From Sidhant Tamrkar | Product & Market Analysis
Thomson Reuters v. Ross: Why 2,243 Headnotes Lost the AI Fair Use Appeal
On this page
The first US appellate ruling on AI training has landed, and it is narrower than the headlines. On 29 September 2026 the Third Circuit held that ROSS Intelligence infringed 2,243 Westlaw headnotes, and that copying them to train a legal research AI was not fair use. Thomson Reuters v. Ross is now the AI copyright lawsuit every vendor cites, but the court went out of its way to say generative AI is a different case.
Key takeaways
- The Third Circuit affirmed that training a competing AI tool on 2,243 Westlaw headnotes was not fair use. A unanimous panel, in an opinion by Judge Tamika Montgomery-Reeves dated 29 September 2026, found three of the four statutory factors against ROSS.
- The court recognised a licensing market for AI training data as a harm that counts. It called the market for licensing headnotes to train AI "rapidly developing" and held that ROSS took an opportunity that belonged to Thomson Reuters.
- Generative AI was expressly set aside, so this is a rule for retrieval and classification tools. The panel stressed that ROSS's system could not generate new expression and distinguished the Anthropic and OpenAI litigation on that ground.
- The case is not over, and the next move belongs to ROSS. ROSS said on 2 October 2026 that it will seek Supreme Court review, and the district court still has damages and the remaining headnotes to decide.
What the Third Circuit held in Thomson Reuters v. Ross
The court answered the two questions it was asked, and it answered both against ROSS. Westlaw headnotes are original enough to be protected by copyright. Using them to build training data for a rival legal research tool was not fair use.
The panel was Judges L. Felipe Restrepo, Tamika Montgomery-Reeves and Emil Bove. Judge Montgomery-Reeves wrote for a unanimous court. The opinion issued under seal on 29 September 2026 so the parties could propose redactions, and the full text was released in October.
The short answer
Is AI training fair use after Thomson Reuters v. Ross? Not when you copy a competitor's editorial work to build a tool that competes with that competitor. The Third Circuit held that ROSS's copying of 2,243 headnotes was commercial, minimally transformative and harmful to a real licensing market. It deliberately left generative AI for another case.
The two certified questions
This reached the appeals court early, through 28 U.S.C. 1292(b), which lets a trial judge certify a controlling legal question before final judgment. Judge Stephanos Bibas, a Third Circuit judge sitting by designation in Delaware, certified two questions in spring 2025. Were the headnotes and the West Key Number System original? And was ROSS's use fair use?
On originality, the panel applied the "extremely low" bar from Feist and found the editors' choices supplied the required spark of creativity. It rejected ROSS's argument that headnotes are the law itself with one line: "Headnotes are not law; judicial opinions are." It did not decide the Key Number System, which ROSS never briefed.
The framing sentence of the opinion is the one worth quoting to your board. "Under ROSS's framing, this case appears to concern the future of AI legal technology," the court wrote. "But appearances can be deceiving. In truth, this is no more than an ordinary copyright case." I read that as a deliberate signal to other courts. Do not treat this as the AI ruling; treat it as a competitor ruling.
The record: how 25,000 memos became an AI fair use appeal
The facts matter more than usual here, because the court leaned on them to keep the holding narrow. ROSS wanted to build an AI search tool for case law. It needed examples of legal questions paired with the passages from judicial opinions that answer them.
ROSS asked to license Westlaw content and was refused. It then hired LegalEase Solutions, a legal services vendor, to produce roughly 25,000 training memos. Each memo paired a question with 4 to 6 passages from opinions, graded from great to irrelevant. The drafters used Westlaw headnotes to frame the questions, which was the step that created the liability.
The finished tool searched a corpus of roughly 10 million judicial opinions, which nobody owns. A footnote noted that attempts to reach Westlaw through borrowed investor and student credentials weighed against ROSS, to the extent good faith matters at all. ROSS shut down its platform in 2021, citing the cost of this litigation.
The 2,830 headnote batch
Thomson Reuters did not have to prove that every memo copied a headnote. It focused on a batch of 2,830 headnotes that ROSS's own expert said closely resembled the memo questions. Judge Bibas concluded no reasonable juror could find that 2,243 of them were not copied, and the appeal accepted that.
That is 79% of the examined batch held copied as a matter of law, by my arithmetic. The other 587 were not resolved on summary judgment. The lesson for anyone building training data is blunt. Your opponent's expert does not need to find the copying, because yours might.
Factor by factor: how the AI fair use appeal changed the score
Fair use under section 107 of the Copyright Act weighs 4 factors. They are the purpose of the use, the nature of the work, the amount taken, and the effect on the market. The interesting part of this appeal is that the scoreboard moved between the two courts.
Judge Bibas split the factors 2 to 2 in February 2025 and decided the case on market harm. The Third Circuit made it 3 to 1. That is a stronger result for rights holders than the trial ruling, not merely a confirmation of it.
| Factor | District court, February 2025 | Third Circuit, September 2026 |
|---|---|---|
| 1. Purpose and character | Against ROSS: commercial, not transformative | Against ROSS: commercial and "minimally transformative, at best" |
| 2. Nature of the work | For ROSS: headnotes not very creative | Slightly for ROSS: published and "more factual than fictional" |
| 3. Amount used | For ROSS: output showed users no headnotes | Against ROSS: each headnote is its own work, copied whole |
| 4. Market effect | Against ROSS: a market substitute for Westlaw | Against ROSS: original market, value of the work and an AI licensing market |
| Result | 2 to 2, market harm decides. | 3 to 1 against fair use. |
Compiled from Venable's February 2025 alert on the district ruling and the Ballard Spahr and Haynes Boone alerts of 2 October 2026 on the appeal.
Factor three flipped, and that matters for training data
The third factor is where most AI defendants expected to win. A model trained on millions of documents uses a tiny share of any one rights holder's catalogue, and the output rarely reproduces the input. Bibas accepted the output argument for this factor.
The appeal rejected it. The panel treated each headnote as a separate work, copied in full, and said ROSS "took more than necessary" because the opinions themselves were available. My view is that this is the most portable part of the opinion. Any vendor arguing that its share of a rights holder's catalogue was small should expect to hear that every item is its own work.
Ease is not necessity: the line for AI training data
ROSS's strongest argument was intermediate copying. Courts have long allowed copying as a step toward something new, most famously in software cases. ROSS said the headnotes were a means to train a model, never a product shown to users.
The panel's answer fits in six words: "Unlike necessity, ease is not a justification for copying." The cases ROSS relied on involved copying that was required to reach unprotected material. ROSS had the unprotected material already, in the form of the opinions. Using headnotes was the easy route, not the only one.
Why Google v. Oracle did not help
The court distinguished Google v. Oracle, Sega v. Accolade and Sony v. Connectix as a single line. In each, copying code was necessary to reach functional elements copyright does not protect. That necessity is what made the copying fair.
On the first factor, the court also relied on Andy Warhol Foundation v. Goldsmith, the 2023 Supreme Court case that asks whether a use shares the original's purpose. Both Westlaw and ROSS used the headnotes to help lawyers find case law. Same purpose, same customers, so the transformation argument had little to work with. I think this test is right, and it is the one I would apply to any product brief before training begins.
The licensing market finding is the part B2B AI vendors should read twice
On the fourth factor the panel found harm three ways. It found harm to the original legal research market, to the value of the headnotes, and to a derivative market. The derivative market is the new part. The court wrote that "the market for licensing headnotes as text to train AI is rapidly developing," and that ROSS took Thomson Reuters's opportunity to enter it.
This is the sentence publishers have wanted from an appeals court since 2023. It turns "you could have licensed it" from a negotiating line into a legal harm. It sits next to deals already struck in the open market, which the tally of AI content licensing deals covers in detail.
A market that does not need a price list
Thomson Reuters had refused to license to ROSS. The court still treated the licensing market as real, and noted that Thomson Reuters trains its own AI search products on headnotes. Several commentators read that to mean a rights holder's refusal to license does not prove the market is absent.
That is uncomfortable for one large class of vendor: the one that builds on a competitor's editorial layer because the competitor will not sell it. The commercial context is not abstract either. Thomson Reuters is selling AI legal products against Harvey and others, a contest covered in the analysis of Harvey and Thomson Reuters in legal AI and in the review of RELX legal growth against Harvey.
| What you trained on | Position after the Third Circuit ruling |
|---|---|
| A competitor's summaries, tags or annotations, to build a rival tool | Inside the holding. Same purpose, same market, licensing harm recognised |
| Public domain primary material, such as opinions, statutes or filings, annotated by your own staff | Outside it. The court pointed to this as the route ROSS should have taken |
| Lawfully acquired works for a generative model | Expressly reserved. Bartz and the OpenAI litigation remain the live authorities |
| Customer data under a contract that permits training | Mostly a contract question. Check that the licence scope actually covers training |
| A contractor's training set of unknown origin | You inherit its choices. ROSS's liability came through a vendor's memos |
Zan Digital classification of the holding as reported. It is analysis, not legal advice, and the generative row depends on cases still being decided.
The briefs: ten amici for ROSS and none for Thomson Reuters
The appeal drew an unusually lopsided set of friends of the court. LawSites counted ten amicus briefs by November 2025, all supporting ROSS, and none supporting Thomson Reuters. The filers ranged from the Electronic Frontier Foundation and library groups to industry bodies and AI legal research startups.
Ten to zero did not move the panel. Amicus volume measures who feels threatened by a rule, not who is right about it. Anyone forecasting litigation outcomes from the weight of the briefing has just watched that method fail.
What the amici argued
The arguments fell into three groups. Library and open access groups argued that "no one can own the law" and that headnotes lack originality. A brief from copyright professors including Pamela Samuelson and Rebecca Tushnet called the use highly transformative and said the trial court misapplied Warhol.
A third group, including the Computer and Communications Industry Association and a coalition of legal AI startups, stressed how different ROSS's tool was from generative models. That last argument is worth noticing. The industry's own amici helped the court draw the generative line that now limits the ruling.
Late filings shaped the record too. In April 2026 the court asked for supplemental briefs on the fair use decision in ASTM v. UpCodes. In early September, ROSS sent the panel a letter citing the United States government's statement of interest on fair use in the OpenAI litigation, and Thomson Reuters replied on 9 September. A footnote in the opinion set generative AI aside partly in response.
What each remaining ruling would mean for B2B AI vendors
The appellate opinion binds federal courts in Delaware, New Jersey, Pennsylvania and the Virgin Islands. Everywhere else it is persuasive. The decision "creates continued uncertainty around the application of copyright law to AI model training," said Yar Chaikovsky, ROSS's lawyer at White & Case. ROSS intends "to seek review by the Supreme Court," Chaikovsky added.
| Next ruling | Legal effect | What it means for a B2B AI vendor |
|---|---|---|
| Rehearing denied, Supreme Court declines | Panel opinion stands as the only appellate answer. | Budget a licence for any non-generative tool trained on a rival's editorial content. |
| Full Third Circuit grants rehearing | Panel opinion vacated pending a new decision. | Uncertainty returns for at least a year. Do not unwind licences already signed. |
| Supreme Court takes it and affirms | National rule for competing, non-generative training. | Licensing market harm gains top court weight, and generative cases will be argued against it. |
| Supreme Court reverses on fair use | Intermediate copying for training gains footing. | Licence prices for training data come under pressure. Indemnities matter less. |
| Supreme Court reverses on originality | Headnote-style summaries fall outside copyright. | Editorial layers become cheaper to train on, and contract terms take over as the guardrail. |
Scenario framing is Zan Digital analysis. The Supreme Court grants a small fraction of petitions, so the first row is the base case.
The cert clock
Supreme Court Rule 13 gives a losing party 90 days from judgment to petition for review. Counted from 29 September, that runs to 28 December 2026. A timely petition for rehearing resets the clock to run from the order that resolves it.
LawSites reported on 12 October that ROSS had said in a court filing it would ask the full Third Circuit to reconsider. I could not confirm from the docket whether that petition was filed or decided by 1 November. Either way, expect the earliest Supreme Court decision on whether to hear the case in 2027. Meanwhile, the district court still has to resolve damages, possible copyright expiry on some headnotes, the unresolved headnotes and a tortious interference claim.
Where this reading of the AI copyright ruling is weakest
This post treats the opinion as a meaningful signal for B2B AI vendors. There are three good reasons to discount that, and I hold the first one strongly.
The facts were unusually bad for the defendant. ROSS sought a licence, was refused, used a contractor that worked from Westlaw content, and built a direct substitute. Most AI vendors cannot be described that way. A court facing a vendor that trained on a broad corpus for a different purpose could reach the opposite result while citing this opinion approvingly.
The record was settled by summary judgment. ROSS did not dispute the copying on appeal. Questions about how a model actually uses training data, raised in one amicus brief about vector embeddings, never got a factual hearing. A case with a full technical record could look different.
It is one panel in one circuit. District courts in California reached split views on generative training in Bartz v. Anthropic and Kadrey v. Meta, and no appeals court has resolved that split. The strongest counter-reading is that the licensing market theory is circular: any use becomes a lost licence once a rights holder decides to sell. That objection has real force. The panel answered it with evidence of a developing market, but the next court may want more than that.
My sources also carry a limit. I could not extract text from the opinion PDF during research, so the quotes here come from the opinion as reproduced in LawSites and several law firm alerts that agree on them. Read the opinion itself before relying on any single phrase in a filing.
Frequently asked questions
What did the Third Circuit decide in Thomson Reuters v. Ross?
On 29 September 2026 a unanimous Third Circuit panel affirmed that ROSS Intelligence infringed 2,243 Westlaw headnotes and that its use was not fair use. Judge Tamika Montgomery-Reeves wrote the opinion. The court held the headnotes original, weighed three of four fair use factors against ROSS, and expressly left generative AI training for future cases.
Is AI training on copyrighted material fair use?
There is no single answer yet. The Third Circuit held that training a non-generative legal research tool on a competitor's headnotes was not fair use. District courts in California have treated training generative models on lawfully acquired books more favourably. No appeals court has ruled on generative training, so the answer depends on the product, the source of the data and the court.
Does Thomson Reuters v. Ross apply to generative AI?
Not directly. The panel stressed that ROSS's system could not create new expression and only returned passages from existing judicial opinions. It distinguished the Bartz v. Anthropic and OpenAI litigation on that ground. Its reasoning on the licensing market and on copying each work in full will still be cited against generative AI developers, so it is relevant even where it is not binding.
Will the Supreme Court hear the Thomson Reuters v. Ross appeal?
ROSS said on 2 October 2026 that it intends to seek Supreme Court review. A petition is due 90 days after judgment, or after any rehearing petition is resolved. The Supreme Court grants a small share of petitions, but a first appellate ruling on AI training is the kind of question it sometimes takes. A decision on whether to hear it is unlikely before 2027.
Are Westlaw headnotes protected by copyright?
Yes, according to the Third Circuit. Applying the low originality standard from Feist, the panel found that the editorial choices in writing each headnote supplied enough creativity. It rejected the argument that headnotes are the law itself, writing that headnotes are not law and judicial opinions are. It did not decide whether the West Key Number System is protected, because ROSS did not brief that issue.
What does the Ross ruling mean for AI companies that license training data?
It strengthens the value of a licence. The court treated the market for licensing content to train AI as a real market that unlicensed copying can harm, even where the owner refused to deal. For vendors, that means a signed licence is now a stronger defence than a fair use argument for non-generative tools trained on a rival's editorial content.
Where to start: three checks on your own training data
Start with an inventory, not a legal memo. List every dataset your non-generative models were trained or fine-tuned on, and mark the ones that include another company's summaries, tags, annotations or curated labels. Those rows are the ones this ruling reaches, and you can usually build the list in an afternoon.
Next, read your data vendors' contracts for one clause: who warrants where the training data came from. ROSS's liability arrived through a contractor's memos. If your vendor does not warrant provenance and indemnify you for it, that gap is the one covered in the comparison of model provider terms. It will also surface in diligence, as set out in the review of AI representations in merger agreements.
Finally, put two dates in the calendar: 28 December 2026 for a cert petition if no rehearing petition was filed, and the Supreme Court's decision on whether to hear the case. For the wider set of live cases, the tracker of AI copyright cases and training data prices keeps the other half of this picture.
Related on the data access side
The same fight over who may train on whose content is playing out in API terms. See how SaaS vendors are rewriting API terms to wall off AI training.
References
- US Court of Appeals for the Third Circuit, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, opinion dated 29 September 2026. The primary source for the holding.
- LawSites, "No more than an ordinary copyright case": 3rd Circuit unseals its opinion, 12 October 2026. Used for the panel, quotes, the 2,830 and 2,243 figures, the Key Number System point and the rehearing statement.
- Haynes Boone, Third Circuit rejects fair use for ROSS's copying of Westlaw headnotes, 2 October 2026. Used for argument date, the 25,000 memos, LegalEase, the 10 million opinion corpus and the credentials footnote.
- Ballard Spahr, Third Circuit addresses fair use in AI training but leaves generative AI questions unresolved, October 2026. Used for the factor analysis and the generative AI distinction.
- LawSites, ROSS says it will ask Supreme Court to review 3rd Circuit ruling, 2 October 2026. Used for the White & Case statement.
- LawSites, "No one can own the law": amici come out in force to support ROSS, 4 November 2025. Used for the ten amicus briefs and their arguments.
- Venable, Judge rejects fair use defense in Thomson Reuters' AI copyright suit against Ross Intelligence, February 2025. Used for the district court's factor split and the reversal of its 2023 ruling.
The weakest part of this source base is that opinion quotes are taken from secondary reproductions that agree with each other, because the court's PDF could not be text-extracted during research. Procedural status, including any rehearing petition, is current as at 1 November 2026 and should be rechecked against the docket.
Related reading