It depends, and a new federal appeals court ruling does not settle the question for chatbots or image generators. In September 2026 the Third Circuit held that a company violated copyright when it used 2,243 Westlaw summaries to train a legal-search tool built to compete with Westlaw. The court called it "no more than an ordinary copyright case" and said the tool could not create new text. Whether training a generative model on your work is legal remains open in other courts.
What did the appeals court actually decide?
The U.S. Court of Appeals for the Third Circuit decided Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153, in an opinion filed September 29, 2026 (read the court's opinion). Judge Montgomery-Reeves wrote it, and Judges Restrepo and Bove were on the panel. The text of the opinion shows no separate concurring or dissenting opinions.
The court answered two questions. First, are the editorial summaries Thomson Reuters writes for Westlaw, called headnotes, original enough to be protected by copyright? Yes, it said. Second, was it a fair use for a competitor named ROSS Intelligence to copy those headnotes to train its AI? No, it said. The court affirmed a lower-court order that granted Thomson Reuters partial summary judgment, which is a ruling for one side on specific issues without a trial.
This was an interlocutory appeal, meaning an appeal taken before the whole case is finished. The opinion does not describe what happens next in the lower court, so this article does not either.
Who were the companies, and what did ROSS copy?
Thomson Reuters owns Westlaw, an online legal-research service. Westlaw editors write headnotes, which are short annotations explaining a point of law made in a judicial opinion. Editors are told to limit each headnote to 800 characters where possible, and each one has to make sense to a reader who has not seen the opinion.
ROSS was founded by three University of Toronto computer science students. It built a search engine that answers plain-language legal questions by returning passages from roughly ten million judicial opinions. Those opinions are not copyrighted, and Thomson Reuters does not claim copyright in them. The headnotes are a different matter.
To teach its software which passages answer which questions, ROSS hired a company called LegalEase Solutions, which wrote about 25,000 training memos. Each memo posed a legal question and listed four to six opinion passages as answers. The memo writers used thousands of Westlaw headnotes to come up with the questions, because the headnotes gave them "an easy way" to frame them. The district court found that for 2,243 headnotes, the memo questions were so close to the headnote wording, and so unlike the opinion wording, that no reasonable juror could find they were not copied.
ROSS did not hide its goal. It ran ads comparing itself to Westlaw, priced its service "in line with" Westlaw, and some law firms switched.
Are short summaries of public court opinions really protected by copyright?
The court said yes, because the bar for originality is low. A work needs only "some creative spark." The court pointed to the choices Westlaw editors make: which points of law are important enough to include, and how to word them so the headnote stands on its own while still matching the opinion.
ROSS argued that protecting headnotes would give Thomson Reuters "a monopoly over the law." The court disagreed: "Headnotes are not law; judicial opinions are," and the opinions stay free for anyone to publish. The court also noted there are many ways to express a point of law, so a headnote is not the only possible phrasing of the underlying idea.
One limit is worth knowing. The court did not decide whether headnotes that quote an opinion word for word could be protected. The lower court had granted judgment only on headnotes that did not copy opinion text verbatim.
What is fair use, and how do courts decide it?
Fair use is a defense written into federal law at 17 U.S.C. § 107. Even when someone copies protected material without permission, the copying is not infringement if it counts as a fair use. Courts weigh four factors: the purpose and character of the use (including whether it is commercial and whether it is transformative), the nature of the original work, how much of it was used, and the effect on the market for the original. The Copyright Office explains these factors on its fair use page, which also says there is no formula or set percentage that guarantees a use is fair and that courts decide these claims case by case.
The side claiming fair use carries the burden of proving it. In this case that meant ROSS.
How did each fair use factor come out?
The court found that three factors weighed against fair use and one slightly favored it, then weighed them together.
Factor one, purpose and character, went against ROSS. The use was commercial. The court said ROSS's use shared the same ultimate purpose as Thomson Reuters's own use, helping users find relevant legal material, making it "minimally transformative, at best." The court also said copying the headnotes was an easy route, not a necessary one, since ROSS had free access to the underlying opinions. It added that, to the extent good faith still matters, the record showed ROSS at times acted in bad faith, including attempts to get into Westlaw with an investor law firm's credentials despite notice that Westlaw's terms of service prohibited it.
Factor two, nature of the work, slightly favored ROSS. The headnotes were already published and are more factual than fictional, because they must accurately convey the law.
Factor three, amount used, went against ROSS. ROSS argued it took only 0.08% of Thomson Reuters's 28 million headnotes. The court answered that the question is whether no more was taken than necessary for the copier's purpose, and copying the headnotes was not necessary to train the AI because the underlying opinions were freely available.
Factor four, market effect, went against ROSS. The court found that ROSS's copying hurt the headnotes' value as a draw for Westlaw subscribers, threatened Westlaw's position in the legal-research platform market, and took away Thomson Reuters's chance to license headnotes as AI training data, a market the court called "rapidly developing." It was not persuaded by ROSS's claimed public benefits, noting that the opinions were already free and that ROSS offered no evidence for its claims about halting AI development or national security.
Overall, the court held that ROSS's use "was not fair."
Why does it matter that this case involved legal research?
The facts lined up unusually well for Thomson Reuters. ROSS built a competing legal-research platform at comparable prices, and ROSS aimed to replace Westlaw. The court contrasted an earlier case in which Google scanned books to build a search tool that could point readers toward books they might buy. Here, the court said, ROSS "aims to replace Westlaw."
The same facts shaped the court's footnote on generative AI. The court stressed that ROSS's tool could not generate original expression and that ROSS trained it to create a commercial substitute for Westlaw. A case with different facts could come out differently.
What does this ruling not decide about chatbots and image generators?
It does not decide whether training a generative AI model on your writing, art, or photos is fair use. The court said so directly in a footnote. It described a statement of interest the Department of Justice filed on September 1, 2026, in a separate pending case in federal court in New York, In re: OpenAI, Inc. Copyright Infringement Litigation. That filing relied on a 2025 California federal ruling in Bartz v. Anthropic for the idea that training a large language model that can "generate original responses" is transformative. The footnote also cited Kadrey v. Meta Platforms, a 2025 California federal decision, as "exploring" whether generative AI programs use training materials transformatively.
The Third Circuit said "the concerns raised in that separate case do not apply here." It explained that ROSS's platform "cannot generate original expression" and that ROSS trained its AI to create a commercial substitute for Westlaw. This article does not report how those other cases came out, because the Third Circuit's opinion does not say, and those cases are in different courts.
Three questions are therefore left open by this ruling:
Whether a company may train a generative model on works it did not license.
Whether a chatbot or image tool that produces output resembling your work infringes, which is a separate question from training.
How courts will treat a model that does not compete with the original in the way ROSS competed with Westlaw.
Who does this ruling bind, and can it change?
A Third Circuit decision is binding precedent in federal courts in Pennsylvania, New Jersey, Delaware, and the U.S. Virgin Islands. Federal courts in other regions are not bound by it, though they may find its reasoning persuasive. The Supreme Court has not ruled on AI training and fair use.
The ruling is also tied to its facts and to an early stage of the lawsuit. Further review is generally possible after a decision like this, and the opinion itself does not say what comes next. Anyone relying on this ruling should check whether anything has changed since September 29, 2026.
What can creators do now to protect their work?
No step below guarantees that an AI company will not use your work, and none replaces advice from a lawyer who handles copyright. These are practical moves that keep your options open.
Decide whether you might ever want to sue. If so, learn about registering with the U.S. Copyright Office, which runs the process through its registration page. Federal law generally bars an infringement lawsuit over a U.S. work until the copyright has been registered or preregistered (17 U.S.C. § 411(a)), with limited exceptions. Our articles on whether you need to register to own your work and on what registration gets you explain the tradeoffs.
Keep dated originals: drafts, source files, raw photos, and the date you first published. They help show that the work is yours and when you made it.
Read the terms of service of every platform where you post. Some address how your uploads may be used, and you agree to them when you sign up.
Look for opt-out settings on platforms and use robots.txt on your own website if you want to ask automated crawlers to stay away. These are requests and settings, not legal guarantees, and a crawler may ignore them.
Talk to a copyright lawyer before sending a demand letter. A weak or overreaching demand can backfire, and our piece on copyright demand letters and image troll scams shows how those letters can go wrong.
If you run a business that uses AI tools, see the legal risks of using AI tools in your small business. If you cannot afford a lawyer, a legal aid office or a court self-help center may be able to point you to low-cost help.
This article is general information about one court decision, not legal advice, and it is not a substitute for advice about your own situation.
Frequently asked questions
Did the Third Circuit rule that AI training is not fair use?
Not in general. The court held that ROSS Intelligence's copying of 2,243 Westlaw headnotes to train a legal-search tool meant to compete with Westlaw was not fair use. It called the case "no more than an ordinary copyright case" and said ROSS's AI could not generate original expression.
Does this ruling apply to AI chatbots or image generators?
The court did not decide that. In footnote 7 it separated ROSS's tool from generative models that can "generate original responses" and said the concerns raised in a separate generative AI case in New York do not apply. Whether training a generative model on your work is fair use is still open.
Can an AI company use my art or writing without asking?
This ruling does not give a general answer. Courts weigh fair use case by case using four factors, and the Copyright Office says there is no set formula. The facts matter, including whether the use is commercial, how much was taken, and whether it harms the market for your work.
Is this ruling binding everywhere in the United States?
No. A Third Circuit decision binds federal courts in Pennsylvania, New Jersey, Delaware, and the U.S. Virgin Islands. Courts elsewhere are not bound, though they may consider its reasoning. The Supreme Court has not ruled on AI training and fair use.
What should a creator do if they think their work was used to train AI?
Keep dated originals, save evidence of where your work appeared, and read the terms of any platform that hosted it. If you want to pursue a claim, consider talking to a copyright lawyer before sending any demand, and learn how registration with the U.S. Copyright Office works.
This article is general legal information, not legal advice, and may not reflect the most current law or the law in your jurisdiction. Laws vary by state and change over time. For advice about your specific situation, consult a licensed attorney.
Knowing your rights is the first step
Join thousands committing to calmly and consistently exercise their constitutional rights.
