An Appeals Court Upheld a Copyright Win Over AI Training. The AI Company Lost.

On September 29, 2026, the US Court of Appeals for the Third Circuit upheld a lower court's ruling in favor of Thomson Reuters against Ross Intelligence, a legal-research startup. Reuters, whose parent company won the case, reported that this is the first US appeals court decision on a copyright dispute over AI training. Judge Montgomery-Reeves wrote the opinion. Judges Restrepo and Bove sat on the panel with her, and the ruling now binds lower courts in Pennsylvania, New Jersey, Delaware and the Virgin Islands.

The dispute goes back to 2020. Thomson Reuters accused Ross of copying thousands of Westlaw "headnotes," the short editorial summaries that pull the key legal point out of a court opinion, and using them to train a competing AI legal-search tool. In February 2025, Judge Stephanos Bibas ruled against Ross and rejected its fair-use defense.

Here is the shape of what just happened, because it is narrower than the headline sounds. There has been no trial. Judge Bibas ruled early, on one slice of the case: 2,243 particular headnotes he found Ross had copied. Both sides then asked him to send two questions about that slice up to the appeals court before the rest of the case moved on: whether those headnotes can be copyrighted at all, and whether Ross's use of them was fair. The appeals court answered both, and sided with Thomson Reuters on both.

The rest of the case stays in the lower court, including the claims about the other headnotes Ross is accused of copying and any money Ross would owe. Ross lost the argument that mattered most. Nobody has set a price on it yet.

On fair use, the court worked through the four factors the statute lists. On the first, it called Ross's use commercial and "minimally transformative, at best." Ross took an extra step that Thomson Reuters did not, feeding the headnotes to a machine-learning program, and the court allowed that this "arguably presents a slight degree of difference in use." But both companies used the headnotes toward the same end, helping lawyers find the right case, so the first factor went against Ross.

The second factor went the other way. Because headnotes have to describe the law accurately, the court found them "more factual than fictional," and in its conclusion said that factor "weighs slightly in favor of fair use." It was the only one that did. On the third factor, the court found that "for each headnote taken, ROSS copied an entire work," and that Ross did not need to copy them, since the underlying court opinions are free to anyone. On the fourth, it found "the market for licensing headnotes as text to train AI is rapidly developing," and that Thomson Reuters not having licensed them yet "does not disprove that a market exists to do so."

The court also decided something the coverage mostly skipped: that the headnotes can be copyrighted at all. Ross had argued that a headnote is just the legal point restated, so protecting it would hand Thomson Reuters a monopoly on the law. The court disagreed. Its comparison was a banana costume, from one of its own earlier cases. A banana can be made into many different costumes, so no one design locks up the idea of a banana. In the same way, the panel endorsed a line from the District Court, that "there are many ways to express points of law from judicial opinions," so protecting one phrasing does not lock up the opinion.

Ross had also been expected to challenge Westlaw's Key Number System, the index Thomson Reuters uses to classify opinions. The question was inside what the lower court sent up for review, but in footnote 4 the court noted that Ross "never challenged the originality of the Key Number System in any of its briefing, despite its inclusion in the certified questions," treated the issue as forfeited, and declined to address it. So it was never argued, and this ruling leaves no appeals-court answer on it.

One correction to the record, because it has traveled: the company said to be gone argued this appeal. The opinion gives the argument date as June 11, 2026, and the counsel listing at the back of the opinion names White & Case as the firm that argued for Ross. Treat any claim that the losing company no longer exists as a reason to check the court record rather than the coverage.

It is worth being careful about what this ruling does and does not settle for everyone else. Ross's tool retrieved passages out of court opinions and pointed lawyers at them. It never wrote new text of its own, and nothing it produced read like Thomson Reuters's writing. The suits still pending against other AI developers mostly involve systems that do write new text, which is a different fact pattern.

The court addressed that difference in footnote 7, and the reason it did is the most interesting thing in the opinion. On September 1, 2026, the US Department of Justice filed a statement of interest in the OpenAI copyright litigation in New York, arguing that training a large language model, which can "generate original responses," is a transformative use. It leaned on Bartz v. Anthropic to say so.

Twenty-eight days later, the Third Circuit answered in a footnote: "The concerns raised in that separate case do not apply here." It gave two answers, taking the DOJ's arguments in the DOJ's own order. To the argument built on Bartz, about models that can generate original responses: "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness." The DOJ also argued that the training in the OpenAI case did not produce "substitutive competition," and to that the court said: "Here, ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw." It then added a pointed line: the DOJ "is tracking these issues and knows how to assert its interests, but the DOJ notably did not do so here."

Two things about that footnote are worth holding on to. It does not rank the two answers, and it never says which one controls. And it is expressly about this record: the concerns raised in the other case, the court wrote, "do not apply here." So it decides nothing about a model that does write new text, and reading it either as a safe harbor for such a model or as a rule against one puts more in it than the court wrote. The court also announced no new rule about AI. It said the opposite about itself: "[t]his is no more than an ordinary copyright case."

One date worth getting right, since coverage mixes it up. What footnote 7 cites in Bartz v. Anthropic is a 2025 district-court fair-use ruling, reported at 787 F. Supp. 3d 1007. A decided 2025 ruling is what the DOJ leaned on and what the Third Circuit was answering, which is not the same thing as a case still waiting for an answer.

Thomson Reuters said, per the reporting, that it "firmly believes that respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions."

For anyone building a product on top of an AI model, the practical lesson is about where the legal risk sits. The subject matter is almost incidental here, and reading this as a case about legal summaries misses the point. The risk lives in training on content you do not have clear rights to, for a product that competes with what that content was for. That risk does not go away because a case settles or a company stops selling. It sits in the design until someone builds around it.

So two questions are worth answering about your own product before a court asks them for you. What did the model train on, and can you show it. And does the product you sell compete with the market that material already served. If the answer to the second is yes and the answer to the first is unclear, the fix is a license or a different source, and it is cheaper now than in a brief.

That is the design choice behind ForIntel, Foragentis's intelligence-products line. Every claim in a ForIntel report traces back to a named, checkable source, rather than to a model trained on material nobody licensed, with no path back to where the information came from. The question this ruling raises for the rest of the industry, whether an AI product's answers can be traced to sources it had the right to use, is the one ForIntel was built to answer from the start: https://foragentis.com