An Appeals Court Says Training on Westlaw Headnotes Was Not Fair Use

A three-judge panel of the Third Circuit has affirmed partial summary judgment for Thomson Reuters against ROSS Intelligence, holding that training a competing legal research product on Westlaw's headnotes was not fair use. The practical reading is narrow and worth stating plainly: this is a holding about one company copying another company's curated editorial layer to build a rival product. It is not a ruling that all AI training on copyrighted text is unlawful, and the court went out of its way to say so.

The opinion came from a panel of Judge Tamika Montgomery-Reeves, who wrote for the court, with Judges L. Felipe Restrepo and Emil J. Bove III concurring. ROSS's counsel, Yar Chaikovsky of White & Case, told Courthouse News the company intends to seek review by the Supreme Court, citing what it called continued uncertainty about how copyright law applies to AI model training.

The headnotes are the product, not the raw text

The first holding is about what a headnote is. Westlaw's headnotes are short summaries of points of law, written by the publisher's attorney-editors and attached to a case. They are not the judicial opinion. That distinction did most of the work in the appeal, because it forecloses the argument that a headnote is a bare copy of a government-authored document that nobody can own.

The panel found the headnotes carry enough expressive content to be copyrightable, describing the required contribution as a creative spark. That is a low bar, and it is deliberately low: the threshold for originality in copyright is famously undemanding. The consequence for anyone building a dataset out of a publisher's editorial apparatus is that the layers added by a human editor are the layers most likely to attract protection. Raw statute and raw opinion text sit in a different position from the annotations, summaries and cross-references a commercial publisher wraps around them.

The mechanics matter for anyone assembling a corpus from a publisher's database. A headnote is a sentence or two of editorial prose attached to a point of law, and it is the thing that makes a large case database usable, which is why it exists as a product rather than a byproduct. Westlaw's value is not only that it holds the opinions, but that an attorney-editor has read each one and written a short indexed summary that a practitioner can scan. ROSS wanted the summaries. The court held that wanting the part of the database that reflects human editorial judgment is different from wanting the judicial text underneath it.

The court's reasoning did not turn on how much text was taken. Once the headnotes were accepted as original expression, the analysis moved to the purpose of the copying and the effect on the market for the original, and those two questions carried the appeal.

A second reason the framing of the claim mattered: the suit was about training material, not model output. The theory was that ROSS copied headnotes on the way into its system. Whether a model's generated answer reproduces protected expression is a different question, governed by different evidence, and the panel did not need to reach it. That separation between input and output is the fault line running through the current wave of copyright litigation against AI developers, and different circuits are now deciding different halves of it.

The word the court chose was "minimally"

Fair use turns on four statutory factors, and the first is the purpose and character of the use. A use that transforms the original by giving it a new meaning or message gets more latitude than one that substitutes for it. The panel held that ROSS's use was minimally transformative at best.

That phrasing matters more than the outcome of the case. A defendant arguing that training is inherently transformative is asserting that the copy serves a different purpose from the original, so the first factor should favour it. The Third Circuit did not adopt that position. It also did not reject it wholesale. On the record before it, the court saw a competing product built from the same material, and treated the use as closer to substitution than to commentary.

The fourth factor, market harm, was where the opinion landed hardest. The panel cited harm to two markets: the existing market for legal research, and what it described as the potential derivative market of licensing headnotes to train AI. The second of those is the one to watch. Copyright law treats a market for licensing a work as a market even when it is new, and a plaintiff does not have to show that licences were already being sold to establish harm. If the licensing market is potential rather than actual, the holding is really a statement that the right to license training data belongs to the publisher by default.

An ordinary copyright case, according to the court

The panel resisted the framing both sides brought to the appeal. Writing for the court, Montgomery-Reeves put it this way: under ROSS' framing, this case appears to concern the future of AI legal technology, but appearances can be deceiving, and in truth it is no more than an ordinary copyright case.

Read that as a signal about scope. The district court had produced a revised fair use ruling, and the appeal asked the Third Circuit to say something larger about AI training. The panel declined and resolved the dispute on the material in front of it. Counsel for ROSS is asking the Supreme Court to take the broader question anyway.

Where the sources disagree about this opinion

One detail in the coverage does not reconcile. IPWatchdog's headline describes the opinion as sealed, while Courthouse News quotes substantial passages and attributes them to the opinion. Both cannot be straightforwardly true of the same document. The full text is published as the opinion PDF, which is the primary source for this article; the document was identified as a search result and not opened in this pass.

The resolution is unconfirmed. It is possible the opinion was filed under seal and later unsealed, that a portion of the record is sealed while the opinion is not, or that one outlet's headline is imprecise. This article relies on the passages Courthouse News quotes and on its account of the holding, and flags the disagreement rather than picking a side. Readers should treat the seal question as open. The sealing of an opinion would itself be unusual, because a published appellate decision is normally a public document, and the Third Circuit's own docket is the place to confirm it.

What this does and does not decide

It does not decide that every training run over copyrighted text infringes. The holding is tied to headnotes, to a competitor's product, and to evidence about two specific markets.

It does not decide the parallel questions in the district courts, where other plaintiffs are pursuing different theories over different kinds of works. A ruling about a legal publisher's annotations is not a ruling about song lyrics or novels.

It does decide that the transformative-use defence is not automatic. A party that wants to rely on it now has to explain what its model does with the material that is new, and has to answer the market-harm question for licensing markets that the plaintiff says it intended to enter.

One practical asymmetry follows. A defendant that scrapes a large web corpus can argue that no single publisher's licensing market is affected. A defendant that takes a specific competitor's curated dataset cannot make that argument, because the dataset is the competitor's product. The theory scales down as easily as it scales up, and it is strongest exactly where the material is most obviously curated.

It also decides something about sequencing. The court resolved a case about training data without needing to reach model outputs, because the claim was about the copies made on the way in. That is a reminder that training and output are separate legal questions, and a win on one does not carry the other.

And it decides that an appellate court will not treat AI as a reason to rewrite the doctrine. The panel applied the ordinary four factors to an ordinary commercial dispute and reached an unremarkable result. For a training programme that had been assuming the technology itself was the argument, that is the most useful thing in the opinion.

What to do before your next training run

If you are compiling a corpus, separate three things in your records: material you generated yourself, material under a licence that permits the use, and material whose provenance you cannot establish. Only the third category is the problem, and it is the one that tends to grow quietly inside a scraped dataset.

Read the terms of the commercial sources you use. A legal publisher's subscription agreement is a contract question before it is a copyright question, and the licence terms will usually decide the dispute faster than fair use will.

Keep the licence file with the dataset. If a source later offers a training licence, the date you acquired the material and the terms in force at that time are the facts that will matter, and reconstructing them two years later is expensive.

Finally, treat the derivative-market theory as the live risk. The part of this opinion most likely to be quoted in the next complaint is the sentence about licensing headnotes to train AI, because it turns a new use into an old market and hands the plaintiff a theory that does not depend on showing lost subscriptions.