It’s not a great idea to get sued by one of the oldest legal research platforms in the universe.
The short version: The Third Circuit held that fair use did not protect an AI company that copied copyrighted Westlaw headnotes to build a competing legal research product. The decision turns on the headnotes’ originality, ROSS’s substitutive commercial purpose, the copying of entire headnotes, and the harm to existing and emerging AI-licensing markets.
What the Third Circuit Decided
The legal boundary between AI training and copyright infringement reached a pivotal turning point on September 29, 2026. That’s when the U.S. Court of Appeals for the Third Circuit held that ROSS Intelligence could not rely on fair use after copying Westlaw’s proprietary editorial headnotes to train a competing legal search platform. In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. (Appeal No. 25-2153) (ROSS), the court affirmed the District of Delaware’s grant of partial summary judgment for Thomson Reuters and held that the copied headnotes were protected by valid copyrights.
Background: Scraping Editorial Value Under the Guise of “AI Innovation”
Thomson Reuters operates Westlaw, organizing judicial opinions via its proprietary Key Number System and human-edited headnotes. While judicial opinions themselves are uncopyrightable public-domain materials, Thomson Reuters drafts original, concise headnotes that synthesize points of law with enough context to stand independently.
ROSS developed an AI-driven legal search engine powered by natural language queries designed to retrieve relevant case law excerpts. Crucially, ROSS’s model was not generative in the creative sense; it did not confect novel paragraphs, but rather retrieved passages from ten million public-domain judicial opinions.
To teach its model which judicial passages answered which legal questions, ROSS engaged LegalEase Solutions to draft approximately 25,000 question-and-answer training memoranda. Rather than drafting questions from scratch or parsing opinions manually, memo-writers signed up for Westlaw accounts and then copied over 2,000 Westlaw headnotes to frame questions because they provided an “easy way” to capture the legal issues. ROSS then positioned itself as a direct, fee-charging substitute for Westlaw.
The Originality Threshold: Headnotes Carry the “Creative Spark”
Before reaching fair use, the court rejected ROSS’s argument that Westlaw headnotes lack copyright protection. Under Feist Publications, Inc. v. Rural Telephone Service Co. (1991), originality requires only a minimal creative spark. The Third Circuit found that Westlaw’s editors met that threshold by deciding which legal holdings mattered, synthesizing the relevant facts, and drafting headnotes that could stand on their own without reference to the underlying judicial text.
The court also dismissed ROSS’s invocation of the “merger doctrine” (which is the idea that the expression and the facts are inseparable). Because there are myriad ways to summarize and express the legal rules articulated in a court opinion, the underlying idea did not merge with Westlaw’s specific editorial expression.
The Four-Factor Fair Use Breakdown
The core of the Third Circuit’s ruling lies in its analysis of 17 U.S.C. § 107. While the second statutory factor tipped marginally toward ROSS because headnotes are predominantly factual, factors one, three, and four decisively favored Thomson Reuters.
Factor 1: Purpose and Character of the Use
Under the Supreme Court’s Andy Warhol Foundation v. Goldsmith (2023) precedent, courts must weigh the degree of transformativeness against the commercial character of the secondary use.
The Same Ultimate Purpose: ROSS argued that using headnotes as computational inputs to train a statistical model was an intermediate, transformative process. The Third Circuit flatly disagreed because Westlaw and ROSS ultimately used the material for the same commercial purpose: helping attorneys locate relevant case law. The presence of an intermediate machine-learning step did not make the use transformative when the resulting product performed substantially the same function as the original.
Distinguishing Google Books: Unlike Authors Guild v. Google, Inc. (2d Cir. 2015)—where books written for reading were digitized to enable full-text indexing and snippet displays that pointed readers toward purchasing the originals—ROSS used the headnotes to replace Westlaw.
Distinguishing Reverse-Engineering Precedents: In landmark software cases like Google v. Oracle (2021), Sega v. Accolade (9th Cir. 1992), and Sony v. Connectix (9th Cir. 2000), intermediate copying was deemed fair because copying declaring code or BIOS was “strictly necessary to access unprotected functional interfaces.” Here, copying Westlaw headnotes was not necessary; the underlying court opinions were freely accessible. ROSS copied headnotes simply because it was cheap and efficient. As the court memorably noted: “Unlike necessity, ease is not a justification for copying.”
The Generative AI Distinction (Footnote 7): The court addressed the DOJ’s amicus position in In re: OpenAI but did not decide whether training a generative AI model is fair use. Instead, it noted that large language models that generate new expressive outputs may raise different transformativeness questions from a non-generative retrieval tool designed to compete in the same market as its training inputs.
Factor 2: Nature of the Copyrighted Work
Because headnotes are factual compilations derived from judicial opinions, they reside closer to the boundary of copyrightable expression than purely imaginative works like novels or songs. While this favored ROSS, the court reiterated that factor two rarely carries dispositive weight.
Factor 3: Amount and Substantiality
Quantitative proportions do not tell the whole story. ROSS protested that copying 25,000 headnotes represented a mere 0.08% of Westlaw’s 28-million headnote database. The Third Circuit held that each individual headnote is a separate copyrightable work, meaning ROSS copied the entirety (100%) of each challenged headnote. Without a valid transformative purpose to justify taking the “heart” of these discrete works, the scale of copying weighed heavily against fair use.
Factor 4: Market Effect and Derivative Licensing
Like I’ve been saying for months, cases live and die by Factor 4. Here, the market-effect analysis favored Thomson Reuters because ROSS threatened both the value of Westlaw’s headnotes and the emerging market for licensing editorial content as AI training data.
- Harm to the Original Work’s Value: Even though Thomson Reuters did not sell headnotes as an isolated, standalone product, the headnotes served as an essential customer draw for Westlaw subscriptions. Relying on Video Pipeline v. Buena Vista (3d Cir. 2003), the court held that copying the headnotes diminished their commercial value as an anchor feature of Westlaw.
- Harm to Derivative AI Licensing Markets: Under Campbell v. Acuff-Rose (1994), courts must analyze markets that creators are likely to develop or license. The Third Circuit recognized that licensing editorial text as AI training data is a real, rapidly developing derivative market. By scraping Westlaw’s headnotes without a license, ROSS usurped Thomson Reuters’s legitimate opportunity to monetize its data for AI development.
- Rejection of Vague Policy Pleas: The court dismissed ROSS’s broad assertions that enforcing copyright would “halt AI development” or threaten “national security” as unsupported by evidence, emphasizing that utilizing machine learning does not grant an enterprise “carte blanche” to bypass copyright law.
Practical Guidance for Industry Stakeholders
Nothing is black & white in the law, so here is a rough risk spectrum for AI developers, data scrapers, content creators, and publishers assessing fair use after Reuters v. ROSS:
Fair Use Exposure Spectrum After Reuters v. ROSS: AI Training Data, Scraping, and Market Substitution
| Higher Exposure of Infringement | Lower Exposure of Infringement |
| Scraping curated/editorial data | Scraping raw facts & public domain |
| Shortcut copying to save costs/time | Copying strictly necessary to interoperate |
| Outputs compete directly with source | Highly transformative, generative outputs |
| Bypassing existing licensing markets | Intermediate analysis with robust guardrails |
Guidance for AI Developers & Data Scrapers (Potential Defendants)
Distinguish “Raw Public Data” from “Curated Value-Add”: You can freely scrape and ingest raw, uncopyrightable facts (e.g., judicial opinions, government reports, raw census data). However, as soon as your scraper targets human-curated taxonomy, tags, annotations, or proprietary summaries, you cross into copyrightable expression.
Avoid Convenience Shortcuts: If the raw, underlying facts are available elsewhere, scraping an intermediary’s editorialized work because it is cleanly formatted or easy to parse destroys any defense of “technological necessity”. Document why specific inputs were chosen.
Target Non-Substitutive Market Objectives: If your model’s commercial purpose is to compete directly with the product from which the training data was extracted, you will fail Warhol’s transformative-use test. Ensure your downstream deployment provides distinct functionality that does not substitute for the source market.
Monitor Clean Hands and Terms of Service: While the court focused its core analysis on copyrightability, footnote 9 explicitly cited ROSS’s bad-faith actions—such as bypassing Westlaw paywalls, misrepresenting credentials, and violating terms of service—as factors supporting liability. Ethical data provenance matters. Every Gen AI case puts the source and acquisition of training data under scrutiny. Paying for curated data and securing the rights required for the intended use can materially reduce risk, but the license terms still matter.
Guidance for Content Creators & Publishers (Potential Plaintiffs)
Separate Unprotectable Data from Proprietary Expression: To prevail in court, articulate the specific human selection, coordination, and synthesis that gives your content its creative spark. Maintain clear guidelines and evidentiary records demonstrating how your editors or creators make independent creative choices.
Establish and Document Active AI Licensing Channels: The Third Circuit recognized data licensing for AI training as an actionable derivative market. Rightsholders should actively establish licensing frameworks, partner with collective management organizations (CMOs), or use data-licensing brokers. Proving that a commercial market for your training data exists—or is likely to be developed—can win Factor Four.
Tie In-House AI Features to Your Core Product Value: Thomson Reuters successfully showed that it used its own headnotes to train internal AI features (like WestSearch Plus). Demonstrating that unauthorized scrapers undermine your own AI product development directly validates your market harm claims.
Conclusion
The Third Circuit has made one principle unmistakably clear: innovation in algorithmic design cannot be subsidized by the unlicensed expropriation of another company’s editorial labor. If an AI developer uses copyrighted editorial content to build a commercial substitute because copying is easier than gathering the underlying public-domain facts, fair use may offer no shield. The practical takeaway is equally clear: document where training data came from, distinguish raw facts from curated expression, and license proprietary content before using it to build a competing product.
About Tim Billick: Tim is a patent and intellectual property attorney at Practus who focuses on IP litigation, trademark and patent prosecution, and technology licensing. His experience spans technologies including software, medical devices, mechanical engineering, semiconductors, and consumer products, and he brings an entrepreneur’s perspective to helping companies protect and commercialize innovation.



