Published on Feb 19, 2025
The recent summary judgment ruling in the first-ever AI training data copyright lawsuit may give courts some guidance on how to consider generative AI copyright cases. If so, that’s good news for copyright holders.
In Thomson Reuters v. Ross Intelligence, Judge Stephanos Bibas, a Third Circuit judge sitting by designation in the Delaware District Court, found largely in favor of copyright holder Reuters (TRI) and against Ross Intelligence, a startup that used Reuters’ copyrighted work to train its AI model. While the lawsuit, filed in 2020, predates the emergence of generative AI and has some key differences, there are a few important elements of Judge Bibas’ ruling that experts say can be applicable to generative AI lawsuits.
“On a macro level, this is a big win for the copyright owners,” said Matt Topic, partner at Loevy + Loevy, which represents several news publishers in copyright cases against AI companies. “Even if there are some differences between the cases, this decision is very useful.”
A brief summary of the case: Reuters owns Westlaw, a leading legal research platform. Ross developed an AI-powered legal search platform, where users asked it legal questions and got relevant sections of judicial opinions in response. Ross was accused of using Westlaw’s copyrighted content—including “headnotes,” or summaries of legal concepts written by Westlaw editors—to train its AI. This was after Reuters refused to give Ross access to Westlaw, seeing Ross as a potential competitor. Ross then contracted a third party that did have access to Westlaw to create a series of legal questions and answers, based on (and, Reuters alleged, including) Westlaw’s headnotes, which Ross’ AI trained on.
For its part, Ross disputed that it copied any protected content, but said that even if it had, it had a fair use defense for doing so. It also had some counterclaims against Reuters for antitrust violations, but those were previously dismissed.
Reuters said in the statement that it was “pleased” with the decision. Ross’ legal team declined to comment, and Ross went out of business years ago, due at least in part to this litigation.
Judge Bibas denied partial summary judgment in September 2023, finding that, while copying had occurred, there were genuine factual disputes around copyright infringement and fair use that had to be decided by a jury. In August 2024, the trial was continued in order to re-hear summary judgment arguments. This time, Judge Bibas found in favor of Reuters.
Experts pointed The Capitol Forum to two key parts of Judge Bibas’ ruling that copyright holders in generative AI training cases should be happy about: his finding that there is a market for training data that can be harmed when AI companies use copyrighted content without authorization; and his finding that intermediate copying, a fair use defense that many generative AI companies are fond of, does not apply to the written word.
Intermediate copying is not a get out of jail free card for AI training. As The Capitol Forum explored in-depth last month, AI companies are hoping they will have an “intermediate copying” fair use defense, often citing Ninth Circuit cases Sega Enterprises v. Accolade and Sony Computer Entertainment v. Connectix. In a December 2024 summary judgement hearing, Reuters’ lawyer said intermediate copying was “not a get out of a jail free card.”
The rest of this article is currently locked and only available to subscribers. Request a trial to receive unlimited access to all articles.
