Summary

On September 29, 2026, a unanimous Third Circuit panel held that ROSS Intelligence did not make fair use of Thomson Reuters' Westlaw headnotes when it used them to train a competing AI legal search tool, and that the headnotes are original enough to be protected by copyright. It is the first federal appellate decision on fair use in AI training. The holding is narrow because the tool was non-generative and built to replace Westlaw, but the court's reasoning gives both AI developers and owners of curated content a test they can apply to their own data today.

The Event

The Court of Appeals for the Third Circuit filed its opinion under seal on Tuesday, September 29, and released it on September 30, according to Reuters (which is owned by Thomson Reuters, the plaintiff). The case is Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153. Judge Tamika Montgomery-Reeves wrote for the three-judge panel and framed the appeal as "no more than an ordinary copyright case."

The panel affirmed a February 2025 partial summary judgment by Judge Stephanos Bibas, a Third Circuit judge sitting by designation in the District of Delaware. As MSK's client alert recounts, Judge Bibas had sent nearly every issue to a jury in September 2023, then reversed himself shortly before trial. His revised ruling found infringement of 2,243 headnotes, rejected fair use, and certified the originality and fair use questions for interlocutory appeal.

The facts, as summarised by FindLaw, MSK and Ballard Spahr: Thomson Reuters refused to license Westlaw to ROSS. ROSS then hired LegalEase Solutions, which prepared about 25,000 "Bulk Memos" whose legal questions were drafted from Westlaw headnotes and paired with ranked passages from judicial opinions. ROSS converted the memos into training data for a tool that answered plain-language questions by returning passages from existing opinions. ROSS shut down its platform in 2021, citing litigation costs, but pursued the appeal.

The court's holdings, per the MSK and Ballard Spahr analyses:

  • Originality. Choosing which points of law to summarise and how to word them clears the low bar set by Feist. The panel left open whether headnotes that quote an opinion verbatim are protectable, because none of the 2,243 did.
  • Factor one (purpose). The use was commercial and "minimally transformative, at best." Training was an intermediate step, but the court judged purpose by the finished product: both parties used the headnotes to build a legal research platform.
  • Factor two (nature). Favoured ROSS, given the headnotes' largely factual content.
  • Factor three (amount). Favoured Thomson Reuters. Unlike Judge Bibas, the panel did not consider whether headnotes appeared in ROSS's output. It asked only whether ROSS took more than its purpose required. Because the opinions themselves were freely available, copying headnotes was not necessary. "Unlike necessity, ease is not a justification for copying." FindLaw reports the court also rejected ROSS's argument that it used only about 0.08% of Westlaw's headnotes, since it copied whole headnotes.
  • Factor four (market). Favoured Thomson Reuters on two grounds: harm in the existing legal research market, and harm to a market for licensing headnotes as AI training data, which the court called "rapidly developing."

Context

The only other merits rulings on AI training came from two Northern District of California judges in June 2025. In Bartz v. Anthropic and Kadrey v. Meta, Judges Alsup and Chhabria each found that training a general-purpose large language model on books was transformative. MSK notes that Bartz later produced a settlement of about $1.5 billion on pirated-books class claims, and that on September 1, 2026 the Justice Department filed a statement of interest in In re OpenAI arguing that LLM training is transformative.

The Third Circuit addressed those cases directly. In footnote 7 it distinguished Bartz on two grounds: ROSS's tool "cannot generate original expression," and it was trained to create a commercial substitute for Westlaw. MSK observes that the panel did not say which ground carries the weight. Reuters reports that the court also stated the Justice Department's concerns in the OpenAI case "do not apply here."

The court also rejected ROSS's reliance on intermediate-copying precedent (Sega v. Accolade, Sony v. Connectix, Google v. Oracle). In those cases copying was needed to reach unprotected functional material. ROSS could have worked from the public opinions.

Implications

The ruling does not decide whether generative AI training is fair use. It highlights considerations for a fact-specific assessment: the trained product's purpose, the justification for the amount copied, and effects on existing and potential licensing markets. In this case, ROSS's minimally transformative, substitutive use and its access to public opinions meant that its necessity argument did not support fair use. The court weighed all four statutory factors together.

For AI developers, a practical starting point is to examine why the chosen training material is needed. Run a provenance check on each training set and ask a specific question: does it contain third-party editorial layers (summaries, annotations, classifications, curated labels) when the underlying public source would have served? If so, rebuild from the primary material or license the layer. This matters most for vertical tools that serve the same users as the source. A patent analytics tool trained on a vendor's written abstracts or classification annotations, rather than on the published documents, sits closer to ROSS than to Bartz. Also review internal strategy documents: the panel relied on ROSS's own admission that it aimed to replace Westlaw.

For content owners, the ruling strengthens the value of curated editorial work. Inventory assets that reflect editorial selection and wording, confirm US copyright registrations where those works are commercialised (the district court relied on the presumption of validity from Thomson Reuters' registrations), and document any training-data licensing programme. A real, priced licensing offer is evidence of the market the court protected.

For Korean and other Asian companies, no Korean party was involved, but many sit on both sides. Korean technology firms building AI products for the US market may face questions about the justification for copying and market substitution as part of a fact-specific fair-use assessment. The same groups often own curated databases, technical documentation or industry content that they could license as training data, and there the ruling helps them. The two positions call for different files: a provenance record defending your own training choices, and a licensing record supporting your claim to a market in your own content.

Outlook

The open question is the one footnote 7 leaves unresolved: a general-purpose generative model that is transformative but also competes with the works it learned from. Watch the generative AI cases, including In re OpenAI, and any appellate review of the California rulings. Until an appellate court addresses that scenario, Bartz and Kadrey remain district-level authority, while ROSS is now circuit-level authority within the Third Circuit.

In the meantime, hold a hedge rather than wait. Segregate training data by source and licence status to help identify a contested set, stop using it in future training, or assess licensing options. Segregation alone does not remove a dataset's influence from an already trained model. Whether that influence can be removed without full retraining depends on the model and training architecture and on the availability of a suitable, validated unlearning method. Where a product competes with a data source, treat licensing as the default rather than relying on a transformative-use argument. And in vendor and data-acquisition contracts, add warranties on how training material was produced, since ROSS's exposure arose from memos that a contractor drafted from Westlaw.

Sources