Third Circuit Poised to Decide Whether Training an AI Model on Copyrighted Content Is Fair Use

In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., argued before the U.S. Court of Appeals for the Third Circuit on June 11, 2026, the parties dispute whether the use of copyrighted materials to train an artificial intelligence model qualifies as fair use under the Copyright Act. The Third Circuit’s forthcoming decision may have significant implications beyond the legal research industry and is likely to influence the evolution of artificial intelligence training models.

While AI in legal practice has drawn significant attention, particularly in recent months, the ROSS Intelligence dispute dates back to a time when few legal practitioners were aware of AI. In 2014, students from the University of Toronto applied IBM’s Watson technology to train a computer model to think like a lawyer. From this idea, Ross Intelligence, an AI legal search engine, was born.

Training AI on Law and Copyright Infringement Claims

To train the model to think like a lawyer, ROSS used both judicial opinions and legal memoranda written by a third party, LegalEase Solutions. Each of the legal memos posed basic legal questions to train the ROSS model to learn to rank potential answers to legal questions. But as ROSS recognized in its brief before the Third Circuit, the legal questions were often adapted from Westlaw headnotes because they provided an easy way to frame the issues. Importantly, Westlaw’s parent company, Thomson Reuters, had denied ROSS a license to use its services — seemingly to prevent this use — and ROSS used LegalEase to get around the ban.

Thomson Reuters sued ROSS in May 2020 in the District of Delaware, alleging copyright infringement and tortious interference. The case was assigned to Judge Stephanos Bibas of the Third Circuit, sitting by designation. In Sept. 2023, Judge Bibas denied the parties’ competing motions for summary judgment, concluding that questions of originality and fair use needed to go to a jury. But two days before trial, the court postponed the proceedings and invited renewed summary judgment briefing.

In Feb. 2025, Judge Bibas issued a second opinion reversing himself on nearly every issue. The court held that at least 2,243 of the headnotes ROSS copied were original and copyrightable as a matter of law, that ROSS’s use was not transformative because ROSS “took the headnotes to make it easier to develop a competing legal research tool,” and that ROSS’s use harmed both the existing market for Westlaw and the potential market for licensing Westlaw content as AI training data. Recognizing that “the controlling law is unclear” and that the questions presented are “novel and difficult,” Judge Bibas certified the case for interlocutory appeal under 28 U.S.C. § 1292(b), and the Third Circuit accepted review.

The dispute before the Third Circuit is twofold: whether headnotes based on non-copyrightable judicial opinions are copyrightable and, if so, whether use of the headnotes to train an AI model is a fair use of the copyrighted materials.

The Questions of Copyrightability and Fair Use

On appeal, ROSS argues that Judge Bibas read copyrightability too broadly and fair use too narrowly. It contends that headnotes are not original because West’s editors are trained to follow the language of judicial opinions as closely as possible — meaning that the headnotes draw on non-copyrightable material belonging to the public. ROSS also asserts that, even if the headnotes are original works, ROSS’s use was fair under applicable standards. In its fair use argument before the Third Circuit, ROSS relies on the doctrine’s common law roots to emphasize that the use of existing technology as a springboard to invent new technology is fair. Even beyond this common law argument, ROSS contends its use satisfies the four-factor test for fair use under Section 107 of the Copyright Act. ROSS argues that if the Third Circuit affirms Judge Bibas’ decision, new uses of AI in the legal industry (and beyond) will be stifled — the opposite of the intention of the Copyright Act’s ultimate aim to enhance the general public good with creative works.

Thomson Reuters counters that this is not a case about who owns the law, but about a competitor copying creative editorial content to build a market substitute. It emphasizes that the Supreme Court has repeatedly recognized headnotes as copyrightable, and that ROSS could have built its product using the collection of judicial opinions it already had in its possession without copying any Westlaw content. On fair use, Thomson Reuters argues that ROSS’s purpose was identical to its own — helping researchers find relevant law — and that ROSS marketed its product as a direct replacement for Westlaw at a lower price point. It also urges that ROSS acted in bad faith by obtaining the content through LegalEase after Thomson Reuters refused a license, and that ROSS’s copying harmed both the existing market for Westlaw subscriptions and the emerging licensing market for AI training data.

At oral argument on June 11, 2026, a three-judge Third Circuit panel appeared skeptical of ROSS’s position. Judge Tamika Montgomery-Reeves pressed ROSS on why Westlaw’s headnotes are not sufficiently original to qualify for copyright protection, while Judge Emil J. Bove pushed back on the argument that ROSS’s product was meaningfully different from Westlaw’s, asking, “Why is it transformative as opposed to just a different type of legal search engine?” Counsel for ROSS responded that the company had “taught [a] machine how to think like a lawyer” and that the use was “spectacularly transformational.” The panel’s questioning, however, suggests that ROSS faces an uphill battle on both copyrightability and fair use.

A Transformative Opinion in the AI Era

Whatever the Third Circuit ultimately decides, the implications will be felt far beyond the legal research industry. A decision affirming Judge Bibas would lend significant weight to copyright holders in pending AI copyright suits against companies such as OpenAI, Meta and Anthropic, and would likely accelerate the trend toward licensing arrangements between AI developers and content owners. A decision reversing the district court could open the door for AI companies to train models on proprietary databases without licensing agreements, so long as the resulting product is meaningfully different from the original. Either way, Thomson Reuters v. ROSS Intelligence will be one of the most consequential copyright opinions of the AI era, and counsel on both sides of the AI copyright divide should monitor the case closely.