Thomson Reuters v. Ross: the first AI fair use appeal turns on substitution

A federal appeals court has ruled for the first time on whether training an AI system on copyrighted material is fair use, and the AI company lost. On Sept. 29, a unanimous Third Circuit panel affirmed that Ross Intelligence infringed 2,243 Westlaw headnotes and rejected its fair use defense, Copyright Lately’s Aaron Moss writes. Headnotes are the short summaries Thomson Reuters editors write for each point of law in a court opinion. Ross won’t feel the loss. The cost of the lawsuit put it out of business in early 2021, nearly two years before ChatGPT launched.

Ross built a legal search tool that answered plain-English questions by returning passages from court opinions. It never generated text of its own. To train it, Ross hired a vendor, LegalEase, to write about 25,000 “bulk memos” that paired legal questions with possible answer passages, and LegalEase used Westlaw headnotes to frame many of the questions. Judge Stephanos Bibas granted Thomson Reuters summary judgment last year and sent the ruling straight to the appeals court.

Fair use is the part of copyright law that lets some copying happen without permission. Courts weigh four factors: the purpose of the use and whether it transforms the original, the nature of the work, how much was taken, and the harm to the original’s market. Ross lost three of the four.

“Under ROSS’s framing, this case appears to concern the future of AI legal technology,” Judge Tamika Montgomery-Reeves wrote for the panel. “But appearances can be deceiving. In truth, this is no more than an ordinary copyright case.”

Both sides of the AI copyright fight claimed the win anyway. The Authors Guild said the ruling “vindicates copyright law.” Adam Eisgrau of the tech industry group Chamber of Progress posted that it “implicitly confirms that highly transformative gen AI training … is likely fair use!”

The panel’s reasoning keeps returning to one finding. It called Ross’s use commercial and “minimally transformative, at best,” because both companies used the headnotes “to create and optimize a legal-research platform,” and Ross admitted its product “aims to replace Westlaw.” Ross pointed to older cases that allowed copying code as a necessary step toward building something new. The court answered that Ross had millions of free judicial opinions to work from and used the headnotes because they were “an easy way” to frame its training questions. “Unlike necessity,” the court wrote, “ease is not a justification for copying.”

The line rights holders will quote sits in the market-harm factor. The court said Ross “usurped” Thomson Reuters’ chance to license its headnotes as AI training data, a market it called “rapidly developing.” Moss points out the court cites no record evidence that the market exists, only that Ross “offers no evidence to disprove this.”

On Sept. 1, the Justice Department filed a statement in the consolidated In re OpenAI cases calling large language model training “exceedingly transformative.” The panel addressed that in a footnote and set Ross apart from the chatbot cases two ways. “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression,” it wrote. It also stressed that Ross trained its system “for the purpose of creating a commercial substitute for Westlaw.”

Those two reasons lead to two different rules. If what protects a model is that it can write something new, every chatbot clears the bar. If what sinks a model is that it was built to take the customers of the work it learned from, the generative cases get harder for the defendants.

Ross’s users never saw a headnote. It lost because its product competed with Westlaw. A chatbot that answers the question a news story answered, or prints the lyrics a music publisher licenses, sits closer to that fact pattern than the AI industry’s victory posts suggest. The substitution reading is also the one that holds up better, because the ability to generate text says nothing about whether the output takes the original’s buyers.

The next court to sort it out is close. Judge Eumi Lee is scheduled to hear summary judgment arguments on Oct. 21 in Concord Music v. Anthropic, the music publishers’ suit over song lyrics used to train Claude, and the publishers have already pointed her to Ross. Briefing in In re OpenAI wraps up in November.

What’s left of Ross has vowed to take the case to the Supreme Court, Futurism reports, which would put the first big AI copyright question in front of the justices on a search tool that never wrote a sentence.

Sources

• Copyright Lately: Everyone Won the Ross Fair Use Appeal. Except Ross.
• Third Circuit opinion, Thomson Reuters Enterprise Centre v. Ross Intelligence (via Copyright Lately)
• Futurism: Court Finds That Using Copyrighted Material as AI Training Data Is Not Fair Use