a label. He said the system used deep neural-network technology similar in kind to the systems later associated with ChatGPT and Anthropic, although he acknowledged that ROSS’s product was not generative AI.

In response to Judge Montgomery-Reeves, he agreed that ROSS did not write briefs or generate legal answers, but instead returned quotes from judicial opinions.

Davies also challenged Thomson Reuters’ market-harm evidence, arguing that the record lacked concrete proof that ROSS harmed Westlaw’s market. Judge Bove reminded him that, under precedent, ROSS bore the burden on fair use, including factor four.

Davies acknowledged the burden but argued that Thomson Reuters was better positioned to produce evidence such as subscriber numbers and customer data. He said the record included evidence that some users left ROSS and returned to Westlaw.

The argument ended with an exchange over whether the intermediate-copying cases were meaningfully analogous. Judge Bove noted that in Sega and Sony, the copying was necessary because humans could not read object code, whereas ROSS employees could read the judicial opinions and create their own training materials from them.

Davies responded that there is no strict necessity requirement for fair use and that ROSS had a legitimate purpose in building a new legal research technology under time pressure as a startup.

As the argument wrapped up, Judge Bove framed the issue more broadly, telling Davies, “I think this case is about balancing what we have and where are the red lines.”

“It’s all about balancing,” Davies agreed, “as all copyright cases are, Your Honor.”

The court did not indicate when it would rule.