What Happened
The Third Circuit has rejected an AI company's fair use defense, holding that training a model on a competitor's copyrighted material constitutes copyright infringement. The decision, issued by a three-judge panel, could set a landmark precedent for AI training more broadly.
At the centre of the case are Thomson Reuters, owner of the Westlaw legal research platform, and ROSS Intelligence, a now-defunct AI legal research startup.
The Background
Thomson Reuters sued ROSS in 2020. The claim: ROSS took headnotes — short descriptions of important legal issues that appear before judicial opinions in Westlaw — and used them to train its AI. Westlaw had refused ROSS direct access, so the startup obtained the material through a third party, LegalEase Solutions.
ROSS's purpose, Reuters argued, was not legal research but to "rush out a competing product" built on Westlaw's headnotes.
In 2025, a federal judge in Philadelphia granted Reuters partial summary judgment and denied the same to ROSS, finding Westlaw's headnotes were copyrightable and that ROSS's use without permission violated copyright law.
ROSS had argued the headnotes were too similar to the underlying judicial opinions — which are not copyrightable — to qualify for protection. The lower court disagreed, holding that Westlaw's selection and arrangement of the notes constituted a "creative spark," comparable to a sculptor chiselling a raw block of marble. That spark made the non-verbatim headnotes original works.
And because ROSS's use was insufficiently transformative and aided its own market competition with Westlaw, fair use did not apply.
The Appeal
ROSS appealed, and the case reached a Third Circuit panel in June 2026.
Writing for the panel, Circuit Judge Tamika Montgomery-Reeves reduced the dispute to its simplest terms:
"Under ROSS' framing, this case appears to concern the future of AI legal technology. But appearances can be deceiving. In truth, this is no more than an ordinary copyright case."
On originality, she agreed with the lower court:
"Each headnote had to convey enough factual background and reasoning to stand on its own without reference to the opinion or any other headnote, while accurately reflecting the corresponding opinion passage. Because these decisions show the requisite minimal 'creative spark,' the headnotes are original."
On fair use, the panel was equally unmoved, calling ROSS's use "minimally transformative at best":
"ROSS took an intermediate step of using the headnotes to train an AI program, which arguably presents a slight degree of difference in use. But the undisputed evidence demonstrates that ROSS used the headnotes to train an AI program for the benefit of its legal-research platform."
Crucially, the panel also found that ROSS's use would harm Westlaw's value — both in the legal-research market and in the potential derivative market of licensing headnotes to train AI. That fourth fair-use factor (market harm) is where the ruling bites hardest for the AI industry.
"The undisputed evidence reflects that Thomson Reuters' materials possess a creative spark, and ROSS aspired to be a direct competitor by using them for a highly similar purpose. We therefore hold that Thomson Reuters' materials are copyrightable and that ROSS' use was not fair."
Judges L. Felipe Restrepo and Emil J. Bove III concurred.
Why It Matters
Three things stand out.
The "ordinary copyright case" framing is the strategy. By refusing to treat AI training as a categorically new kind of use, the panel applied settled doctrine — originality plus the four fair-use factors — instead of carving out an AI exception. That is precisely what rights holders have been hoping for.
Market harm now reaches the AI-licensing market. The panel recognised a derivative market for licensing content to train AI. If that reasoning travels, the argument that training is fair use because it does not reproduce the work for human readers weakens considerably — the harm is to a market the rights holder may itself want to sell into.
One circuit is not the last word. The Third Circuit binds federal courts in Delaware, New Jersey, Pennsylvania and the Virgin Islands. Other circuits — and the Supreme Court — are not bound. Expect this cited in every AI copyright dispute from here, and expect it to be appealed or distinguished.
For anyone building on scraped or licensed third-party data, the practical question sharpens: not "is the output transformative?" but "am I competing with the source — or destroying a market it could license?"