First Appeals Ruling on AI Training: What ROSS Really Means

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit became the first federal appeals court to rule on whether copying copyrighted material to train an AI system is fair use. The answer, in Thomson Reuters v. ROSS Intelligence, was no. Headlines framed it as a blow to AI training. So the question every builder, publisher and AI user is now asking is simple: does this ruling mean training AI on copyrighted work is illegal?

The short answer

No, not as a general rule. The court decided a narrow case: a non-generative legal search tool, built to replace Westlaw, trained on material derived from Westlaw's own editorial summaries when the underlying court opinions were free to use. The judges went out of their way to say that generative models may present a different picture. What the ruling does do is harden three pressure points that every AI developer should now take seriously: whether the product competes with the source, whether the copying was actually necessary, and whether a market for licensing training data exists.

What the court actually decided

The published opinion (No. 25-2153, argued June 11, 2026, filed September 29, 2026) was written by Judge Tamika Montgomery-Reeves for a panel that also included Judges Restrepo and Bove. It affirmed the February 2025 summary judgment by Judge Stephanos Bibas on two certified questions: whether Westlaw headnotes are original, and whether ROSS's use of them was fair use. The judgment was first entered with the opinion under seal; the sealing was reported on September 29 and the text became public on September 30.

The facts that mattered

ROSS wanted an AI search engine that would answer plain-language legal questions with relevant passages from judicial opinions. To train it, ROSS hired LegalEase Solutions, which wrote about 25,000 training memos. According to the opinion, memo drafters "used thousands of Westlaw headnotes to create the memos' legal questions." The infringement finding covers 2,243 headnotes. Crucially, the court stressed that "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."

Headnotes are protected

On the first question, the court held that "all 2,243 independently created headnotes at issue in this appeal have 'some creative spark,'" because editors "made creative decisions about which points of law were important enough to include in a headnote and how to word those points of law." It rejected the argument that the ideas and the expression merge, noting "there are many ways to express points of law from judicial opinions."

How the four fair-use factors fell

• Purpose and character: against ROSS. "Given that ROSS's use was highly commercial and minimally transformative, we conclude that the first factor weighs against fair use." The court stated plainly: "ROSS aimed to replace Westlaw."
• Nature of the work: for ROSS. Headnotes are largely factual, and the court concluded, "This factor thus favors fair use." It noted this factor rarely drives the outcome.
• Amount used: against ROSS. The court pointed out that "ROSS had access to the underlying judicial opinions and could freely copy them," but chose the headnotes because they offered "an 'easy' way to create its training memos."
• Market effect: against ROSS. The court found that "the market for licensing headnotes as text to train AI is rapidly developing," and concluded, "Thus, the fourth factor weighs against fair use."

The bottom line: "Evaluating these factors together, and in light of copyright's purposes, we hold that ROSS's use was not fair."

Why this is not a verdict on ChatGPT, Claude or Gemini

The most important sentence for the wider AI industry is the one distinguishing generative models. The court noted that the U.S. Department of Justice had filed a statement of interest in the consolidated OpenAI copyright litigation in New York, relying on Bartz v. Anthropic to argue that training a large language model is transformative. The panel's response: "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."

That leaves the existing trial-court map intact. In June 2025, Judge William Alsup ruled in Bartz v. Anthropic that training on lawfully acquired books was "exceedingly transformative" fair use, while refusing fair-use protection for millions of pirated copies. On September 1, 2026, the DOJ told the Southern District of New York that training on copyrighted material "in and of itself" does not violate copyright law, though its filing did not address pirated datasets and treated outputs separately.

Commentators have drawn the same line. LawNext noted that ROSS, now shut down, copied the material before generative AI was widespread and built a competing research product, not a generative system. Copyright Lately called it the first federal appellate ruling on fair use in AI training, with implications for pending generative AI cases.

Where the ruling still bites generative AI

Being narrow is not the same as being irrelevant. Three parts of the reasoning travel well beyond ROSS:

1. Substitution beats technique. The court did not treat "it was only an intermediate training step" as a shield. What mattered was that the finished product competed in the same market as the source. A generative tool built to stand in for a specific publisher's product may face the same logic.
2. Licensing markets count. The finding that a market for licensing text "to train AI is rapidly developing" matters for every sector where AI data-licensing deals are being signed. The more such deals exist, the easier it becomes for rights holders to argue lost licensing revenue under factor four.
3. Convenience is not necessity. Because free public-domain opinions were available, copying the paid editorial layer looked like a shortcut, not a requirement. Developers who train on premium, curated versions of freely available material should expect the same question.

What we know vs. what we're inferring

Known (from the opinion and cited sources):

• The Third Circuit affirmed that 2,243 Westlaw headnotes are copyrightable and that ROSS's use was not fair use.
• Three of four factors went against ROSS; the nature-of-the-work factor favored it.
• The opinion expressly distinguished generative models in Bartz and the OpenAI litigation.
• The DOJ has argued in New York that LLM training is transformative, without addressing pirated data.

Inferred (our analysis, not holdings):

• Rights holders are likely to cite the market-substitution and licensing-market reasoning against generative AI companies, especially where outputs compete directly with the source.
• The decision binds only courts in the Third Circuit (Delaware, New Jersey, Pennsylvania and the U.S. Virgin Islands); other circuits may reason differently.
• The weight a court gives to licensing markets will likely grow as more AI data deals become public.

What to watch next

• Further review in ROSS: whether ROSS seeks rehearing by the full Third Circuit or Supreme Court review. Because this was an interlocutory appeal, the remaining proceedings return to the District of Delaware.
• In re: OpenAI (S.D.N.Y.): how the court handles the DOJ's September 1, 2026 statement of interest, and whether it treats the ROSS opinion as persuasive.
• Other circuits: the first appellate ruling on a generative model will be the real test; until then, ROSS is the only appellate data point.
• Licensing deals: each new AI content-licensing agreement strengthens the factor-four argument this court accepted.

For anyone building or buying AI tools, the practical takeaway is clear: know where your training data came from, whether a licensing market exists for it, and whether your product substitutes for the source.

Educational information only.