Reuters v. ROSS: District Court First to Consider Whether Training Generative AI Model Is Fair Use
Reuters v. ROSS: District Court First to Consider Whether Training Generative AI Model Is Fair Use
A federal district court in Delaware has issued a summary judgment ruling in Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc. The decision addresses several important issues:
(2) whether a plaintiff who has registered the copyright in a compilation can allege separate infringement claims for each individual item within that compilation if the plaintiff has not registered the copyright in the individual items; and
(3) the big open question regarding generative AI: whether reproducing third-party copyrighted materials to train a generative AI model constitutes fair use, at least in some cases (and how the Supreme Court’s decision in Warhol v. Goldsmith affects this analysis).
Here, we focus on the last of these three issues.
ROSS Intelligence, a now-defunct (thanks to this and other litigation) legal-research startup tried to develop an AI-powered natural language search engine. Users would input legal questions and the ROSS search engine would output verbatim extracts from judicial opinions—and only verbatim extracts.
For generative AI models to work well, they must be trained on a massive amount of training data. Westlaw, owned by Reuters, provides access to a large database of judicial opinions and other potentially useful training data, notably including “headnotes” that summarize specific points of law from the cases in the database. Reuters registered the copyright in its overall compilation of legal materials but did not individually register each headnote.
Reuters refused to grant ROSS a license to use the Westlaw database to train the AI model powering the ROSS search engine, so ROSS approached LegalEase Solutions instead. LegalEase engaged subcontractors to produce the “Bulk Memo Project,” a collection of approximately 25,000 memos each containing one question and four to six answers, some of which were created using a text-scraping bot. Reuters alleged that the memo questions were essentially just Westlaw headnotes with question marks at the end. ROSS reproduced the memos in the course of training its AI model, and Reuters sued for copyright infringement (direct and secondary). Reuters’ direct infringement claim raised two central questions:
(1) Was Reuters’ copyrightable expression reproduced in the memo questions?
(2) If so, was ROSS’s use of that material for training its AI model to power ROSS’s search engine a fair use?
We will skip over the arguments about whether the Westlaw headnotes are, in fact, protected by copyright. The short answer is that the court declined to hold that the headnotes were copyrightable as a matter of law and left that issue for the jury to decide on a headnote-by-headnote basis. (One intriguing question is whether any of Reuters’ attorney-editors used (with or without Reuters’ knowledge) generative AI tools to help draft the headnotes, which could render those headnotes unprotectable even if they appear to contain expression. This is a question we expect to see raised in many infringement cases moving forward. But we digress.)
Assuming some of the memo questions contained protectable expression taken from Reuters’ headnotes, the court moved on to address the parties’ cross-motions on ROSS’s fair use defense. The court ultimately concluded that the fair use question must also go to the jury. While the court did not rule one way or the other on the fair-use question, its analysis nonetheless provides some insight into the way courts consider fair-use issues related to AI.
The fair use analysis balances four factors. Factor one, the second most important factor, considers the purpose and character of the use. In practice, the factor one analysis focuses on two things: (1) commerciality and (2) transformativeness.
The Supreme Court has fluctuated over the years in the significance it ascribes to whether an allegedly infringing use is for a commercial or profit-making purpose, with this consideration arguably becoming less important in recent years. Nonetheless, the Supreme Court in Warhol made clear that, while a commercial purpose does not make a use presumptively unfair (as the Court suggested, albeit in dicta, in the Betamax Case), commercial use does matter. That said, although Reuters cited Warhol for the proposition that commercial use weighs heavily against a finding of fair use, the ROSS court was not convinced, stating that it would not “overread” Warhol.
Instead, the court here distinguished between Warhol, where both commerciality and transformativeness weighed against a finding of fair use, and Google, where the use was commercial but transformative as a matter of law and the case dealt with new technology. The ROSS court clearly found Google the more apt of the two cases. As a result, the court focused squarely on transformativeness.
The more the use of an allegedly infringing work is for a purpose, or of a character, that diverges from that of the original, the more “transformative” that work is. Here, the parties disputed how transformative ROSS’s use of the headnotes was.
Reuters argued that ROSS intentionally used the Westlaw headnotes to train ROSS’s model in order to copy the creative decisions of Westlaw’s attorney-editors. ROSS argued that it used the headnotes to create a search tool that would produce relevant quotes from judicial opinions in response to natural language questions, which radically departs from the original purpose and character of the headnotes. According to ROSS, its reproduction of headnotes during training was a form of intermediate copying analogous to the intermediate copying the Court looked upon favorably in Google.
While the court was unwilling to pick between the parties’ characterizations at the summary judgment stage, it stated that ROSS’s intermediate copying “was transformative . . . if ROSS’s AI only studied the language patterns in the headnotes to learn how to produce judicial opinion quotes.”
It is noteworthy that the court’s language is consistent with a finding of fair use (with respect to reproductions in training) even if the ROSS search engine sometimes, by accident, generates verbatim headnotes.
Ultimately, the jury will determine ROSS’s purpose in training its AI model on the Westlaw headnotes, but the court’s unqualified statement suggests that if the jury agrees with ROSS’s characterization of the purpose of its training, it should find for ROSS on transformative under factor one.
Factor two considers the nature of the copied work. All else being equal, the less creative the work, the more the factor favors the defendant. While the jury must determine how creative the headnotes are, the ROSS court made clear that the factor “seems to favor fair use.”
Typically, the extent to which a work is copied (both in terms of amount and substantiality) determines the result of the factor three analysis. However, the ROSS court highlighted that even large amounts of copying can be fair use if the amount was connected to a transformative purpose. Citing the Second Circuit’s decision in Authors Guild v. Google, the court stated that “verbatim intermediate copying has consistently been upheld as fair use if the copy is ‘not reveal[ed] . . . to the public.’”
The parties disputed what the correct “work” was for purposes of Reuters’ infringement claim: ROSS argued that the infringement analysis should focus on the compilation (i.e., the entire Westlaw database), while Reuters argued that ROSS infringed each individual headnote. The court sided with Reuters, treating each headnote as a separate “work.” As a consequence, the jury must determine whether outputs of the ROSS search engine included Reuters’ copyrightable expression (if there is any) from the individual headnotes and whether the “scale of copying (if any) was practically necessary and furthered [ROSS’s] transformative goals.”
The court did not clarify what “scale of copying” the jury would be required to determine whether “practically necessary.” Is the question whether (1) the copying of each specific individual headnote was practically necessary, (2) ROSS needed to copy the headnotes more broadly, or (3) ROSS needed to copy a large quantity of data to train its AI model more generally? In many cases, developers of AI models need massive amounts of training data but do not need any specific piece of training data. It would be surprising if the fair use analysis favors defendants who train their models on particular works because of the unique value of those works, perhaps indicating the importance of the copyrightable expression contained in such individual works, over defendants who have no particular need for the plaintiff’s individual works but simply require a large quantity of data.
Interestingly, the court went beyond the standard factor three analysis, which looks at the extent to which the allegedly infringed work is copied, and more generally concluded that “if ROSS’s AI works the way that it says, it is likely fair use because it produces only the opinion, not the original expression.” This comment suggests that, regardless of the extent to which a work is copied at the training stage, whether or not the output of an AI model reproduces copyrightable expression from the underlying work may also influence the fair use analysis with respect to reproductions at the training stage.
The ROSS court, citing Authors Guild, noted that the fourth fair use factor requires an assessment of “whether the use had a ‘meaningful or significant effect’ on the value of the original [work] or its potential market.” The court clarified that “not all losses are created equal,” and “[i]f the source of the loss is not that the original’s expression is being appropriated,” then the fourth factor may favor the defendant. Put differently, while technological innovation reshapes markets, plaintiffs cannot fasten the losses they may suffer as a result of that change to the dependable wagon of copyright litigation and expect innovators to foot the bill.
Here, the parties disputed what the relevant market was. The court identified two possible markets:
(1) Westlaw itself, as a legal research platform; and
(2) Westlaw’s data.
As for (1), the court concluded that the ROSS search engine would not be a market substitute for Westlaw if it created a new research platform serving a different purpose from Westlaw’s purpose. Reuters, though, alleged that some ROSS customers canceled their Westlaw subscriptions.
On (2), while ROSS argued that Westlaw would never license its data as training data so there was no market for it, the court again said it was up to the jury to make that determination.
Finally, the court looked at the public benefits ROSS’s copying will likely produce. Once again, the parties disagreed. ROSS argued the ROSS search engine increased access to legal resources at a lower cost, while Reuters argued all ROSS did was disincentivize companies like Reuters from creating headnotes. The court found that “each side presents a plausible and powerful account of the public benefit that would result from ruling for it,” so a jury must decide between them.
ROSS draws attention not only to the fact-intensive nature of fair use analysis, but also to the underappreciated role played by the fact-finder’s normative evaluation of generative AI. As discourse progresses around what an AI-filled future would or should look like, prevailing cultural narratives on generative AI may begin to coalesce around either side’s preferred story, with great legal consequence. It will be interesting to see who recognizes that fact and, of those, who proactively seeks to shape it.