Part of the AI Regulation News hub.
The Third Circuit, on interlocutory appeal, affirmed partial summary judgment for Thomson Reuters, holding that 2,243 Westlaw headnotes are original and that ROSS Intelligence's copying of them to train a non-generative legal search AI was not fair use
The court took pains to say this was "no more than an ordinary copyright case". ROSS's tool returned passages from existing opinions and generated nothing, and the panel set it apart from the generative-model disputes in Bartz and In re: OpenAI on exactly that point.
Bottom line: A ruling of the Third Circuit on an interlocutory appeal, filed Sept. 29, 2026. It affirms the District Court's order granting Thomson Reuters partial summary judgment on two certified questions: whether the headnotes are original, and whether ROSS's use was fair. It is not a final judgment in the case, it binds no court outside the Third Circuit, and it concerns a tool the court says was not generative AI.
Who this affects: IP and copyright litigators, legal-tech vendors that train search or retrieval models on third-party editorial content, and law-firm innovation and procurement counsel who license or build legal research tools.
Issue date: Opinion filed Sept. 29, 2026; argued June 11, 2026. The opinion sets no compliance date and does not describe further proceedings in the District Court.
What changed: On this interlocutory appeal from partial summary judgment, the panel held that the 2,243 Westlaw headnotes at issue are original enough for copyright protection, and that ROSS's copying of them into AI training memos for a competing legal-research platform was not fair use. Factor two favoured ROSS; factors one, three and four went against it.
Analysis: On our reading, the decisive facts were substitution and ease. ROSS built a product to replace Westlaw, and it used the headnotes because they were an easy way to frame questions, although the underlying opinions were free to copy. The court said ease is not a justification for copying. Those facts, not AI training as such, carried the result.
Primary sources: Opinion of the Court, Thomson Reuters v. ROSS Intelligence, 3d Cir. No. 25-2153 (PDF)
- Instrument (EN)
- Opinion of the Court, Thomson Reuters Enterprise Centre GmbH; West Publishing Corp v. ROSS Intelligence Inc, Appellant, No. 25-2153
- Authority
- United States Court of Appeals for the Third Circuit. Panel: Restrepo, Montgomery-Reeves and Bove, Circuit Judges; opinion by Judge Montgomery-Reeves
- Jurisdiction
- United States, Third Circuit. Appeal from the U.S. District Court for the District of Delaware, No. 1:20-cv-00613 (Judge Stephanos Bibas, sitting by designation)
- Status
- Decided. Interlocutory appeal under 28 U.S.C. 1292(b); the District Court's partial summary judgment order is affirmed
- Bindingness
- A court decision. It binds no court outside the Third Circuit and does not decide whether training generative AI models is fair use
- Issue date / next deadline
- Filed Sept. 29, 2026; argued June 11, 2026. No deadline or remand terms are stated in the opinion
- Scope
- 2,243 Westlaw headnotes that do not copy opinion text verbatim. The Key Number System was not addressed because ROSS forfeited that issue
- Primary source
- https://www2.ca3.uscourts.gov/opinarch/252153p.pdf
What exactly did the Third Circuit decide?
It decided two certified questions on an interlocutory appeal, and nothing broader. The District Court had granted Thomson Reuters partial summary judgment, holding as a matter of law that ROSS's use of 2,243 Westlaw headnotes infringed Thomson Reuters's copyright and was not fair use. At the parties' request it then certified for interlocutory review whether the headnotes and the Key Number System are original, and whether ROSS's use of the headnotes was fair.
The court set its own frame before reaching either question. On page 2 of the opinion it wrote: "Under ROSS's 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". The phrase tells the reader that the AI context did not change the analysis: the panel treated the dispute as two standard copyright questions, originality and fair use, decided on this record.
The panel answered both questions against ROSS on the headnotes and affirmed the partial summary judgment on the materials at issue. In the court's words: "Thus, we will affirm the District Court's order granting Thomson Reuters partial summary judgment." The tool involved was not generative. The opinion says: "ROSS's AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions."
Two limits sit in the footnotes. ROSS never challenged the originality of the Key Number System in its briefing, so the court did not address that forfeited issue (footnote 4). And the District Court's suggestion that headnotes quoting opinions verbatim might still be original was left unanswered, because summary judgment covered only the 2,243 headnotes that do not copy opinion text verbatim (footnote 5). For this appeal ROSS also did not dispute that LegalEase and Morae Global copied Westlaw headnotes, or that their actions are attributable to ROSS (footnote 2).
Thomson Reuters's suit also alleged tortious interference with contract. The opinion does not say what happens to that claim, or to the rest of the case, on return to the District Court, and we do not speculate.
Was ROSS's tool generative AI?
No, and the opinion says so in plain terms: "ROSS's AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions."
As the court describes it, ROSS built a search engine that answered plain-language legal questions with relevant passages drawn from a bank of approximately ten million uncopyrighted judicial opinions. To train it, ROSS engaged LegalEase Solutions, which wrote approximately 25,000 legal memos, each pairing a legal question with four to six opinion passages graded great, good, topical or irrelevant. Memo drafters used thousands of Westlaw headnotes to frame the questions, because the headnotes provided, in the record's words quoted by the court, "an easy way [to] fram[e] questions."
Footnote 7 draws the line expressly. The court notes a Department of Justice statement of interest filed Sept. 1, 2026 in In re: OpenAI, Inc. Copyright Infringement Litigation (S.D.N.Y.), which relied on Bartz v. Anthropic to argue that training a large language model is transformative. The panel said the concerns in that separate case "do not apply here", and that "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression". Anyone citing this decision for a proposition about generative-model training is reading past that footnote.
Why are the headnotes copyrightable when the opinions are not?
Because, on the court's account, an editor's choices about which points of law to include and how to word them clear the low originality bar. All 2,243 headnotes at issue were held to have "some creative spark". Each had to convey enough background and reasoning to stand alone while accurately reflecting the corresponding passage.
The opinions themselves stay free. Answering ROSS's argument that protection would give Thomson Reuters a monopoly over the law, the court wrote: "Headnotes are not law; judicial opinions are." Thomson Reuters, the opinion records, disclaims copyright in work prepared by government officers as part of their official duties, such as the opinions underlying the headnotes.
The court rejected the merger doctrine argument, holding there are many ways to express points of law from an opinion, and distinguished Matthew Bender v. West (2d Cir. 1998) as a case about party names and parallel citations, not independently composed headnotes. Footnote 10 adds that the court affirms the District Court's holding that each individual headnote is a copyrightable work.
How did the four fair use factors come out?
Three of four went against ROSS, on the undisputed record at partial summary judgment. Factor one, purpose and character: ROSS's use was commercial, and because both companies used the headnotes to build legal research platforms, the use shared the same ultimate purpose and was "minimally transformative, at best". The court found the intermediate step of AI training "arguably presents a slight degree of difference in use" and no more.
ROSS relied on Authors Guild v. Google, where the Second Circuit found a highly transformative purpose in a feature for searching and reading small snippets of books, a feature that may also have driven users to buy the original books. The court distinguished it on purpose: "No such transformative purpose is present here." ROSS's platform helps users find opinions, something Thomson Reuters already does with its headnotes, and "as ROSS admits, it aims to replace Westlaw." ROSS also relied on three intermediate-copying cases, Google v. Oracle, Sega v. Accolade and Sony v. Connectix. In each, the court said, copying was necessary to access the unprotected functional aspects of computer code. Here, "ROSS does not need to copy Thomson Reuters's headnotes to access the underlying unprotected information."
Factor two favoured ROSS, because the headnotes were published and are more factual than fictional. Factor three went against it: ROSS copied whole headnotes, each of which the court treats as a complete work, without a transformative purpose to justify that much. Factor four went against it as well. ROSS's copying diminished the headnotes' value as a draw to Westlaw, and the court said widespread copying of this kind would harm Westlaw's position in the legal-research platform market; and it held that ROSS usurped Thomson Reuters's opportunity in what the evidence showed, on the court's account, to be a "rapidly developing" market for licensing headnotes as AI training text.
On public benefit, the court found no evidence for ROSS's claims that the ruling would halt AI development or harm national security, and said that does not give ROSS "carte blanche to violate copyright law merely because it incorporates AI." The court also recorded, in footnote 9, undisputed evidence that ROSS at times acted in bad faith in trying to access Westlaw, and said that to the extent good faith remains relevant it weighs against fair use.
What should legal-tech vendors and firm counsel take from it?
This part is our interpretation, not the court's. The holding is fact-bound: a commercial, non-generative search tool, trained on a competitor's editorial annotations, to build a substitute for that competitor's product, when the public-domain source material was available. Each of those facts appears in the court's reasoning on factors one, three and four.
For a vendor, the practical question the opinion invites is provenance. Where training data includes headnotes, summaries or other editorial layers written by a competitor, this panel treated using them for convenience, rather than necessity, as weighing against fair use. For firm innovation and procurement counsel, the same question belongs in diligence on any legal research tool a firm licenses: what was it trained on, and was that material licensed. The opinion itself imposes no duty on law firms or tool buyers.
The decision says nothing about whether training a generative model on copyrighted text is fair use. Courts outside the Third Circuit are not bound by it, and within the Third Circuit its reach to different facts is a question for later cases.
What we did not verify
What we opened: a pdftotext extraction of the Third Circuit's opinion in No. 25-2153, saved in our evidence file as the text of www2.ca3.uscourts.gov/opinarch/252153p.pdf. We read the whole extraction, from the caption to the list of counsel and amici. Every fact and quotation on this page comes from it; quotations are folded to straight ASCII quotes and rejoined across line-end hyphenation, with no words changed.
What we did not open: the District Court's partial summary judgment opinion, the certification order, the parties' briefs, the amicus briefs, the appendix the court cites, the DOJ statement of interest in In re: OpenAI, and the Bartz and Kadrey decisions mentioned in footnote 7. We did not check the court's docket for a mandate, rehearing petition or later proceedings.
What we refuse to claim: we do not say this decision resolves fair use for AI training generally, that it applies to generative AI, that it binds courts in other circuits, or that judicial opinions are copyrightable; the court said the opposite on that last point. We do not describe a remand or next steps, because the opinion does not. We have not verified from the text whether the court designated the opinion precedential, and we do not label it either way. We also note, without resolving it, that the factor-three discussion refers to ROSS copying "the 25,000 Westlaw-written headnotes", while the holding and conclusion concern 2,243 headnotes; we report the 2,243 figure as the scope of the judgment.
We make no claim about how this ruling ranks or compares in novelty against other appellate decisions on copying for AI training.
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
On an interlocutory appeal, the Third Circuit affirmed that copying a competitor's editorial headnotes to train a commercial, non-generative legal search tool built to replace that competitor was not fair use. The transferable point is provenance: on this record, convenience did not justify copying when the public-domain source was available. The ruling does not decide generative-AI training.
Source File
https://www2.ca3.uscourts.gov/opinarch/252153p.pdf
Open the Third Circuit opinion in No. 25-2153 and confirm five things: the filing date of Sept. 29, 2026 in the caption, the court's description of the case as "no more than an ordinary copyright case" in the second paragraph of the opinion, at page 2, the holding and affirmance at page 2, the statement at page 5 that ROSS's AI was not generative AI, and the conclusion at page 28 that the 2,243 headnotes are original and ROSS's use was not fair.
The undisputed evidence reflects that Thomson Reuters's 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's materials are copyrightable and that ROSS's use was not fair. ยท Opinion of the Court, Thomson Reuters v. ROSS Intelligence, 3d Cir. No. 25-2153, filed Sept. 29, 2026
FAQ
Does this ruling decide whether training generative AI on copyrighted works is fair use?
No. The opinion says: "ROSS's AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions." In footnote 7 it said the concerns raised about large language model training in In re: OpenAI do not apply here, and distinguished the AI models in Bartz and In re: OpenAI.
Did the court say judicial opinions are copyrightable?
No. It wrote that "Headnotes are not law; judicial opinions are", and that opinions remain free for publication to all. What it held original were 2,243 Westlaw headnotes written by Thomson Reuters's editors that do not copy opinion text verbatim.
Is the case over?
The opinion does not say. It resolves an interlocutory appeal of two certified questions and affirms the District Court's partial summary judgment order. It does not describe further proceedings, and the suit also alleged tortious interference with contract.
Which fair use factors favoured ROSS?
Only factor two, the nature of the work, because the headnotes were published and are more factual than fictional. The court held that factors one, three and four weighed against fair use, and that ROSS's use, weighed as a whole, was not fair.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.