Fair Use Analysis in ‘Thomson Reuters’ Should Have Generative AI Defendants Worried
On September 29, the Court of Appeals for the Third Circuit issued its decision in Thomson Reuters v. Ross Intelligence, affirming the district court’s summary judgment order in favor of Thomson Reuters and rejecting Ross’s fair use defense. Importantly, the court rejected the notion that “intermediate step” copying during the training process automatically qualifies as a transformative use under the first fair use factor. Instead, the court explained that the analysis must look to the ultimate purpose of AI models and their outputs.
The decision also confirmed a critical point that’s at the center of dozens of ongoing generative AI cases: that harming a copyright owners’ position in potential licensing markets represents cognizable harm under the fourth factor. And while the decision makes clear that the technology at issue in the case is not generative AI, the court’s reasoning on factors one and four (and others) is equally applicable to ongoing generative AI litigation—something that should have defendants in those cases second guessing their arguments.
Don’t Believe the Spin That the Decision is Good News for Generative AI Companies
Soon after the decision came out, some generative AI supporters claimed it was a victory for their side. The Chamber of Progress said the decision is “actually good news” for generative AI companies, which it claimed are the clear winners because the Thomson Reuters court distinguished generative AI from the technology at dispute in the case. They claimed that in doing so, the court was implicitly signaling that fair use would “easily go the other way.” That incorrect interpretation is likely based on a footnote in the decision that references a recent Department of Justice (DOJ) statement of interest in an ongoing generative AI infringement case that says “[u]nlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression.” But distinguishing a case on the type of technology at issue alone does not decide it, and the court’s passing remark in the footnote doesn’t come remotely close to adopting the DOJ’s position on generative AI training. To think after reading the Third Circuit’s decision that generative AI defendants will clearly benefit from it is some serious wishful thinking.
It’s worth noting that other organizations like the Electronic Frontier Foundation (EFF), an organization supportive of broad fair use applications in the AI context, said that the decision “misses the mark” and was generally more critical in its response. So too was the Authors Alliance, which said the opinion “represents several deformations of copyright law” and “hands litigants a set of perniciously pernicious tools.” Like these organizations, supporters of generative AI companies should indeed be concerned, contrary to those who are spinning the decision as a clear win for generative AI and are living in an alternate (or maybe artificial?) reality.
Training is Not Categorically Transformative
One of the most impactful parts of the Third Circuit’s fair use analysis is its confirmation that a transformative use analysis under the first fair use factor does not begin and end at the copying-for-training, or intermediate copying step. The court acknowledged that Ross’s copying of Thomson Reuters’ Westlaw headnotes for training purposes was an intermediate step but looked past it to analyze the step in the context of Ross’s ultimate purpose: developing a competing legal-research platform that serves the same purpose as the original works. The decision explains that while the defendant “took an intermediate step of using the [works] to train an AI program, which arguably presents a slight degree of difference in use,” its use “shares the same ultimate purpose” as the plaintiff’s use, making it “minimally transformative,” at best.
The court also rejected the defendant’s reliance on past intermediate copying cases—like reverse engineering cases—because in those cases “copying was necessary to access the unprotected functional aspects of computer code.” The decision explains that, unlike in those cases, Ross had access to the judicial opinions but copied protectable Westlaw headnotes because doing so was “easy,” and “[u]nlike necessity, ease is not a justification for copying.” That’s a crucial point because generative AI defendants cite to many of the same past fair use cases to make the “necessity” argument. But when licensing alternatives exist, and the market for licensing for AI training continues to grow, unlicensed copying is a matter of convenience and cost, rather than a justifiable necessity.
Cognizable Market Harm Takes Many Forms
What might have the most far-reaching impact in generative AI litigation is the Third Circuit’s discussion of market harm under the fourth fair use factor, which correctly recognizes three major forms: harm to (1) the original market for copyrighted works, (2) the value of the work, and (3) potential derivative markets (such as those for licensing works for training.)
Before getting to potential derivative markets, which is rightfully getting a lot of attention, it’s worth looking closer at harm to the value of a work because it’s sometimes overlooked in fair use analyses. The plain language of Section 107 of the Copyright Act requires consideration of “the effect of the use upon…the value of the copyrighted work,” but courts often collapse “market” and “value” into one undifferentiated inquiry. The reality is that the value of a work can take forms that don’t fit neatly into recognized markets.
The Third Circuit uses the example of the unauthorized publication of movie trailers that damaged the trailers’ value as draws to other cross-market works. The decision explains that while there may not be a standalone market for Thomson Reuters Westlaw headnotes, the headnotes are a feature that attracts users to buy a Westlaw subscription, and “by copying the headnotes and using them to build its own competing legal-research platform, ROSS appropriated the headnotes’ value for finding and analyzing judicial opinions and diminished their value as a draw for users to Thomson Reuters’s legal-research platform.” This same reasoning can and should be applied to any context where the value of a work is not directly tied to an existing or potential market, whether that be related to generative AI or not.
The highlight of the court’s factor four analysis comes in its discussion of harm to potential derivative markets, confirming that harm to the emerging market for AI training licenses is cognizable under the fourth factor. Ross had argued, as many generative AI defendants do, that no such market exists (or has been exploited) for the works at issue in the case. But the Third Circuit explains that “a potential derivative market is not illusory just because an author ‘has evidenced little if any interest in exploiting this market for derivative works,’” and the fact that the copyright owner “did not license its [works] to others does not disprove that a market exists to do so.”
Finding that “the market for licensing [the works] as text to train AI is rapidly developing,” the court held that by copying the works “for use as training data without . . . authorization,” the defendant “usurped [the copyright owner’s] opportunity to enter that derivative market and license its [works] for that purpose,” and that this “market harm . . . weighs against fair use.” The same is true in generative AI litigation, where some defendants claim that plaintiffs suffered no harm because they have not themselves licensed their books for AI training. That argument is on much shakier ground after Thomson Reuters, regardless of the differences in the underlying AI model technology in the cases, in no small part due to the fact that the generative AI training-license market is far more developed now, with hundreds of deals across news, publishing, music, and images.
This key holding also indirectly refutes the “circularity” argument invoked by many generative AI defendants. That argument basically claims that a lost licensing fee can’t count as market harm because any unlicensed use could be licensed, so the harm assumes the use isn’t fair. But the Third Circuit correctly treated the training-license market as it would any other derivative market. That approach is especially applicable in situations where a licensing market already exists, which is clearly the case with generative AI licensing.
Conclusion
Defendants will keep saying Thomson Reuters is not a generative AI case. That is true, and it also doesn’t matter. The Third Circuit rejected arguments, not a technology, and they are the same arguments generative AI defendants are making in courtrooms across the country: that copying for training is an intermediate step that transforms the use, that copying is justified when it’s the easiest path, and that a licensing market doesn’t count until the copyright owner has entered it. Courts hearing those cases now have appellate reasoning on each point. Ross tried to convince the Third Circuit that its case was about the future of AI, and the court answered that it was “no more than an ordinary copyright case.” Generative AI defendants would be wise to take note.
To stay up to date with the latest news in artificial intelligence (AI) and copyright, sign up for our AI Copyright Alert. You can also visit our AI and Copyright hub for additional resources on federal court cases, current licensing, and more. Additionally, if you aren’t already a member of the Copyright Alliance, you can join today by completing our Creator Membership form! Members gain access to monthly newsletters, educational webinars, and so much more — all for free!