· research

Thomson Reuters v. ROSS: headnotes and fair use

On February 11, 2025, the federal district court for the District of Delaware revisited its own ruling in the dispute between Thomson Reuters and ROSS Intelligence and sided with Thomson Reuters. The court found copying of 2,243 protected Westlaw headnotes proven and rejected the fair use defense. This is the first decision on the merits on whether a legal AI system may be trained on the editorial layer of someone else’s legal database, and it went against the developer. Below is an analysis based on the text of the decision and the appellate record.

What the case is

The plaintiff is Thomson Reuters Enterprise Centre GmbH, owner of Westlaw. The defendant is ROSS Intelligence, developer of an AI legal research system. Case No. 1:20-cv-00613 was heard by the federal district court for the District of Delaware, Judge Stephanos Bibas. The key ruling is the summary judgment of February 11, 2025, in which the judge reconsidered his earlier decision on the same questions.

What ROSS did

ROSS was building a machine-learning legal search system. It obtained its training data through a contractor in the form of so-called Bulk Memos - questions and answers drafted with reliance on Westlaw headnotes. Headnotes are editorial annotations to court decisions: brief statements of legal positions written by Thomson Reuters editors. Court decisions themselves are not protected by copyright in the US, but the editorial layer on top of them is.

The ROSS model was not generative: in response the user received excerpts from court decisions, not generated text and not the headnotes themselves.

What the court decided

QuestionCourt’s conclusion
Was there copyingYes, copying of 2,243 protected headnotes was proven
Nature of the useCommercial and not sufficiently transformative: editorial materials were used to create a competing legal search product
Fourth fair use factorThe weightiest: harm to the Westlaw market and to the potential market for licensing data for training
OutcomeFair use rejected, infringement found

The court gave the greatest weight to the fourth factor of the fair use analysis - the effect on the market. Thomson Reuters not only sells Westlaw but could potentially license its data for training models. ROSS’s product competed with Westlaw directly, and that outweighed the other considerations.

What the decision does not establish

The decision does not introduce a general rule for training large language models. Two circumstances limit its reach. First, the ROSS model was not generative, and the court did not assess a situation where training data dissolves into the weights of a model that generates new text. Second, ROSS was a direct competitor of Westlaw, and the market analysis was built around that competition.

The decision may be appealed, and the district court’s conclusions may still change.

Update. ROSS filed an interlocutory appeal; the case is in the Third Circuit as No. 25-2153. This entry will be updated after the appellate decision.

The case directly concerns the provenance of training data. Even if the end interface shows the user only court decisions, intermediate use of editorial annotations may lead to liability, especially when competing with the database owner. Four practical requirements follow from the decision:

  • separate court decisions in the public domain from the protected editorial layer on top of them;
  • keep an audit trail of the corpus: where each fragment came from and on what terms;
  • check the licenses of data suppliers, including contractors who prepare training sets;
  • test the model for memorization so that protected fragments are not reproduced in answers.

For Lexometrica this is a reference point when assembling a corpus of Russian court practice: the court acts themselves are open, but any annotations, digests and classifiers on top of them must be accounted for separately with a documented source.

Sources

← Journal