The first big artificial intelligence fair use legal opinion is out,[1] and it was a surprising loss for AI developer ROSS Intelligence,[2] whose AI training was found to infringe copyrights held by Thomson Reuters,[3] owner of legal publisher Westlaw.

Attorneys and others who care about fair use and emerging technology have watched this case closely as a possible harbinger of things to come in AI litigation.

The Feb. 11 opinion in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., in the U.S. District Court for the District of Delaware, may have little predictive value, however, because U.S. Circuit Judge Stephanos Bibas failed to engage with an important line of case law on intermediate copying, and misapplied the concepts of commercial substitution and superseding use.

These central ideas in fair use law have been part of fair use’s protection of free expression and innovation for decades, and other courts that engage and properly apply them may be more likely to find fair use in AI training cases.

ROSS Intelligence developed a search engine that could answer natural language legal questions with lists of cases most likely to provide useful legal answers, like “When is a dog bite negligent in Illinois?” or “When is a website a ‘public accommodation’ under the [Americans with Disabilities Act]?”

To teach its search engine how to match queries to cases, ROSS trained an AI model using bulk memos prepared by a legal research firm based in part on headnotes written by Westlaw editors. The headnotes themselves are simple restatements of principles in public domain legal opinions. Users of ROSS’ search engine never see any of this intermediary text, which was used only behind the curtain to teach ROSS’ AI tool how to parse legal queries.

The search results delivered to users consisted of public domain legal opinions.

Thomson Reuters’ lawsuit alleged that the intermediate copying involved in training ROSS’ search engine was infringing, and Judge Bibas’ recent opinion agreed, at least with respect to a substantial subset of the allegedly infringed headnotes.

Judge Bibas’ first fair use mistake was in evaluating ROSS’ training process. Making behindthe- curtain uses of copyrighted works as part of a process that results in noninfringing outputs, sometimes referred to as intermediate copying, has been recognized as fair use in a variety of contexts.

Cases involving plagiarism detection,[4] digital research,[5] software development,[6] scholarly manuscripts[7] and search engines[8] all show that fair use analysis focuses on the protected expression a use provides to the public rather than what’s used behind the curtain, especially, but not exclusively, in technology cases.

Although the parties raised some of these cases in their briefs, Judge Bibas’ opinion Brandon Butler mentions only the software development cases and concludes that the intermediate copying idea cannot be applied outside the realm of computer code.

Ignoring the vast majority of the other intermediate copying cases is more than a technical legal mistake — it represents a significant expansion of copyright control at the public’s expense.

As then-U.S. District Judge Pierre N. Leval of the U.S. District Court for the Southern District of New York wrote in Authors Guild v. Google,[9] intermediate copying in that case allowed Google to “augment public knowledge … without providing the public with a substantial substitute for matter protected by the Plaintiffs’ copyright interests.”

When the Authors Guild said these uses should be licensed, Judge Leval responded that “an author’s derivative rights do not include an exclusive right to supply information … about her works.”

Likewise, Thomson Reuters’ copyright in its headnotes does not include an exclusive right to provide researchers with cases that are responsive to natural language queries. Giving copyright holders a veto over intermediate copying deprives the public of new knowledge, and blocks creators and innovators from bringing useful new tools and new works to the marketplace.

The second big fair use idea that Judge Bibas misapplied also involves competition, in particular the role of competition in the U.S. Supreme Court’s 2023 fair use opinion in Andy Warhol Foundation for the Visual Arts Inc. v. Goldsmith.[10] That case involved artist Andy Warhol’s stylized portrait of the artist Prince, based on a more traditional black-and-white portrait by the photographer Lynn Goldsmith.

The Supreme Court ruled that Warhol’s use was not “transformative,” a key part of the firstfactor analysis, citing the direct competition between Warhol’s derivative print and Goldsmith’s photograph in the market for magazine cover illustrations. Competition, in this circumstance, showed that Warhol’s uses were unfairly superseding — and replacing — Goldsmith’s in the marketplace.

Judge Bibas held that ROSS’ competition with Thomson Reuters in the market for legal research platforms was similarly fatal, but this conclusion misreads Warhol and, more importantly, misunderstands the kind of competition that weighs against fair use.

Fair use disfavors superseding uses, which displace a work in the marketplace by offering consumers copies or derivatives of that work. In Warhol, the court started from the premise that Warhol’s print would be an infringing derivative work unless fair use applied. It observed that “[t]o preserve the [derivative work] right, the degree of transformation required to make ‘transformative’ use of an original must go beyond that required to qualify as a derivative.”

The court noted evidence that “photographers generally license others to create stylized derivatives of their work in the vein of the Prince Series.” Warhol’s competition with Goldsmith in the market for magazine covers was fatal because the court had already concluded that Warhol’s work was a derivative of Goldsmith’s, offering the public substantial portions of Goldsmith’s protected expression without a license from Goldsmith.

Not all competition is a superseding use, however, nor is every market effect an unfair one. Fair use encourages some forms of competition and even protects uses that undoubtedly harm their subjects’ markets. The paradigm case in this category is criticism, which may demolish its subject so thoroughly that it undermines public appetite for the work, cratering its sales.

But fair use protects more complex competition and substitution effects as well. A critical essay about a book competes with the book in the market for readers’ finite time and attention. A good essay may provide readers with enough information about the (unprotected) facts and ideas in a book that they no longer feel they need to read the book. A competing Yellow Pages directory organized according to a more useful set of business categories may put the directory from which it took unprotected listing data out of business.

None of these forms of competition or substitution are cognizable by copyright because they don’t compete by offering the protected expression of the work, either in its original form or in a derivative form, to consumers. What’s more, fair use competition creates real public benefits, which the Supreme Court has said should be considered in the fair use calculus in Google LLC v. Oracle America Inc. in 2021.[11]

Taken together, these mistakes make the ROSS opinion a poor bellwether for what’s to come in AI fair use litigation. If Judge Bibas had engaged with the full body of intermediate copying cases and laid out a persuasive case for distinguishing AI training, the outlook for AI companies litigating other cases would be darker today.

Similarly, if Judge Bibas had connected ROSS’ use to actual market substitution, rather than the broader, more amorphous concept of market competition, the opinion would cast a longer shadow. As it stands, however, this first judicial walk through the AI fair use labyrinth has hit a dead end.

Brandon Butler is executive director at Re:Create and a partner at Jaszi Butler PLLC.

Disclosure: Brandon Butler submitted an amicus brief on behalf of accessibility groups BeneTech and Learning Ally in Authors Guild Inc. v. HathiTrust. Butler also helped draft an amicus brief for the American Library Association and other library groups in support of Google in Google LLC v. Oracle America Inc.

The opinions expressed are those of the author(s) and do not necessarily reflect the views of their employer, its clients, or Portfolio Media Inc., or any of its or their respective affiliates. This article is for general information purposes and is not intended to be and should not be taken as legal advice.

[1] Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc., No. 1:20-cv-613-SB (D. Del. Feb. 11, 2025).

[2] https://blog.rossintelligence.com/.

[3] https://www.thomsonreuters.com/en.

[4] A.V. ex rel. Vanderhye v. iParadigms LLC, 562 F.3d 630 (4th Cir. 2009).

[5] Authors Guild Inc. v. HathiTrust, 755 F.3d 87 (2nd Cir. 2014).

[6] Sega Enterprises Ltd. v. Accolade Inc., 977 F.2d 1510 (9th Cir. 1992).

[7] Sundeman v. the Seajay Society Inc., 142 F.3d 194 (4th Cir. 1998).

[8] Kelly v. Arriba Soft Corp., 336 F.3d 811 (9th Cir. 2003).

[9] Authors Guild Inc. v. Google Inc., 804 F.3d 202 (2nd Cir. 2015). The case is a landmark in part because it finally gave a Second Circuit imprimatur, from no less a fair use authority than Judge Pierre Leval, to the digital fair use reasoning that had been deployed in other

[10] Andy Warhol Found. for the Visual Arts Inc. v. Goldsmith, No. 21-869 (U.S. 2023).

[11] Google LLC v. Oracle America Inc., No. 18-956 (U.S. 2021). Judge Stephen Breyer also explained in Oracle v. Google that fair use provides a “context-based check” on the copyright owner’s monopoly, protecting the public and competitors against excess market power derived from copyright. Judge Bibas failed to apply that check.