Last week saw the release of the first appellate court opinion on fair use in AI training, and it was a dud. In Thomson Reuters GMBH v. Ross, Third Circuit Judge Montgomery-Reeves makes a series of errors that leads inexorably to the wrong outcome on both copyright eligibility and fair use. As Adam Eisgrau explained shortly after the opinion was published, a lengthy disclaimer in footnote 7 should contain the potential fallout of this case for the broader AI ecosystem. But bad law is bad law, and this one is bad.
When the District Court issued its opinion siding with Thomson Reuters (TR), I wrote a piece for Law360 explaining the two big ideas that he got wrong: intermediate copying and market substitution. Sadly, the Circuit Court doubles down on both mistakes.
First, they follow the District Court in taking an arbitrarily narrow view of what is sometimes called “intermediate copying.” Intermediate copying is a form of fair use that happens all the time: it’s making a copy as part of an internal process of research and development that results in the creation of an original, non-infringing work or tool. In Sega v. Accolade, for example, Accolade copied the software running on the Sega Genesis video game console in order to reverse engineer its source code and develop new games that could be played on the console. Accolade never published the Sega code; they only published their new games. When Sega sued, the court rejected their attempt to turn a copyright in the console code into a monopoly over the entire market of Genesis-compatible games. Making copies of the Genesis code to learn how it worked and what it takes to make a compatible game is protected by fair use. Intermediate copying often has this pro-competitive aspect: it allows new creators to operate in a realm that would otherwise be dominated by incumbents.
Unfortunately, Judge Montgomery-Reeves followed Judge Bibas in applying the concept of intermediate copying only when copying was “necessary” for a subsequent use. Both judges took the software interoperability cases like Sega, where copying was seemingly necessary, to define the outer limit of that form of fair use. But there’s no reason to read the concept so narrowly.
Many, many fair use cases follow the intermediate use pattern, copying protected works as part of a process that results in new, non-infringing works. Any search engine, whether for the web or for books or images, is an intermediate fair use: the search corpus is copied as an intermediate step to enable the creation of the search tool. Plagiarism detection software is another example, where again the copying of previous works is an intermediate step to creation of the tool.
Ross itself is a good example of why intermediate copying should be understood expansively. The West headnotes at issue in Ross are like a map of a public domain territory—the land of the law. Just as any cartographer is free to consult previous maps for the factual information they contain, so Ross should be free to consult the West headnotes as a source of information about the law, and to use that information to inform its own attempt to map the same legal territory using AI. Requiring each new mapmaker to go out and re-explore the same territory before they can depict it in a new map is wasteful, even dangerous. It’s also bad for the public. Once one map has reliably documented the facts, other mapmakers have always been free to re-represent those facts using new aesthetic and technological approaches. The Ross search tool is just a new map of the law, and there’s no shame in using the old map as a reference in the process.
The second big mistake is closely related: like Judge Bibas, the Third Circuit got tripped up on the fact that Ross is competing with Thomson Reuters in the market for legal research software, equating that competition with the kind of unfair substitution that is “copyright’s bête noir” according to the Supreme Court in Warhol v. Goldsmith. But as Chad Rutkowski points out in his breakdown of the opinion, competition isn’t the same as substitution.
Plenty of fair uses compete with the works they use. A documentary about movies is…a movie, and it competes with all of the movies it quotes from in the market for your scarce movie-watching dollars. For that matter, a critical essay about a movie also competes with that movie for your scarce entertainment dollars, time, and attention. The Supreme Court recognized in Campbell that the market effect of a scathing review is not the kind of effect that copyright cares about. Neither is the “substitution” that always results when someone chooses one product over another. The substitution that matters is when your use of the work provides copies of that work to consumers, satisfying their interest in that work, not just their general interest in information or entertainment.
One way the Third Circuit could have avoided this confusion is by relying on the most important holding in Warhol: that each use must be carefully defined and evaluated on its own terms. Ross’s specific use of the West headnotes was as part of a training process for an AI tool. Through a series of steps, information from headnotes was distilled and combined into a new dataset that was in turn used to train ROSS’s AI tool. The process resulted in a new work: a model that was not a copy of the headnotes, nor did it offer consumers copies of the headnotes. Rather, by training on the data derived from the headnotes (among other resources), the model developed the ability to match questions about the law with relevant opinions. That use—transforming static data into a dynamic model—is indistinguishable from the use found highly transformative in both Bartz and Kadrey.
Ross has said they will appeal this decision to the Supreme Court. If the Supreme Court takes the case, it may be because they want to correct the 3rd Circuit’s misapplication of its rulings on specific uses and market substitution in Warhol.