Full profile
As of July 4, 2026, the copyright impact of Thomson Reuters v. ROSS Intelligence is still an appellate question, not settled law. The district court’s February 2025 summary judgment ruling went sharply against ROSS Intelligence on infringement and fair use, finding infringement as to 2,243 of 2,830 Westlaw headnotes analyzed and rejecting ROSS’s fair-use defense after weighing the statutory factors.[1] The Third Circuit heard argument on June 11, 2026, before Judges L. Felipe Restrepo, Arianna Freeman Montgomery-Reeves, and Emil Bove, and no decision has issued yet.[2]
That posture matters more than the label “landmark AI case.” The eventual Third Circuit decision may become the first federal appellate ruling on AI training and fair use, but the district court ruling was deliberately confined. Judge Stephanos Bibas wrote that “only non-generative AI is before me today.”[3] For lawyers evaluating generative drafting, summarization, research, or contract-analysis tools, that sentence is not decorative. It is the boundary line.
Where the Case Stands Now
The operative status is simple enough to state and easy enough to overread: ROSS lost important issues at summary judgment, appealed, argued the case, and is waiting for the Third Circuit. The CourtListener docket remains the cleanest place to monitor filings and orders rather than relying on recycled descriptions of the ruling.[4]
For a general counsel or legal-ops lead, the pending appeal creates two different risks. One is legal risk: an appellate court may affirm, reverse, or narrow the district court’s treatment of AI training and fair use. The other is procurement risk: vendors, commentators, and internal stakeholders may speak as if the appeal has already answered more than it has.
The case is also not a direct referendum on every AI system used in law practice. ROSS built a legal research product that used machine learning to return legal answers, not a generative AI system that produces new text in the manner now associated with large language models. That difference does not make the case unimportant. It makes the possible appellate holding more precise.
What Thomson Reuters Said ROSS Copied
Thomson Reuters and West Publishing claimed that ROSS used copyrighted Westlaw editorial material, including headnotes, to train and develop a competing legal research system. The February 2025 ruling treated a large group of asserted headnotes as protectable and found that ROSS infringed 2,243 of the 2,830 headnotes the court analyzed.[1]
ROSS’s story has always carried an access-to-law undertone. It presented itself as a smaller legal technology company trying to build a more accessible research product in a market dominated by incumbents. That context explains why the case attracted amici concerned with competition, legal information access, and beneficial uses of AI. It does not, by itself, answer whether the copied material was original enough for copyright protection or whether the use was fair.
The practical consequence arrived before the appellate merits did. ROSS announced in December 2020 that it would shut down, saying the litigation had become too costly and that the suit was “heavy for ROSS to bear.”[5] That shutdown is not proof that Thomson Reuters was right on the merits. It is a reminder that copyright litigation can decide the fate of a legal technology company long before an appellate court supplies a rule.
The February 2025 Ruling Turned on Fair Use More Than on AI Hype
The district court’s fair-use analysis is the part procurement and risk teams should read slowly. The court did not say that every AI training use is infringing. It also did not say that AI training is categorically fair. Instead, it walked through the four statutory fair-use factors and split them two and two: two factors favored ROSS, and two favored Thomson Reuters.[1]
| Fair-use factor | District court treatment | Why it mattered operationally |
|---|---|---|
| Factor one: purpose and character | Favored Thomson Reuters | The court did not accept ROSS’s use as sufficiently transformative in the way ROSS needed. |
| Factor two: nature of the copyrighted work | Favored ROSS | The headnotes related to legal information, but that did not carry the outcome. |
| Factor three: amount and substantiality | Favored ROSS | The court’s treatment of amount did not overcome its view of purpose and market effect. |
| Factor four: effect on the market | Favored Thomson Reuters | The court treated competitive market harm as central to the fair-use balance. |
The important move was weighting. Although the factor count was even, the court gave practical weight to factor one and factor four, the two factors that favored Thomson Reuters.[1] That is why shorthand descriptions such as “two factors for each side” can mislead. Fair use is not a box score.
On factor one, the court rejected ROSS’s argument that using Westlaw headnotes to train a legal research system was transformative enough to support fair use. The product ROSS was building mattered: it was not merely analyzing text for a remote technical purpose; it was developing a legal research competitor in the same broad market where Westlaw operates.[1]
On factor four, the court treated market effect as more than speculative. Thomson Reuters argued that ROSS’s use threatened markets for Westlaw content and legal research products, and the court credited market-harm concerns in the fair-use balance.[1] For legal buyers, that is the part that travels most directly into vendor diligence. If a tool is trained on proprietary legal editorial content, the question is not only whether the output looks useful. It is whether the vendor can explain the provenance, license basis, and market-substitution risk of the training material.
ROSS has tried to reframe the scale of the dispute on appeal. It argues that the 2,243 headnotes represent only 0.076% of total Westlaw headnotes.[6] That number should be read as ROSS’s appellate characterization, not as a neutral judicial measurement of harm or substantiality. Thomson Reuters disputes the framing, and the district court did not make fair use turn on that percentage.
Why the “Non-Generative AI” Limitation Does Real Work
The line “only non-generative AI is before me today” limits what the February 2025 ruling can responsibly be used to prove.[3] It does not make the case irrelevant to generative AI. Training-data disputes, fair-use arguments, and market-substitution theories overlap across AI systems. But overlap is not identity.
Generative AI tools raise additional questions about output, memorization, substitution, retrieval-augmented generation, and professional use that were not resolved by the ROSS ruling. A legal research tool that uses machine learning to return answers from legal materials is closer to the facts of this case than a general-purpose model used to draft briefs, summarize depositions, or generate contract clauses. For a more practice-oriented distinction, see General-Purpose vs. Legal-Native AI: What Every Lawyer Needs to Know About the Risks.
That distinction should change how legal teams ask questions. A vendor selling a non-generative legal research product may face one set of copyright-provenance issues. A vendor selling a generative drafting assistant may face those issues plus additional concerns about generated text, hallucinated citations, confidentiality, supervision, and unauthorized practice boundaries. The ROSS appeal may inform those conversations, but it will not finish them.
The Appeal: Argument Heard, UpCodes Briefing Added, Decision Pending
At the June 11, 2026 Third Circuit argument, reporting from Stevens & Lee described a panel that appeared skeptical of ROSS on both copyrightability of Westlaw headnotes and transformative use.[2] Oral argument skepticism is not a holding. Judges test positions, and panels sometimes decide cases more narrowly than questioning suggests. Still, for risk purposes, the argument did not remove the possibility that the district court’s core result will survive.
The Third Circuit also ordered supplemental briefing on its April 2026 UpCodes decision, where a building-code AI startup was found likely to have made fair use.[2] That development gives ROSS a live comparison point: another technology company, another information-heavy legal or regulatory setting, and an appellate court willing to recognize fair use in a data-intensive product context.
UpCodes does not make reversal predictable. The materials described in the ROSS dispute involve Westlaw editorial headnotes and a legal research competitor; UpCodes involved a different record and different claimed works. The supplemental briefing matters because it may shape the Third Circuit’s articulation of fair use, not because it supplies a one-line answer.
That is the appellate question worth watching: not whether the Third Circuit announces that AI training is always lawful or always unlawful, but whether it writes a rule tied to competitive substitution, the nature of the copied legal editorial content, and the specific non-generative system ROSS built. A narrow affirmance would still be consequential. A broad affirmance would be more disruptive. A reversal could become an important defense-side citation, but the record would still matter.
The Amicus Line-Up Shows the Stakes, Not the Answer
The appeal has drawn support from groups and companies with very different institutional interests. EFF, copyright law professors, Free Law Project, Paxton AI, and Trellis Research have supported ROSS, while Disney, the Recording Industry Association of America, the Association of American Publishers, and standards organizations have supported Thomson Reuters.[7]
That coalition map is useful, but only up to a point. ROSS’s supporters emphasize access to law, competition, and the risk that copyright claims can entrench information incumbents. Thomson Reuters’s supporters emphasize control over protected expression, licensing markets, and the danger of allowing AI developers to build competing products from proprietary content. Both sets of concerns are real. Neither substitutes for the Third Circuit’s analysis of this record.
For readers tracking the broader litigation environment, ROSS now sits beside other district-court AI fair-use rulings such as Bartz and Kadrey. Those comparisons are useful, but they should not be flattened into a single “AI copyright law” takeaway. The better working file is a live map of cases, orders, and appeals; our AI Copyright and Fair Use Research guide collects tools and trackers for that purpose, and our companion analysis of the 2026 fair-use divide treats the district-court landscape separately.
What Legal Teams Should Do With the Case Before the Decision
The safest current reading is disciplined and narrow. The February 2025 ruling is a serious copyright loss for ROSS and a serious fair-use win for Thomson Reuters at the district-court level. The June 2026 argument is complete. The Third Circuit decision is pending. Until that decision issues, legal teams should not treat the case as appellate law.
- Monitor the Third Circuit docket and the eventual opinion language, not just headlines about a “landmark AI case.”
- Separate non-generative legal research systems from generative AI tools when assessing copyright and professional responsibility risk.
- Ask vendors for training-data provenance, licensing representations, indemnity terms, and any known disputes involving source materials.
- Treat district-court AI fair-use rulings as important signals, not settled appellate rules.
- Include lawyers responsible for professional responsibility, confidentiality, and supervision in AI procurement review, not only IT and innovation teams.
The case also creates an awkward procurement issue for buyers evaluating Thomson Reuters’s own AI products. The existence of this litigation does not answer whether Westlaw CoCounsel, Lexis+ AI, or another legal-native product is appropriate for a particular legal department or firm. It does make copyright provenance and vendor representations fair questions in any serious evaluation. For that product-side comparison, see Westlaw CoCounsel vs. Lexis+ AI.
When the Third Circuit rules, the first reading should be procedural before it is philosophical: what claims were before the court, what facts mattered, how broadly the panel wrote, and whether it repeated or relaxed the district court’s non-generative limitation. That is the difference between useful legal risk analysis and turning a pending appeal into a slogan.
References
- Court rules on AI fair use in Thomson Reuters v. ROSS Intelligence — Reed Smith
- Third Circuit Poised to Decide Whether AI Training on Copyrighted Material Is Fair Use — Stevens & Lee / JD Supra, June 29, 2026
- Reuters-ROSS Court Ruling on AI Copyright Fair Use — Davis Wright Tremaine, February 2025
- Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. — CourtListener
- Announcement — ROSS Intelligence, December 2020
- ROSS Intelligence Appeals Originality, Fair Use Rulings in Thomson Reuters AI Legal Tool Lawsuit — IPWatchdog, April 15, 2025
- Protecting Access to Law and Beneficial Uses of AI — Electronic Frontier Foundation, September 2025
Comments
Join the discussion with an anonymous comment.