12:01AM, October 10, 2026

In Thomson Reuters v. ROSS Intelligence, decided by the Third Circuit on September 29, 2026, the court became the first federal appellate court to rule on whether using copyrighted material to train an artificial intelligence program constitutes fair use. For ROSS’s use, the court said no. No sooner had the ruling been issued than media headlines described a first appellate loss for AI training. A key limit appears in a footnote of the decision. 

By technology standards, these facts feel ancient. ROSS aimed at developing a legal search engine that would respond to a question posed in natural language by quoting passages from judicial opinions. To train this engine, ROSS needed some examples of legal questions together with appropriate answers. After Thomson Reuters refused to license its content to ROSS because it was a competitor, ROSS hired a contractor to prepare approximately 25,000 “bulk memos,” question-and-answer sets created using Westlaw headnotes, that is, summaries of legal points of a judicial opinion written by the editors. Thomson Reuters sued. In February 2025, Judge Stephanos Bibas (sitting by designation in the District of Delaware) granted partial summary judgment concerning 2,243 headnotes and rejected ROSS’s fair-use defense. 

Judge Montgomery-Reeves, writing for a unanimous panel, affirmed. Headnotes have “some creative spark,” because choosing which points of law to extract and how to phrase them involves creative judgment. The merger doctrine cannot help ROSS since “just as a banana costume can take myriad designs, so too can headnotes have differing expressions.” 

In terms of the first factor, both companies used the headnotes in creating their tools for legal research, making ROSS’s use of headnotes “minimally transformative, at best.” In the intermediate-copying cases dealing with software (Sega, Connectix, Google v. Oracle), ROSS’s case was not strengthened since the copying in those instances was needed to get to the unprotected functional components of the computer code. Moreover, ROSS could have used the opinions. However, according to the court, it chose to copy the headnotes because that offered an easy way to create training memos. In the fourth factor, ROSS competed with the current market for legal research and usurped Thomson Reuters’s opportunity to enter a developing market for licensing headnotes as AI training data. The judgment was affirmed, although not all of the reasoning used: Judge Bibas had ruled in ROSS’s favor on the third factor, and the panel disagreed on that factor. 

The first appellate word on AI training turned out to be about a search engine. The question the headlines were asking, what happens when the model writes, is still waiting for its appellate court.

Now the footnote. On September 1, 2026, the Department of Justice filed a statement of interest in the OpenAI litigation in New York, telling that court that training a large language model is transformative because the model goes on to produce new responses of its own. The Third Circuit did not say the government was wrong. It said the argument does not fit ROSS: unlike the models in Bartz and In re: OpenAI, ROSS’s “AI platform cannot generate original expression,” and its tool did the same job as Westlaw. With that, the court set aside the questions that make the generative cases hard. Is a model that learns from millions of books in order to write new text doing something transformative? Is a flood of machine-written output the kind of market harm copyright recognizes? Both remain open. 

This is significant because the trial courts have taken different approaches to generative AI training. In June 2025, Judge William Alsup in Bartz v. Anthropic, sitting in the Northern District of California, concluded that training Claude on books was “exceedingly transformative,” like someone who reads with the intent of becoming an author, while also concluding that downloading pirated books to build a permanent library was not fair use. It is the latter conclusion that led to the $1.5 billion settlement approved in July 2026. Two days after Alsup’s decision, Judge Vince Chhabria issued summary judgment for Meta in Kadrey v. Meta, but only “on this record,” noting that “in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission” due to their ability to cause market dilution. All three cases illustrate different views of transformation and market effects, with both generative AI cases finding training fair use, while the appellate ruling concerned a search engine. 

Still, parts of the reasoning will remain relevant. Firstly, “AI training” is not a magic phrase. The court considered the function of the tool rather than the means used to create it, and a defendant who created a similar end product will struggle on factor one regardless of the technology. Secondly, necessity has significance. Where the unprotected element (the opinions) is accessible through a means other than reproduction of the protected element (the headnotes), the intermediate-copying doctrine will be of no help to the defendant. Thirdly, and most importantly for the plaintiffs, the court recognized a potential market for licensing Westlaw headnotes for AI training and treated ROSS’s copying as harming it. Such reasoning has been criticized as circular: copyright owners can always claim they would have charged for the disputed use. Whether circular or not, it remains the law in the Third Circuit, and a valuable argument for the plaintiffs in the generative cases, even if its reach beyond headnotes is contested. 

The first appellate word on AI training turned out to be about a search engine. The question the headlines were asking, what happens when the model writes, is still waiting for its appellate court. 

Esther Dong

Esther Dong is a 2L at the University of North Carolina School of Law and a staff member of the North Carolina Journal of Law & Technology. Esther graduated with highest distinction from UNC Chapel Hill with a B.A. in Communication Studies and is interested in technology transactions, intellectual property, and artificial intelligence law. When not in class, you can find her playing tennis, hiking, or being bossed around by her cat, Onyx.