AI Fair-Use Ruling Lays First Marker as OpenAI, Meta Cases Loom

AI companies and the copyright owners suing them in dozens of cases across the country got their first glimpse into how judges will weigh fair use arguments Tuesday in a Delaware court’s opinion rejecting an AI search engine developer’s invocation of the defense.

Generative AI providers including OpenAI Inc., Meta Platforms Inc., and Anthropic PBC have pointed to the fair use doctrine to fend off accusations they infringed copyrights by using material to train their large language models without permission. Judge Stephanos Bibas partly granted Thomson Reuters’ bid for summary judgment Tuesday, finding Ross Intelligence Inc. can’t claim fair use.

The case involves non-generative artificial intelligence, but it will nevertheless reverberate in the dozens of pending cases against generative AI companies.

Bibas found that two of four fair-use factors weighed in Thomson Reuters’ favor: that Ross’ copying of legal summaries called headnotes for a competing product wasn’t transformative, and that it would have a negative commercial effect on Westlaw.

The opinion is the first clear signal that plaintiffs such as New York Times Co. and authors including comedian Sarah Silverman can overcome tech companies’ fair-use arguments, said Bhamati Viswanathan, assistant professor at New England Law.

While plaintiffs still face “real challenges” to win their suits, she said, “what is different now is that at least one court has shown that it is possible to make a successful case, at least at the summary judgment stage, against the defendant tech companies’ fair-use line of defense.”

Plaintiffs in one generative AI suit already latched onto Bibas’ decision.

Music publishers on Tuesday sought permission to highlight it in their pending case in California federal court accusing Anthropic of training its large language models using copyrighted song lyrics. Anthropic quickly fired back opposing the request, calling Bibas’ opinion “irrelevant” to the generative-AI case.

Transformative Use

In 2023, Bibas authored an opinion largely denying Thomson Reuters summary judgment on infringement and fair use. Then, at the last minute, he postponed the trial slated for August 2024. In the Tuesday opinion he said he realized his “prior summary-judgment ruling had not gone far enough.”

This time around, he found the first and fourth fair use factors favored Thomson Reuters. The other two factors—the nature of the original work and the amount used—favored Ross, but ultimately carried less weight, he said.

Santa Clara School of Law professor Eric Goldman disagreed with that weighting and said the second factor is critical to the fair-use analysis because it “provides the necessary breathing room for copying of fact-like materials” such as legal summaries.

“The court provides no rationale for why this ‘tie’ breaks for the copyright owner other than to say that factor two isn’t often important to fair use,” he said in an email.

Ross argued Westlaw’s information didn’t appear as a part of the final product and that the copying occurred only as an intermediate step during the development of its software, which transforms the purpose of the original work. Defendants in other AI cases, including music-generator Suno Inc., have similarly argued that intermediate copying is fair use.

But unlike cases where intermediate copying was permissible, Ross’ copying wasn’t “reasonably necessary to achieve the user’s new purpose,” Bibas wrote, citing the 2023 US Supreme Court ruling in Andy Warhol Foundation for the Visual Arts Inc. v. Goldsmith.

Bibas’ conclusion that Ross’ use wasn’t transformative shows AI copyright cases are “not simply a question of, ‘does the output infringe,’” said Aaron Moss of Greenberg Glusker Fields Claman & Machtinger LLP.

“That is the part of the case that I do think that—notwithstanding the caveats—lawyers are going to be looking to in litigating the 40-plus other cases that do involve generative AI,” he added.

Market Harm

Pointing to the 1985 high court ruling in Harper & Row v. Nation Enterprises, Bibas called the competitive dynamic under the fourth factor the most important in his fair-use analysis.

Lawyers who reviewed Bibas’ opinion disagreed on how his market-harm analysis would translate to cases involving large language models. Ross’ product was “meant to compete with Westlaw by developing a market substitute,” Bibas wrote.

Unlike Ross’ software, AI models such as ChatGPT can be used for many purposes, Moss said, even if it can occasionally spit out copies of articles as The New York Times has alleged.

“It’s certainly not a substitute in the same way that the court seemed to think that Ross’ tool was,” he said. “That obviously gives defense lawyers in these other cases a point of differentiation.”

But Justin Hughes, a professor at Loyola Law School, said tech firms may compete with artists, publications, and authors in a different way: by licensing their models to companies that compete directly with plaintiffs.

AI companies’ profit model isn’t to provide all these services directly to consumers, he said, but to be the back-office provider for something that could use the LLM to compete with or substitute for these creators. A judge may consider who OpenAI licenses their product to when weighing the market factor, he said.

Bibas’ decision isn’t the end of the case, as questions around more than 5,000 headnotes are set for trial. Unless the judge signs off on an interlocutory appeal, partial summary judgment decisions typically aren’t reviewed by a circuit court until the case is fully decided, Hughes said. That would mean this opinion could stand as the only AI fair-use analysis until later, higher-profile cases reach summary judgment.

“If you were counsel for an AI company, you would not like that idea,” he said. “Even though this is not generative AI, it is a fly in the ointment for the generative AI companies.”

Bloomberg Law competes with Westlaw in providing legal research services.

The case is Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., D. Del., No. 1:20-cv-00613.