Training a competing search tool on copied headnotes is not fair use, and the market for training data counts

2025-02-11Copyright

Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. Ross Intelligence Inc., No. 1:20-cv-613-SB (D. Del. Feb. 11, 2025) (Bibas, Circuit Judge, sitting by designation), revising Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., 694 F. Supp. 3d 467 (D. Del. 2023)

Ross Intelligence built a legal research search engine using artificial intelligence. Refused a Westlaw license, Ross bought about 25,000 Bulk Memos of legal questions and answers from LegalEase, whose lawyers wrote them from Westlaw headnotes, and trained its tool on them. Judge Bibas, sitting by designation, revised his 2023 summary judgment ruling. He held that the headnotes and the Key Number System are original, comparing an editor's choice of what to keep from an opinion to a sculptor's choice of what to cut away, granted summary judgment that Ross copied 2,243 headnotes, leaving copyright validity and expiry for trial, and rejected the innocent infringement, misuse, merger and scenes a faire defenses. On fair use under Section 107 of Title 17 of the United States Code, the first factor favored Thomson Reuters because the use was commercial and not transformative: Ross took the headnotes to build a competing legal research tool, and the intermediate-copying cases did not apply because they concern computer code and copying needed to reach unprotected functional elements. The second and third factors favored Ross. The fourth factor favored Thomson Reuters because Ross meant to build a market substitute and its use affected a potential derivative market in data for training legal artificial intelligence tools. The court noted that only non-generative artificial intelligence was before it.

What it changesThe source of training data is now a liability question the patent program cannot ignore, because the patent specification is where many companies describe their training data in public. An application stating that a model was trained on a named commercial database or on a competitor's content is published evidence of copying, available to the content owner years before any lawsuit. Describe training data in specifications by type, structure and the properties that matter to the technique, name a source only after its license status is confirmed, and keep the provenance record in the confidential invention file. The court's recognition of a market for training data also puts licensing into the cost of every model built on third-party content, which makes a training method that needs less licensed data a cost advantage worth claiming. The decision is a district court summary judgment on a non-generative tool.