Skip to content
LegalTech Lab LawTechnologyBusiness

Blog

Copying Westlaw headnotes to train an AI was not fair use

On 29 September 2026 the United States Court of Appeals for the Third Circuit decided Thomson Reuters Enterprise Centre GmbH v ROSS Intelligence Inc, No. 25-2153. The panel affirmed a ruling that a legal research company infringed copyright when it used Westlaw editorial material to train an artificial intelligence system. LawNext calls it the first federal appellate decision to address fair use in the context of AI training.

What was copied

ROSS Intelligence built a legal search engine that answered plain-language questions with passages from judicial opinions. Its system was not generative. The opinion records that it created no new expression and returned only text from existing opinions. To train it, ROSS engaged a firm called LegalEase Solutions, which wrote about 25,000 training memos. The drafters used Westlaw headnotes, the short annotations that state the point of law an opinion decides, to frame the memo questions. The judgment concerns 2,243 of those headnotes.

Thomson Reuters sued in the District of Delaware. Judge Stephanos Bibas granted it partial summary judgment, then certified two questions for interlocutory appeal, an appeal brought before the case has finished. The questions were whether the headnotes are original, and whether the use was fair.

Originality of the headnotes

The court held the headnotes original. Editors choose which points of law matter and how to word them, and they are told to keep each headnote to 800 characters where possible. That is enough creative judgement to clear the low originality bar. ROSS argued that protecting headnotes hands Thomson Reuters control over the law itself. The court answered that headnotes are not law, and that judicial opinions remain free to publish. It also rejected the merger doctrine, the rule that denies copyright where an idea can be expressed in only one or a few ways, because a point of law can be put in many different ways.

The fair use factors

Fair use is the United States defence codified at 17 U.S.C. 107, and it turns on four statutory factors. On the first, purpose and character, the court found the use highly commercial and minimally transformative at best, because both sides used the headnotes to run a legal research platform. The second factor, the nature of the work, favoured ROSS, since headnotes are more factual than fictional. On the third, the amount taken, the court treated each headnote as a whole work and found that ROSS had copied the text of each one in full. Copying was not necessary, because ROSS could have worked from the freely available opinions instead. Judge Montgomery-Reeves wrote that “[u]nlike necessity, ease is not a justification for copying.”

The fourth factor, market effect, is the part of the judgment with the widest consequences. The court found harm in the legal research market, and also in a derivative market: licensing headnotes as AI training data. It held that this licensing market is rapidly developing, and that ROSS took away Thomson Reuters’ chance to enter it. Three factors against and one in favour meant the defence failed, and the court affirmed.

How far the ruling reaches

The panel limited its own reach. In footnote 7 it set this case apart from Bartz v Anthropic and Kadrey v Meta Platforms, on the ground that the ROSS system cannot generate original expression. So the decision does not settle whether training a generative model on protected works is fair use. Its practical force lies elsewhere. A rights holder who can show a real or developing market for licensing its material as training data now has an appellate authority to cite on the fourth factor. For lawyers advising either side of a training-data dispute, that is the paragraph to read first.

Sources