The most important sentence in the first U.S. appeals ruling on AI training is not in the body of the opinion. It sits at the bottom of page 17, in footnote 7 of the Third Circuit's decision in Thomson Reuters v. ROSS Intelligence, filed September 29, 2026:
"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."
Both are chatbot cases. Bartz v. Anthropic is the authors' case against Anthropic in California; In re OpenAI gathers before one judge in New York the cases news organizations and authors brought against OpenAI and Microsoft. ROSS built something else — a legal search engine that pointed you to passages in real court opinions. It wrote nothing of its own, and it went dark in 2021.
Two earlier pieces here named this appeal as pending, and the footnote grades both. In June, I wrote that "ROSS was a non-generative legal-search tool, so its eventual outcome won't cleanly settle the chatbot question." That holds. In September, I called the appeal "the first federal appellate test of the whole question" — and footnote 7 says it was not the whole question. I would rather correct that myself than leave it to a headline.
One disclosure, because the footnote names Anthropic's case: this newsletter is written with AI assistance, and the assistant is made by Anthropic. Both cases get the same treatment below. Hold me to that.
So what did the court decide? Less than the headlines say, and, in one respect, more than ROSS's allies would like.
A search engine that wrote nothing
Thomson Reuters owns Westlaw, whose editors write headnotes (short notes that state a point of law from a court opinion and sit above it on screen). ROSS Intelligence wanted Westlaw's customers and made no secret of it.
To train its system, ROSS's contractor wrote approximately 25,000 memos, each a legal question paired with passages from court opinions. The drafters built the questions from thousands of Westlaw headnotes, because the headnotes provided "an easy way [to] fram[e] questions." The court was explicit: "ROSS's AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions."
The panel upheld the trial court on both questions: "We therefore hold that Thomson Reuters's materials are copyrightable and that ROSS's use was not fair." The ruling covers 2,243 headnotes the trial court found copied. Of fair use's four factors (roughly: why you used the work, what kind of work it is, how much you took, and what your use does to its market), the court found that "the second factor weighs slightly in favor of fair use, but the first, third, and fourth factors weigh against it."
Why wasn't using the headnotes as training material a different enough purpose? ROSS pointed to three software cases that allowed copying computer code. The court read them as cases about need: there, "copying was necessary to access the unprotected functional aspects of computer code." ROSS, by contrast, "had access to the underlying judicial opinions and could freely copy them to make the memos needed to train its AI." It used the headnotes because they offered an "easy" way to build the memos. Hence the line the headline borrows, in full: "Unlike necessity, ease is not a justification for copying."
Judge Tamika R. Montgomery-Reeves, a Biden appointee, wrote for a unanimous panel that included an Obama appointee and a Trump appointee. The court first filed the opinion under seal, then made it public once both sides agreed nothing needed redacting.
To be clear about the limits: the holding concerns a tool that generated nothing and was built to replace the service it copied from, and it binds only the federal courts in Pennsylvania, New Jersey, Delaware and the Virgin Islands, as Authors Alliance points out. The group, which backed ROSS on appeal, calls this "the first federal appellate ruling on whether copying works to train an AI system is fair use." (The Ninth Circuit's Doe v. GitHub ruling, on September 16, turned on claims about AI outputs, not on fair use in training.)
Unless further review intervenes, the case goes back to the trial court in Delaware. ROSS, out of business but still a litigant, said on October 2 that it will ask the Supreme Court to take the case, arguing that the decision "creates continued uncertainty around the application of copyright law to AI model training." A Thomson Reuters spokesperson told Reuters (which Thomson Reuters owns) that the company was pleased with the ruling.
Why would a court answer a case it wasn't hearing?
Because the government had argued the question in another courtroom. On September 1, the Justice Department filed a statement of interest in the New York case, arguing that "the use of copies to train LLMs is extraordinarily transformative." Footnote 7 answers that filing up front — "The concerns raised in that separate case do not apply here" — and only then notes that the department "relied on Bartz" and argued that the OpenAI training "did not result in 'substitutive competition.'"
Why don't those concerns apply? Two reasons, in one breath. ROSS's platform "cannot generate original expression" — and "ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw." Which matters more? "The panel never says which ground carries the weight," writes Josh Geller of the entertainment-law firm Mitchell Silberberg & Knupp, who says footnote 7 "may be the most important passage in the opinion."
Authors Alliance calls the footnote "a welcome caveat" but warns that it "may also be weaker than the opinion appears to intend." The reason, in its words: "Much of the work in footnote 7 is done not by the generative/non-generative distinction but by substitution." And: "Plaintiffs in generative AI cases will surely use this case to argue that any model competing in their market belongs on ROSS's side of that substitution line."
The footnote decides neither case. Bartz produced a June 2025 trial-court ruling on fair use and later settled. In re OpenAI is in summary-judgment briefing, and as of October 3, 2026, replies are due November 20 (my September piece said November 6, before a September 24 extension).
Then why are publishers calling it a powerful precedent?
Because of the fourth factor — the part of this opinion most likely to travel. The statute asks about "the effect of the use upon the potential market for or value of the copyrighted work," and this case leans hard on "potential."
Thomson Reuters argued that ROSS harmed its position in "the potential derivative market for licensing headnotes as AI training data," though it had not licensed its headnotes to others. The court counted the harm anyway: "Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing." Thomson Reuters, it noted, "is using its headnotes as training data for its own AI search products." And: "That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so." In short, a licensing market the owner never entered still counts as harmed.
Why should anyone outside legal research care? Because the same fight runs through the generative cases, and as the law firm Ballard Spahr notes, "district courts have already divided over the market-harm question," and "No federal appellate court has resolved that disagreement."
Authors Alliance makes the strongest case against the court's reasoning, and it deserves its own words:
"The opinion cites nothing in the record for the first proposition. It never addresses the circularity problem the Supreme Court flagged in Oracle and that other decisions guard against: every rightsholder loses a licensing fee if the relevant market is defined as the market for licensing the very use at issue."
(The "first proposition" is the court's line that the market is "rapidly developing.") Put plainly: define the market as the very use being challenged, and every unlicensed use is a lost sale by definition, so the fourth factor favors the owner every time. The group predicts that "Rightsholders will cite its acceptance of an unlicensed, thinly evidenced AI training market."
The rightsholders' reading is the mirror image. The Copyright Alliance, which calls itself "the unified voice of the copyright community," says the decision makes "unequivocally clear" that harm to a potential AI-training market counts, and that an owner who "has not yet entered such a market" is harmed all the same. Geller grants that the panel "did not address the circularity argument," but notes that "licensing of training data is a commercial reality" either way.
I think both camps are describing something real, and what separates them is evidence. I expect that is where the chatbot cases will be fought.
What if a legislature had answered first?
Japan's did — years before any American appeals court. Article 30-4 of Japan's Copyright Act permits exploiting a work "in any way and to the extent considered necessary" when no one's purpose is to enjoy the thoughts or sentiments it expresses, naming data analysis as one such case. Then comes the exception to the exception: "provided, however, that this does not apply if the action would unreasonably prejudice the interests of the copyright owner…" (That is the Justice Ministry's English translation, a reference text; only the Japanese has legal effect.) On September 13, I called that proviso "a further brake"; this ruling makes its design worth a closer look.
Japan chose that shape on purpose. Drafting the 2018 amendment, whose Article 30-4 took effect January 1, 2019, the government weighed American-style fair use and chose specific provisions instead, citing "how roles are shared between the legislative and judicial branches" and Japan's litigation system, among other reasons.
In 2024, an expert subcommittee advising Japan's Agency for Cultural Affairs set out how the rule applies to AI: using works for "AI development or other forms of data analysis" may, "in principle, be allowed without the permission of the copyright holder." The guidance "is not legally binding," the agency notes.
So when does the brake engage? The guidance asks "whether it will compete in the market with the copyrighted work" and "whether it will impede the potential sales channels of the copyrighted work in the future." Then it ties that future market to signals an owner controls: a copyrighted database "available on the internet for a fee," technical measures such as a robots.txt file (the instruction a website uses to turn crawlers away), and "the past sales record" of such databases.
As I read the two, both look forward, at markets that may not exist yet, but Japan's guidance anchors the potential market in what the owner has done, while the Third Circuit accepted a market Thomson Reuters had not entered. Japan's signals look like an attempt to head off exactly the circularity Authors Alliance describes. That compares designs, not outcomes: nothing in the guidance says how a Japanese court would treat Westlaw's headnotes, and in 2024 the agency itself said there were "very few court precedents" on AI and copyright.
As of October 3, 2026, Article 30-4 remains in force, unchanged, and it has critics at home: the Japan Newspaper Publishers & Editors Association complains that the rule "contains no explicit opt-out provision" and wants consent before news is used to train generative AI.
When "necessary" becomes a question for engineers
Here is where things get interesting, and where I start extrapolating; treat this as my guesswork.
Suppose you wrote a well-loved field guide, and a model trained on it ends up in a court that follows this opinion. The fight may turn on whether its builder could have learned the same things from material it was free to copy, the way ROSS could have used the opinions. Geller expects as much: what training data "is actually necessary to make a functional model remains an open technical question that will likely continue to be litigated." Imagine two expert witnesses arguing over whether your book was necessary, or merely easy.
Or suppose you hold an archive (a photo library, say). The court credited Thomson Reuters's evidence that it trains its own AI on its headnotes, and the company now calls itself "an AI and technology company" on its About page. Follow that logic and every archive owner has a reason to train something on its own archive — partly to show some future court that a training market is "rapidly developing." The circularity Authors Alliance worries about would arrive as a business plan.
And the Supreme Court's first AI-training case could be about a search engine that went dark in January 2021. On ROSS's plan, the litigation tracker Chat GPT Is Eating the World wrote: "All eyes will be on the position of the U.S. Solicitor General" (the government's lawyer at the Court). Authors Alliance, in an October 2 post that does not mention ROSS's announcement, said the odds of rehearing or review "seem low to us," and it sees the likelier route to the Court running through the generative AI cases.
Who's cheering, who's worried, and why the right is on both sides
Publishers' and authors' groups are cheering. "We applaud the court for this powerful precedent," said Maria Pallante, who heads the Association of American Publishers. Mary Rasenberger of the Authors Guild said the ruling "vindicates copyright law."
Before the ruling, the Electronic Frontier Foundation, whose brief for ROSS was joined by library and public-interest groups, argued the other way: "The law should encourage the creation of AI tools to digest and identify facts for use by researchers, including facts about the law."
On the right, the split is real. It was the Trump Justice Department that called training "extraordinarily transformative." Mike Davis of the conservative Article III Project urged the department to withdraw that filing, writing, "In a free market, businesses pay for the inputs they need." And the Foundation for American Innovation (whose programs include a fellowship for conservative policy professionals) backed ROSS, arguing that "a bare nonpublic 'license to be trained on' is not a market for expression; no cognizable harm appears on this record."
Ballard Spahr's alert frames the question this ruling leaves for a future appeals court: "should copyright law treat generative AI training more like human learning, or does the model's capacity to generate competing expression at unprecedented scale require different analyses for purpose of use and market harm?"
What does this mean for you?
Don't read it as a verdict on chatbots. It covers 2,243 headnotes copied for a tool that generated nothing, in one circuit. If someone says a court ruled AI training isn't fair use, ask which court, and which AI.
Read footnote 7 yourself. The opinion is free and linked at the top; the footnote is on page 17.
If you make things, find out who holds the AI rights in your contracts. The court counted a market for licensing headnotes as AI training data as Thomson Reuters's to enter. If courts apply that logic to your work, that market belongs to whoever holds the AI rights. The Authors Guild's free model contract clauses say "any and all grants of AI rights to publishers should be expressly negotiated" (an advocacy group's advice; weigh it with your agent).
Register what you would want to defend. "You will have to register, however, if you wish to bring a lawsuit for infringement of a U.S. work," and registered works "may be eligible for statutory damages and attorney's fees in successful litigation."
If you build datasets, use sources you are entitled to copy. In the Third Circuit, at least, convenience will not justify lifting someone else's editorial layer. Authors Alliance's text-and-data-mining resources are written for researchers.
Put the dates down (as of October 3, 2026). The deadline for a rehearing petition is 14 days after entry of judgment, by my count October 13. A Supreme Court petition is due "within 90 days after entry of the judgment" (by my count December 28, 2026, unless a rehearing petition or an extension moves it). In New York: oppositions October 23, amicus briefs October 30, replies November 20, with no hearing or decision date set.
Follow the New York case directly. CourtListener's free docket alerts send you updates when new filings appear (its copies can lag the court's official system).
The lesson, as I see it
The footnote is the most honest sentence in the opinion: the court declined to decide the chatbot question on a record that never presented it, and said so where the government would look. The necessity line will last; I expect "ease is not a justification for copying" to be quoted for years (along with Authors Alliance's reply: "transformative use has never required that copying be the only route to a purpose"). The fourth factor is where the real argument starts.
A market for training data is real, and an owner should not lose it just because it moved more slowly than the copier. But if a "potential market" exists the moment anyone copies without paying, the fourth factor stops asking a question. The answer is evidence a stranger could check: what was offered, to whom, at what price. Japan's guidance spells out its version of that list. American courts will build theirs one case at a time; as of October 3, 2026, the last scheduled New York summary-judgment briefs are due November 20.
My vote? Let the market count, and make the owner show it.
Convenience is a fine reason to take the elevator. In the Third Circuit, it is not a reason to copy.