The U.S. Court of Appeals for the Third Circuit has affirmed that Ross Intelligence’s use of Westlaw’s editorial headnotes to train a competing AI legal-research platform did not qualify as fair use. The ruling, which affirmed the lower court’s decision, carries a specific practical message for AI developers building products on top of structured editorial data: the source and function of your training material carry legal weight, not just what your users see in the final output.
The Headnote Is Not the Opinion
The distinction between a public-domain judicial opinion and the editorial summary a human editor wrote to accompany it is the conceptual core of this case.
Judicial opinions are public-domain material. Anyone can read them, copy them, and build tools around them. Westlaw’s headnotes are something different.
A headnote is a short editorial summary placed alongside a court opinion. Think of it as the difference between a raw dataset and the proprietary index a company builds to make that data searchable: the underlying data may be free, but the index is not. Westlaw’s editors write those summaries by selecting which legal issue matters, deciding what factual and analytical context to include, and expressing the point concisely enough to help a lawyer locate the relevant passage quickly.
That editorial judgment, according to the Third Circuit, supplies the “creative spark” copyright law requires. The court applied copyright’s low originality threshold and found sufficient independent expression in the editors’ choices of selection and presentation.
Ross Intelligence partnered with LegalEase Solutions to build training materials connecting legal questions to passages from judicial opinions. Those materials incorporated Westlaw headnotes, and the headnotes were not incidental: they were functional to the product Ross was building.
“The undisputed evidence reflects that Thomson Reuters’ materials possess a creative spark, and ROSS aspired to be a direct competitor by using them for a highly similar purpose.” Judge Tamika Montgomery-Reeves, U.S. Court of Appeals for the Third Circuit.
The court’s reasoning draws a clear line. A public-domain source does not automatically place a privately created summary or classification in the public domain.
Internal Use Did Not Save Ross
Ross’s argument that users never saw the headnotes failed because the court focused on what the copied material enabled the competing product to do.
Ross argued that it used the headnotes only internally, as training data, and that the material was not presented to users in the same form as Westlaw headnotes. That argument failed.
The court described the use as “minimally transformative” at best. The technical process differed from Westlaw’s presentation of headnotes to subscribers, but the commercial purpose was highly similar: helping users conduct legal research. Changing the format, or keeping the training material out of user-facing output, does not by itself make a use transformative.
“But appearances can be deceiving. In truth, this is no more than an ordinary copyright case,” Judge Montgomery-Reeves wrote, according to Bloomberg Law’s account of the opinion.
The court also considered harm to two markets. Ross’s copying threatened the existing market for Westlaw’s legal-research services. It also threatened a potential market in which Thomson Reuters could license its headnotes directly to AI developers. Both considerations weighed against fair use.
What This Ruling Does Not Cover
The Third Circuit’s decision is fact-specific and does not establish a blanket prohibition on training AI systems with copyrighted material.
The court’s analysis depended on particular circumstances: protectable editorial content, direct commercial competition, copying of a significant portion of the relevant headnotes, and a product targeting substantially the same market function. Those facts shaped every step of the fair-use analysis.
The Electronic Frontier Foundation and other digital-rights groups argued for Ross, warning that the ruling could give rights holders excessive control over functional or factual material. That concern remains relevant as courts encounter training-data disputes in other contexts.
Ross Intelligence is now defunct. The ruling may nonetheless be relevant to developers assembling training datasets from structured information sources such as legal databases, news archives, annotated scientific literature, or commercial data services.
Summaries, annotations, labels, and metadata built around public records may be protectable if they contain sufficient original expression. The practical signal from this case is direct: what matters to a court is what the training material enabled your product to do, and whether that product competes with the source it came from.
