Thomson Reuters v. ROSS on Appeal: What the First Appellate AI Fair Use Ruling Means in the Sixth Circuit and for the Music Industry

On September 29, 2026, a unanimous Third Circuit panel affirmed Judge Stephanos Bibas's ruling that ROSS Intelligence infringed Thomson Reuters's copyrights by using Westlaw headnotes to train an AI legal research tool, and that the use was not fair. The opinion, written by Judge Tamika Montgomery-Reeves for a panel that included Judges Restrepo and Bove, was filed under seal and released to the public on September 30.

Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. 2026), is the first decision from any federal court of appeals on whether training an AI system on copyrighted material is fair use. We covered the district court's ruling, along with the 2025 decisions in Bartz v. Anthropic and Kadrey v. Meta, in our earlier article on AI training and fair use. This piece covers what the Third Circuit added, what it left open, how the reasoning lines up with Sixth Circuit law that governs our Tennessee clients, and what it means for the music industry, which filed one of the more pointed amicus briefs in the appeal.

The facts in brief

ROSS set out to build a legal search engine that would answer plain-language questions with passages from judicial opinions. To teach its model which passages answer which questions, ROSS hired LegalEase Solutions to write about 25,000 training memos, each pairing a legal question with four to six opinion excerpts graded from "great" to "irrelevant." LegalEase's drafters used Westlaw headnotes to frame the questions, and the "great" answers were usually the passages Westlaw linked to the headnote.

ROSS's AI was not generative. It returned excerpts from public-domain opinions and never showed users a headnote. But ROSS marketed itself directly against Westlaw at comparable prices, and some firms switched. Thomson Reuters sued in 2020, and ROSS shut down in 2021.

Judge Bibas, sitting by designation in Delaware, granted Thomson Reuters partial summary judgment on 2,243 headnotes that the memo questions tracked closely while departing from the opinion text. He then certified two questions under 28 U.S.C. § 1292(b): whether the headnotes are original, and whether ROSS's use was fair. ROSS did not brief the originality of the Key Number System, so the Third Circuit treated that issue as forfeited.

Holding one: the headnotes are original

The court called the Feist originality standard an "extremely low" bar and held that every one of the 2,243 headnotes cleared it. Westlaw's editors decide which points of law matter and how to state them so that each headnote stands on its own while still accurately reflecting the opinion. That is enough of a "creative spark."

The court rejected each of ROSS's counterarguments:

• Monopoly over the law. Headnotes are not law; judicial opinions are, and opinions remain free for anyone to copy. The court leaned on Callaghan v. Myers (1888) and Georgia v. Public.Resource.Org (2020), both of which recognize a private reporter's copyright in headnotes and similar explanatory material.
• Merger. Because there are many ways to express a point of law (the court pointed to Lexis's different headnotes for the same opinions), the idea and expression do not merge. The court analogized to its own decision holding that a banana costume can be designed many different ways.
• Matthew Bender v. West. The Second Circuit's 1998 decision denied protection to West's parallel citations and party-name arrangements because industry convention dictated them. That reasoning does not reach headnotes, which the Second Circuit itself described as independently composed.

One question remains open. Judge Bibas suggested that even a headnote quoting the opinion verbatim might be original through selection. The Third Circuit called that "an interesting question" and declined to answer it, because the judgment covered only headnotes that do not copy opinion text word for word.

Holding two: ROSS's use was not fair

The court found that factors one, three, and four weighed against fair use and factor two weighed slightly in ROSS's favor.

Factor one (purpose and character). Applying Andy Warhol Foundation v. Goldsmith, the court asked whether ROSS's use had a purpose meaningfully different from Thomson Reuters's. It did not. Thomson Reuters uses headnotes to help researchers find relevant opinions; ROSS used them to build a platform that helps researchers find relevant opinions. Training the model was "an intermediate step" that the court said "arguably presents a slight degree of difference in use," but the ultimate purpose was the same, which made the use "minimally transformative, at best." Combined with ROSS's commercial aims, factor one weighed against fair use.

The court distinguished the two lines of cases ROSS relied on most heavily:

• Authors Guild v. Google involved a search tool that pointed readers toward books they might then buy. ROSS's product was designed to replace Westlaw, not send users to it.
• Google v. Oracle, Sega v. Accolade, and Sony v. Connectix involved copying computer code that was necessary to reach unprotected functional elements and achieve interoperability. ROSS had the public-domain opinions and could have written its training memos from them. It used the headnotes because they were easier. In the opinion's most quotable line: "Unlike necessity, ease is not a justification for copying."

In a footnote, the court also noted undisputed evidence of bad faith, including attempts to access Westlaw with an investor's credentials and through a student account, and said that to the extent good faith still matters after Oracle, it cut against ROSS.

Factor two (nature of the work). Headnotes are published and largely factual, so this factor favored ROSS. As usual, it did little work.

Factor three (amount and substantiality). ROSS argued it took only 0.08% of Westlaw's 28 million headnotes. The court treated each headnote as its own copyrighted work, so ROSS copied entire works, and with no transformative purpose and no need to copy, it took more than necessary. (The factor-three discussion refers to ROSS copying "the entire text of the 25,000 Westlaw-written headnotes," which does not match the 2,243 headnotes actually at issue. It does not affect the outcome, but expect it to come up if ROSS seeks rehearing.)

Factor four (market effect). The court found harm on several levels:

• Value of the work. Even if no standalone market for headnotes exists, Thomson Reuters uses them to draw subscribers to Westlaw. Under the Third Circuit's Video Pipeline decision (movie trailers), harm to a work's value as a draw for the owner's core product counts.
• The original market. ROSS copied the headnotes to compete in the legal research platform market, and widespread copying of that kind would hurt Westlaw.
• The derivative market for AI training data. The court found that the market for licensing headnotes as AI training data "is rapidly developing," that Thomson Reuters uses its own headnotes to train its AI products, and that Thomson Reuters's decision not to license them to others does not mean the market does not exist. ROSS usurped that market by taking the headnotes without permission.

The court gave little weight to ROSS's claimed public benefits. The opinions were already free, ROSS charged Westlaw-level prices, and ROSS offered no evidence that the ruling would halt AI development or affect national security.

What the court did not decide

The most important footnote in the opinion is footnote 7. The court noted the Department of Justice's September 1, 2026 statement of interest in In re OpenAI, Inc. Copyright Infringement Litigation (S.D.N.Y.), which relies on Bartz v. Anthropic to argue that training a large language model capable of generating original responses is transformative and does not create "substitutive competition." The court said those concerns "do not apply here" because ROSS's system could not generate original expression and was built as a commercial substitute for Westlaw. It also pointedly observed that DOJ did not weigh in on ROSS.

So the decision is narrower than some headlines suggest. It does not hold that AI training is infringement, and it does not resolve how fair use applies to general-purpose generative models. What it does establish at the appellate level is a framework: courts may look past the training step to what the model is built to do; copying a rightsholder's material to build a direct competitor is unlikely to be transformative; "we needed it" is a real argument but "it was easier" is not; and a developing market for training licenses counts under factor four even if the plaintiff has not entered it yet.

A few procedural caveats. The appeal was interlocutory, damages and the tortious interference claim remain in the district court, and ROSS no longer operates. ROSS could still seek rehearing en banc or certiorari, and commentators have already criticized the opinion for comparing the parties' businesses rather than their uses of the work. Expect the decision to be cited heavily, but also distinguished heavily.

Implications for the Sixth Circuit

Federal courts in Tennessee, Kentucky, Ohio, and Michigan are not bound by the Third Circuit. But ROSS is now the only appellate decision on point, and as far as we can tell, no Sixth Circuit appeal currently presents the AI training question. District judges in this circuit facing the issue will reach for ROSS first, then ask whether Sixth Circuit precedent points the same way. On most of the key moves, it does.

Licensing markets count under factor four. In Princeton University Press v. Michigan Document Services, Inc., 99 F.3d 1381 (6th Cir. 1996) (en banc), the Sixth Circuit held that a commercial copy shop's unlicensed coursepacks were not fair use, in large part because the copying deprived publishers of permission fees in an established licensing market. The en banc court also looked to the copy shop's own commercial reason for copying rather than the educational end use its customers claimed. ROSS's treatment of the AI training-data licensing market, and its focus on the copier's real commercial objective, follows the same logic. A Sixth Circuit court is unlikely to accept the argument that a training-data market is illusory simply because the plaintiff has not yet sold into it.

Same-purpose commercial copying fails. Balsley v. LFP, Inc., 691 F.3d 747 (6th Cir. 2012), rejected fair use where a magazine republished a photo for essentially the same purpose it was originally published. Zomba Enterprises v. Panorama Records, 491 F.3d 574 (6th Cir. 2007), rejected fair use for karaoke discs that copied entire songs for a commercial product. Both predate Warhol but anticipate its emphasis on purpose and commerciality, which is exactly what drove factor one in ROSS.

Necessity and interoperability. The Sixth Circuit has been receptive to copying dictated by function and compatibility. In Lexmark International v. Static Control Components, 387 F.3d 522 (6th Cir. 2004), the court held that a short printer program was likely unprotectable because compatibility and efficiency constraints dictated its form. That is the same line ROSS drew between Oracle and Sega on one side and ROSS's convenience-driven copying on the other. AI defendants in this circuit who can build a genuine necessity record (no practical way to obtain the functional, unprotected information without copying) will have something to work with. Defendants who copied curated editorial content because it was the cheapest training set available will not.

Originality and filtration. The Sixth Circuit has denied protection to rigidly systematic material, as in ATC Distribution Group v. Whatever It Takes Transmissions & Parts, Inc., 402 F.3d 700 (6th Cir. 2005) (parts numbers and catalog arrangement). That is the closest local analog to the arguments ROSS made about headnotes, and ROSS explains why written summaries of legal points are different. Under Kohus v. Mariol, 328 F.3d 848 (6th Cir. 2003), Sixth Circuit courts also filter out unprotectable elements before comparing works. In AI cases built on public-domain or factual source material, expect defendants to press filtration hard, and expect plaintiffs to do what Thomson Reuters did: show that the copied material tracks the plaintiff's expression rather than the underlying public material.

Where the cases will be litigated. The most prominent AI music case, Concord Music Group v. Anthropic, was filed in the Middle District of Tennessee in October 2023 and transferred to the Northern District of California in 2024, where Bartz and Kadrey were decided. That transfer is a reminder that venue fights in these cases are also fights over which circuit's law will develop first. Rightsholders based in Nashville who have a viable basis for venue here should think carefully before conceding a transfer.

Implications for the music industry

The Recording Industry Association of America and the National Music Publishers Association filed an amicus brief supporting Thomson Reuters (Doc. 93, filed November 25, 2025). Their members are plaintiffs in suits against AI developers over sound recordings and compositions, including the Suno, Udio, and Anthropic cases, and they used ROSS to argue two points that matter far more to music than to legal research.

Argument one: purpose is judged by what the model is built to do. ROSS and its amici argued that its purpose was purely technical: converting memos into machine-readable training data. The RIAA and NMPA argued that this truncates the analysis to the first step of the system and ignores why the copying happened. The Third Circuit agreed in substance. It treated training as an intermediate step and held that ROSS's use shared Westlaw's "ultimate purpose." That framing is the most useful thing in the opinion for music plaintiffs, because the standard defense in music AI cases is that training only extracts statistical patterns and is therefore transformative regardless of what the model later produces.

Argument two: market harm does not depend on infringing outputs. The brief argued that AI can flood a market with non-infringing substitutes, citing Kadrey's "market dilution" discussion, the Copyright Office's May 2025 report on generative AI training, and industry data, including Deezer's report that it receives more than 50,000 fully AI-generated tracks a day, over a third of daily deliveries. It also pointed to licensing deals that did not exist when the music suits were filed: UMG and WMG with Stability AI and Udio, all three majors with KLAY, and Kobalt with ElevenLabs.

The court went part of the way. It held that ROSS harmed Thomson Reuters even though ROSS's outputs contained no headnotes, and it recognized a developing market for AI training licenses. The music industry's licensing record is far more developed than Thomson Reuters's was, which had no outside licenses at all. If a nascent, unexploited training-data market was enough in ROSS, the documented label and publisher deals described in the brief should be more than enough to establish a cognizable market in the music cases.

What the court did not adopt. The brief asked the court to hold that training a model to produce competing substitutes "can never be fair use." The court did not say that. It decided the case on direct substitution, not the broader market-dilution theory, and footnote 7 expressly set generative AI aside. Suno, Udio, and other music AI defendants are generative, and they will cite footnote 7 and Bartz to argue that ROSS says nothing about them.

Our read. The generative/non-generative label is less important than footnote 7 makes it sound. The DOJ's argument that the court distinguished turned on the absence of "substitutive competition," and the court distinguished it precisely because ROSS was building a substitute. A general-purpose language model that can discuss song lyrics (the Concord v. Anthropic posture) looks more like Bartz. A music generator trained on recordings to produce recordings for the same streaming, sync, and advertising markets looks much more like ROSS: same ultimate purpose, commercial, and competing for the same listeners and licensees. The "ease is not necessity" holding also matters. Licensed catalogs for training now exist, which makes it harder for a music AI company to claim it had no practical alternative to copying everything available online.

The music cases will still turn on their records. The summary judgment ruling on fair use in the labels' case against Suno in Massachusetts has reportedly slipped to early 2027, so a district court's application of ROSS to a generative music model is still months away.

A Tennessee note. Copyright is not the only tool. Tennessee's ELVIS Act, effective July 1, 2024, extended the state's right of publicity to cover a person's voice, including AI-generated simulations. ROSS does not affect it, and for artists whose concern is a soundalike output rather than training on their recordings, it may be the more direct claim.

Practical takeaways

For rightsholders and content companies:

• Document your AI licensing activity, including inbound inquiries, deals, internal AI uses of your own content, and comparable industry deals. ROSS shows that the training-data market can count under factor four even if you have not licensed yet, but evidence still wins these cases.
• Register your works, including individual items you would want treated as separate works. The court's treatment of each headnote as a whole work drove factor three.
• Enforce your terms of service and access controls. ROSS's use of borrowed credentials appeared in the court's good-faith analysis, and Thomson Reuters also brought a tortious interference claim.
• Review existing licenses and artist, writer, and publishing agreements for whether they grant, reserve, or are silent on AI training rights.

For AI developers and the businesses that hire them:

• Know where your training data came from and keep records of it. Provenance is now central in Bartz, Kadrey, and ROSS.
• Be most cautious with curated editorial or creative content from a company whose product yours will compete with. That is the ROSS fact pattern.
• If you are relying on necessity, build the record before litigation: why the unprotected information could not be obtained another way.
• If you buy or deploy third-party AI tools, negotiate training-data representations and IP indemnities in the vendor agreement.

The bottom line

ROSS is a narrow decision on unusual facts: a non-generative tool, a direct competitor, and a defendant that copied curated editorial content because it was convenient. But it is the first appellate word on AI training and fair use, and its framework (look past the training step to the product, distinguish necessity from convenience, and count training-data licensing as a real market) is consistent with how the Sixth Circuit has long approached fair use. For music companies in Nashville and elsewhere, the opinion strengthens the arguments that matter most while leaving the hardest question, generative AI, for the next court.

Our copyright and entertainment and media attorneys advise rightsholders, music companies, and technology businesses on AI licensing, training-data disputes, and copyright litigation.

---

This article is for informational purposes only and does not constitute legal advice. The law in this area is developing quickly, and the analysis above reflects the decision as issued on September 29, 2026.