Third Circuit Rejects Fair Use for AI Trained on Westlaw Headnotes

The court held two things. First, the 2,243 Westlaw headnotes are original enough to be protected by copyright. Second, ROSS's use of those headnotes to train a competing AI legal-research tool was not fair use.

The court went out of its way to call this "no more than an ordinary copyright case." But the opinion supplies the first appellate roadmap for how courts will evaluate AI training data, and it draws a line between non-generative tools built to replace a competitor and the generative AI models now being litigated across the country.

Background

Westlaw organizes judicial opinions through its Key Number System and headnotes, which are short editorial summaries of the points of law in an opinion. Thomson Reuters's editors draft headnotes under detailed guidelines: each should be concise, ideally under 800 characters, and able to stand on its own for a reader who has not seen the opinion.

ROSS, a startup founded by three University of Toronto computer science students, built an AI search engine that answered plain-language legal questions by returning passages from roughly ten million public-domain judicial opinions. Importantly, ROSS's tool was not generative; it did not create new text.

To train the system, ROSS hired LegalEase Solutions to prepare about 25,000 training memos. Each memo posed a legal question and ranked four to six opinion passages as great, good, topical, or irrelevant. LegalEase drafters built the questions from Westlaw headnotes because, in their words, headnotes offered an easy way to frame questions. The "great" passages were most often the very passages Westlaw linked to the headnote. ROSS advertised itself as a Westlaw alternative at comparable prices, and some law firms switched.

Headnotes Are Copyrightable

Applying the "extremely low" originality bar of Feist Publications v. Rural Telephone,² the court held that each of the 2,243 headnotes has the required creative spark. Editors chose which points of law mattered and how to express them concisely while still standing alone. The court found support in Callaghan v. Myers,³ and Georgia v. Public.Resource.Org,⁴ both recognizing that a private reporter's headnotes can be protected.

The court rejected each of ROSS's counterarguments:

• Monopoly over the law. Headnotes are not law; judicial opinions are, and they remain free for all to use.
• Merger. There are many ways to express a point of law, as LexisNexis's own differently worded headnotes demonstrate.
• Matthew Bender v. West.⁵ The Second Circuit denied protection to West's parallel citations and party-name arrangements, which industry convention dictated. It described headnotes as independently composed.

The court also affirmed that each headnote is a separately copyrightable work, a point that later mattered under the third fair use factor. It left open whether headnotes that quote an opinion verbatim are protected.

AI Training Was Not Fair Use

Three of the four statutory factors weighed against ROSS. ROSS bore the burden on this affirmative defense.

Necessity versus ease. ROSS relied on the intermediate-copying cases: Google LLC v. Oracle Am., Inc.,⁶ Sega Enters. Ltd. v. Accolade, Inc.,⁷ and Sony Comput. Ent., Inc. v. Connectix Corp.⁸ The court distinguished them because copying in those cases was necessary to reach unprotected functional elements. ROSS had the public-domain opinions and could have built its training memos from them. It used headnotes because it was easier, and the court said bluntly that "[u]nlike necessity, ease is not a justification for copying."

Search is not enough. Authors Guild v. Google, Inc.,⁹ did not help either. Google Books served a different function and could steer readers back to the original. ROSS was built to replace Westlaw.

Percentages do not save the copier. ROSS argued it used only 0.08 percent of Westlaw's 28 million headnotes. The court held that qualitative importance and the lack of necessity controlled.

The licensing market counts. The court found the market for licensing headnotes as AI training data to be rapidly developing. Thomson Reuters already uses its headnotes to train its own AI products. That it had not yet licensed them to others did not make the market illusory.

Conduct matters. In a footnote, the court noted evidence that ROSS "attempt[ed] to access Westlaw with law-firm investor credentials," and a "ROSS employee inquired about a Westlaw account under the guise of a solo practitioner." A ROSS employee gained access using a student account. To the extent good faith remains relevant, it weighed against ROSS.

What About Generative AI?

Footnote 7 may be the most-cited part of the opinion. The court acknowledged the Department of Justice's September 1, 2026 statement of interest in In re: OpenAI, Inc. Copyright Infringement Litig.¹⁰ Relying on Bartz v. Anthropic PBC,¹¹ the DOJ argued that training a large language model capable of generating original responses is transformative and does not create substitutive competition.

The Third Circuit said those concerns do not apply here. ROSS's tool generated no new expression, and ROSS trained it specifically to build a commercial substitute for Westlaw. The court also observed that the DOJ chose not to weigh in on ROSS.

The result is a decision that does not resolve the generative AI cases but frames them. Developers of generative models will emphasize the distinction. Rights holders will emphasize the market-substitution and licensing-market reasoning, which does not depend on whether the model generates text.

Practical Takeaways

For companies building or fine-tuning AI:

• Audit training-data provenance. Know where every dataset came from, including data prepared by vendors and subcontractors. ROSS did not contest that LegalEase's copying was attributable to it.
   
• Go to the source when you can. If unprotected material (public records, statutes, opinions, your own data) can do the job, use it. Convenience will not justify copying a competitor's editorial layer.
   
• Mind your competitive purpose. Fair use is weakest when the trained product competes in the same market as the source work. Internal marketing that positions a product as a substitute will be evidence.
   
• Respect terms of service. Shortcuts in obtaining data, such as borrowed or misrepresented credentials, can color the entire fair use analysis.
   
• Budget for licensing. The court treated AI training licenses as a real market. Expect rights holders to demand them.

For content owners and publishers:

• Register and document. Thomson Reuters's registration and editorial guidelines helped establish originality.
   
• Develop the licensing market. Using your own content to train internal AI, and offering licenses to others, strengthens a factor-four argument.
   
• Review your terms of use to address scraping and AI training expressly.

For businesses using AI tools: Ask vendors how their models were trained and seek indemnities covering training-data infringement claims.

What's Next

Because the appeal was interlocutory, the case returns to the District of Delaware for the remaining issues, and ROSS may seek rehearing en banc or Supreme Court review. Meanwhile, the generative AI cases continue in other circuits. Until those courts rule, ROSS is the only appellate guidance available, and it should inform every AI training-data decision now. If you have any questions, please contact Edward D. Lanquist, Nicole Imhof or any member of the Intellectual Property Group.

---

¹ No. 25-2153 (3d Cir. Sept. 29, 2026) (precedential).

² 499 U.S. 340, 345–46 (1991).

³ 128 U.S. 617 (1888). 

⁴ 590 U.S. 255, 265 (2020). 

⁵ 158 F.3d 674 (2d Cir. 1998).

⁶ 593 U.S. 1, 6–7 (2021). 

⁷ 977 F.2d 1510, 1514 (9th Cir. 1992).

⁸ 203 F.3d 596, 599, 606–08 (9th Cir. 2000).

⁹ 804 F.3d 202 (2d Cir. 2015).

¹⁰ No. 1:25-md-3143 (S.D.N.Y. Sept. 1, 2026).

¹¹ 787 F. Supp. 3d 1007, 1014 (N.D. Cal. 2025).