Third Circuit: AI Training Copyright Fair Use Denied for Competitive Use

• The Third Circuit affirmed that training AI models on copyrighted material for competitive purposes does not constitute fair use.
• The case involved Thomson Reuters suing ROSS Intelligence for using Westlaw's proprietary headnotes to train its AI without permission.
• Courts found Westlaw's headnotes copyrightable due to the "creative spark" in their selection and arrangement.
• ROSS's use was deemed "minimally transformative at best" and directly competitive, negating its fair use defense.
• This ruling sets a significant precedent for AI training data, highlighting intellectual property risks and the need for proper licensing.

The Third Circuit's decision establishes a significant precedent, clarifying that training artificial intelligence models on copyrighted content, especially for direct market competition, typically falls outside the bounds of fair use.

The Third Circuit recently issued a pivotal ruling, affirming that training artificial intelligence models on copyrighted material for competitive purposes does not qualify as fair use, thereby establishing a significant AI copyright infringement precedent. This decision stems from a lawsuit initiated in 2020 by Thomson Reuters, the parent company of Westlaw, against ROSS Intelligence, an AI startup that has since ceased operations. Thomson Reuters alleged that ROSS had unlawfully utilized Westlaw's proprietary headnotes—conise summaries of critical legal issues preceding judicial opinions—to develop its own competing legal research product.

According to the complaint, ROSS gained access to these headnotes through a third-party vendor, LegalEase Solutions, after Westlaw had previously denied ROSS direct access. Rather than using the material for traditional legal research, ROSS specifically employed the headnotes to train its artificial intelligence system, aiming to rush a rival product to market. In 2025, a federal judge in Philadelphia sided partially with Thomson Reuters, granting them summary judgment. The lower court concluded that Westlaw's headnotes were indeed copyrightable and that ROSS's unauthorized use constituted a violation of copyright law.

ROSS had argued that the headnotes lacked sufficient originality for copyright protection, claiming they were too similar to the uncopyrightable judicial opinions they summarized. However, the district court found that the unique selection and arrangement of these notes by Westlaw's editors demonstrated a "creative spark," akin to an artist's original work. This creative element rendered Westlaw's non-verbatim headnotes original works, thus eligible for copyright protection. Furthermore, the lower court determined that ROSS's application of the headnotes was insufficiently transformative and directly supported its market competition with Westlaw, negating any fair use defense.

ROSS Intelligence subsequently appealed the district court's findings, bringing the case before a Third Circuit panel in June 2026. Following several months of deliberation, the appellate court, with U.S. Circuit Judge Tamika Montgomery-Reeves—a Joe Biden appointee—writing for the panel, upheld the lower court's judgment. Judge Montgomery-Reeves distilled the complex arguments, stating that despite appearances, the case was "no more than an an ordinary copyright case," rather than solely concerning the future of AI legal technology.

The Third Circuit meticulously re-examined the originality of Westlaw's headnotes. Judge Montgomery-Reeves concurred that the editors' work exhibited the necessary creative spark for legal protection. She highlighted that each headnote was crafted to convey sufficient factual background and reasoning to stand independently, accurately reflecting the corresponding opinion passage without requiring reference to the full opinion or other headnotes. This careful crafting, she noted, demonstrated the minimal "creative spark" required for the headnotes to be considered original and thus copyrightable.

Addressing ROSS's fair use defense, the appellate panel reached conclusions consistent with the lower court. Judge Montgomery-Reeves characterized ROSS's use of the headnotes as "minimally transformative at best." While acknowledging that ROSS took an intermediate step of using the headnotes to train an AI program, which might suggest a slight difference in use, the undisputed evidence confirmed that ROSS leveraged these materials to train an AI for its legal-research platform, directly competing with Thomson Reuters. The panel further determined that ROSS's actions would negatively impact Westlaw's market value in the legal-research sector and diminish its potential revenue from licensing headnotes for AI training, further undermining the fair use argument.

This ruling from the Third Circuit creates a significant precedent regarding Third Circuit AI training copyright fair use, particularly for companies developing artificial intelligence solutions. It clarifies that merely using copyrighted material as training data for a competing product, even if an "intermediate step," is unlikely to be considered fair use if the original work possesses a "creative spark" and the new use is not substantially transformative. The court emphasized that ROSS aspired to be a direct competitor, utilizing Thomson Reuters’ materials for a highly similar purpose, which directly contributed to the finding of infringement.

The decision underscores the substantial legal AI intellectual property risk involved in developing AI models without proper data licensing. It signals to the industry that the use of proprietary content, such as Westlaw headnotes copyrightability, for AI training data without explicit permission or a robust fair use defense, can lead to costly litigation and adverse judgments. This outcome in Thomson Reuters v ROSS Intelligence reinforces the critical need for comprehensive data acquisition strategies that prioritize licensing or ensure genuinely transformative uses to avoid future AI copyright infringement precedent.

This Third Circuit ruling creates a significant precedent, clarifying that training AI models on copyrighted material, particularly for competitive purposes, does not constitute fair use. Lawyers must now advise clients developing AI on the heightened risk of copyright infringement and the critical need for comprehensive data licensing strategies to avoid legal challenges.

Source: Original reporting via Courthouse News