Fair Use on Trial: What the Third Circuit's ROSS Ruling Means for AI Training Data
On September 29, 2026, the Third Circuit issued a decision in Thomson Reuters v. ROSS Intelligence, affirming a lower court's partial summary judgment and rejecting the fair use defense for the copying of Westlaw headnotes. The decision, rendered by Judge Montgomery-Reeves, clarifies the application of fair use in the context of training datasets used by legal research tools.
The Threshold of Copyrightable Expression in Legal Summaries
A central component of the litigation involved whether Westlaw headnotes, which summarize legal points within cases, qualify for copyright protection. Under the standard established in previous copyright jurisprudence, facts themselves are not copyrightable. However, the "modicum of creativity" required to meet the Feist bar can be found in the selection, coordination, and arrangement of those facts. The court found that the Westlaw headnotes met this threshold, providing a level of creative expression that distinguishes them from the raw, unprotectable facts of the underlying judicial opinions.
This finding has significant implications for the AI industry. If headnotes are considered sufficiently creative, they move from the realm of "data" into the realm of "protected expression." This distinction is vital. For developers of specialized AI models, the distinction between training on a public statute (a fact) and training on a proprietary summary of a statute (a creative work) determines the legal viability of their entire data acquisition strategy.
The recognition of headnotes as copyrightable works effectively shifts the legal risk for AI developers from the acquisition of raw data to the acquisition of structured, human-curated summaries.
Why It Matters
Deconstructing the Four-Factor Fair Use Test
The Third Circuit's analysis focused on the four factors of the fair use doctrine. The court's application of these factors provides a roadmap for how future courts might treat specialized datasets.
| Fair Use Factor | Application in ROSS Intelligence Case | Court's Determination |
|---|---|---|
| 1. Purpose and Character of Use | Commercial use; minimally transformative | Against ROSS Intelligence |
| 2. Nature of the Copyrighted Work | Published, highly structured, and creative headnotes | Slightly favors fair use |
| 3. Amount and Substantiality of the Portion Used | Substantial copying of the core headnotes | Against ROSS Intelligence |
| 4. Effect on the Market for the Work | Potential harm to the licensing market for headnotes | Against ROSS Intelligence |
The first factor was particularly decisive. While ROSS argued that their use of the data was transformative because it facilitated a new way of searching legal information, the court found the use to be "minimally transformative." The court noted that the primary purpose of the copying remained the same: to provide a tool for legal research. When a use is purely commercial and adds very little new expressive meaning to the original, the first factor weighs heavily against the user.
The Distinction Between Generative and Extractive Models
A critical aspect of the decision was the court's refusal to conflate the ROSS case with the broader debates surrounding generative AI. Many proponents of fair use argue that Large Language Models (LLMs) are transformative because they do not merely "reproduce" data, but "learn" from it to create entirely new content. This is the essence of the generative AI defense.
However, the Third Circuit distinguished the ROSS matter from these generative cases. Reports describe the dispute as an ordinary copyright case rather than a generative AI case. One reading is that when an AI product serves the same function as the copyrighted source, the "generative" label may not shield it if the input was a direct, non-transformative copy of a structured work. That is an interpretation of the ruling, not a holding the court stated in those terms.
Reports indicate that certain portions of the court's reasoning remain sealed pending redactions. This secrecy may hide crucial nuances regarding how the court views the "learning" process itself versus the "copying" process. If the sealed portions address whether a model's "weights" constitute a derivative work, the implications for the wider AI industry could be significant.
The Economic Reality of Market Displacement
The fourth factor, which examines the effect of the use upon the potential market for or value of the copyrighted work, was a cornerstone of the ruling. The court counted harm to the market for licensing Westlaw headnotes as AI training data. In other words, the unlicensed use competed with a licensing business that Thomson Reuters could pursue.
This finding points toward a licensing expectation for specialized AI. If a data provider has a viable business model based on selling access to its curated summaries, an AI company cannot simply argue that "scraping" is a fair use. The harm is not just to the sale of books or software, but to the emerging economy of data-as-a-service (DaaS). As AI training shifts from general web crawling to specialized, high-quality vertical datasets, the "market harm" argument will become the primary battlefield for copyright litigation. Courthouse News reports that the court held AI training on the copyrighted headnotes was not fair use in this case.
Strategic Outlook for AI Stakeholders
The ROSS decision provides several indicators that investors, managers, and policymakers should monitor as the legal terrain settles.
- For Investors: The ruling strengthens the case for treating curated datasets as protected assets. Companies that own highly structured, human-curated datasets possess a significant defensive advantage. The risk of unlicensed training is now a concrete litigation risk for vertical AI startups.
- For Managers: Data provenance is paramount. Organizations must audit their training pipelines to distinguish between public domain facts and protected, structured summaries. Reliance on "transformative" arguments for commercial, extractive models is a high-risk strategy.
- For Policymakers: The distinction between "learning" and "copying" remains the central tension. Regulators will likely face pressure to define whether the "weights" of a model constitute a reproduction of the training data, a question that the ROSS decision leaves partially open.
The primary indicator to watch is the next round of appellate or Supreme Court challenges. The tension between the "minimally transformative" standard and the need for AI innovation is a conflict that may require a legislative remedy rather than a purely judicial one. Haynes Boone's alert summarizes the holding: the headnotes were copyrightable and the fair use defense failed.
Sources: haynesboone.com, courthousenews.com.