Thomson Reuters v. ROSS Intelligence: training an AI tool on Westlaw headnotes was not fair use

Status

• 29 Sep 2026Opinion of the Court filed. The court will “AFFIRM” the District Court’s order granting Thomson Reuters partial summary judgment.
• BelowUnited States District Court for the District of Delaware, Judge Stephanos Bibas, a Circuit Judge sitting by designation. The District Court certified two questions for interlocutory appeal: whether the headnotes and the Key Number System are original, and whether ROSS’s use of the headnotes was fair use.
• Not decidedThe originality of the Key Number System, which ROSS did not brief and the court treated as forfeited; and whether headnotes that quote an opinion word for word can be original, which the court left open.
• The AI“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 trainingA contractor wrote about 25,000 training memos, each a legal question with four to six passages from judicial opinions in response. Its drafters used thousands of Westlaw headnotes to frame the questions, because the headnotes provided “an easy way [to] fram[e] questions.”

Originality: the headnotes are protected

“All 2,243 independently created headnotes at issue in this appeal have ‘some creative spark.’”

For each headnote, the court said, the editors decided which points of law mattered and how to word them so that each would stand on its own while reflecting the opinion accurately. It rejected the argument that protecting headnotes would give a monopoly over the law: “Headnotes are not law; judicial opinions are.” It also rejected the merger doctrine, because there are many ways to express a point of law, and held that each individual headnote is a copyrightable work.

Fair use: three factors against

• PurposeCommercial and “minimally transformative, at best”. Training an AI program was an intermediate step, but “ROSS’s use of the headnotes shares the same ultimate purpose as Thomson Reuters’s use”: a platform to help users find legal material. Against fair use.
• NecessityThe judicial opinions were freely available, and ROSS could have built its training memos from them. “Unlike necessity, ease is not a justification for copying.”
• NatureThe headnotes were published and are more factual than fictional. This factor favored fair use, “slightly” in the court’s final weighing.
• AmountEach headnote copied was an entire work, and the copying was not reasonable without a transformative purpose. Against fair use.
• MarketThe court found harm to the value of the headnotes to Westlaw’s users, to the market for legal-research platforms, and to a derivative market: “the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing.” Against fair use.

“By copying the headnotes for use as training data without Thomson Reuters’s authorization, ROSS usurped Thomson Reuters’s opportunity to enter that derivative market and license its headnotes for that purpose.”

On the public benefit argued for AI development, the court found no evidence to support it, adding: “While some AI technology may implicate national security concerns, this does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.”

What the court set aside

The court distinguished generative AI. In footnote 7 it noted a Department of Justice statement of interest, filed in a pending generative AI case in the Southern District of New York, arguing that training a large language model is transformative. “The concerns raised in that separate case do not apply here.” The reason it gave: “ROSS’s AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness.” In this case, the court said, “ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw.”

Why it is on this list

A federal court of appeals has now decided a question about AI training on the record of what went into the training: which materials were copied, by whom, into which documents, for which product. The court did not need to look inside the model. It looked at the training memos, the headnotes they were built from and the purpose they served, and it relied on undisputed evidence of each.

It also recognised a market in licensing text as AI training data, and treated taking that text without authorization as harm to it. Where the training material came from, and whether its owner had authorized that use, were central to the court’s analysis.

Sources

• Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc, No. 25-2153 (3d Cir. Sept. 29, 2026), Opinion of the Court, thirty-two pages, read in full at the Third Circuit’s own site

Related

• The law library
• California · AB 2013, training data transparency
• EU AI Act · Articles 53 and 55, general-purpose AI models

Checked against the opinion, read in full. Information, not legal advice.