Note: This article was originally published on the Copyright Alliance blog: copyrightalliance.org/ai-training-not-fair-use. It is republished here with permission.
On February 11, a judge in the U.S. District Court of Delaware issued an opinion in a long-running AI copyright infringement case, rejecting the defendant’s fair use defense and finding it liable for direct copyright infringement. The case was one of the first brought by a copyright owner for the unauthorized use of its works for AI training, and, after many twists and turns, it’s the first to provide an answer to the fair use question. That answer is a clear rejection of the fair use defense as applied to AI training. And while the court explains that Ross’s technology is different than the generative AI technologies at issue in the dozens of ongoing infringement cases, the decision addresses important fair use issues surrounding AI training that could have implications for those other cases.
Background
Back in 2020, Thomson-Reuters (TR), the parent company of the Westlaw legal research service, sued Ross Intelligence, a competitor legal research service, for copyright infringement. The complaint alleged that Ross accessed and copied copyrighted legal content, including Westlaw’s attorney-authored case headnotes, from a Westlaw subscriber to develop a competing product based on machine learning. Days before the trial was scheduled to begin in August of 2024, a hearing was held during which Judge Bibas found as a matter of law that Ross had committed acts of infringement. However, that oral order was rescinded, the trial was postponed, and the parties were directed to bring new summary judgment motions on copyrightability, validity, infringement, and fair use. In October, the parties filed their renewed motions for summary judgment, with Ross arguing, among other things, that the first fair use factor weighs in its favor because its use was highly transformative.
Copyrightability Confirmed and Infringement Found
In the opinion issued this week, Judge Bibas begins by admitting he was wrong to deny TR’s earlier motions for summary judgment on copyrightability and fair use. First, he makes clear that the headnotes and Key Number System copied from Westlaw meet Feist’s low threshold of originality and are therefore protected by copyright—these works were not simply, as Ross argued, uncopyrightable judicial opinions. Then he moves on to the big issue: Is Ross’ copying for AI training an infringement, and if so, is it allowed anyway under the fair use doctrine?
As to infringement, Judge Bibas finds that actual copying occurred for nearly all the headnotes at issue. Judge Bibas says that he himself “slogged through all 2,830 headnotes” and found that for 2,243 of them, “copying is so obvious that no reasonable jury could find otherwise.” Bibas then confirms that those 2,243 headnotes are substantially similar enough to Ross’s “Bulk Memos” that, again, no reasonable jury could find otherwise.
Moving on to Ross’s asserted defenses, Judge Bibas rejects them all methodically, saying that “[n]one of Ross’s possible defenses holds water.” He explains that:
Innocent infringement doesn’t apply because the headnotes bear copyright notices.
Copyright misuse doesn’t apply because the court already ruled against anticompetitive counterclaims.
The merger doctrine was rejected because “there are many ways to express points of law from judicial opinions.”
Scènes à Faire “does not fit” because the headnotes are not stock elements that are required when a judicial opinion is slimmed down.
Clearly, Ross tried throwing the kitchen sink of defenses at the court, but Judge Bibas wasn’t having it.
AI Training Is Not Fair Use
Turning to the main event, the opinion first makes clear that Ross bears the burden of proof for the affirmative defense of fair use, and that here, where there are undisputed facts, the fair use question is a legal one appropriate for a judge to determine. Judge Bibas then analyzes the first fair use factor, explaining that while Ross’s use was admittedly for a commercial purpose, he must also consider whether it was for a transformative purpose. Relying heavily on the Supreme Court’s Andy Warhol Foundation v. Goldsmith decision, the court finds that Ross’s use is not transformative because it does not have a “further purpose or different character” from TR’s. Judge Bibas notes that the parties are competitors and that Ross set out to create a legal research tool that would serve the same purpose as TR’s tool.
The opinion then addresses a claim that’s being made by many AI companies that have been sued for infringement: that training an AI model on copyrighted works constitutes “intermediate step” copying that has been permitted under the fair use exception in past cases—specifically Google v. Oracle, Sony v. Connectix, and Sega v. Accolade. Judge Bibas finds that “those cases are inapt,” mainly because they all involve the copying of computer code, not written works. Importantly, the opinion explains that intermediate step cases that have found in favor of fair use involve copying of functional computer code (usually through the process of reverse engineering) that’s necessary for a competitor to innovate.
Here, the headnotes were expressive works that do not contain unprotected “underlying ideas” that need to be reached through copying TR’s headnotes. If unprotected facts about cases (or the full copies of the cases themselves) were all that Ross was after, it could have easily found and copied those elsewhere—there was no justification for copying the headnotes that would tilt the analysis of this factor in Ross’s favor. That’s also true for almost all the generative AI cases being litigated today, and it’s a fact that AI companies will find hard to refute. Ultimately, Judge Bibas finds that Ross simply took the headnotes to make it easier to develop a competing legal research tool—a use that is in no way transformative.
The opinion then moves to factors two and three, finding that they both favor Ross. Considering the nature of the works, Judge Bibas says that while the headnotes have more than a minimal spark of originality, they’re not as creative as a novel or something an artist creates from scratch. It’s not the strongest part of the opinion, as Judge Bibas gets close to a subjective criticism—of which the Warhol court warned—over the creativity of a work of authorship. Regardless, he notes that factor two rarely plays a significant role in fair use determinations.
As for factor three, the amount or substantiality of the work used, Judge Bibas finds that it favors Ross because, even though the headnotes were copied by Ross, they were not made available to the public. But that question is given disproportionate consideration and weight in his analysis of this factor. He would have been better off focusing on whether Ross copied the heart of the work, which he admits favors TR, but then inexplicably concludes that “Ross wins this factor anyway.”
The most important part of the fair use analysis comes at the end when Judge Bibas considers the fourth factor. Judge Bibas recognizes that factor four, the effect of Ross’s use on the market for TR’s original works, has historically been treated as the most important of the four factors. He explains that courts must not only consider existing markets, but also potential markets. The opinion then confirms a critical point that could have a big impact on all AI litigation: there is an “obvious” potential market for using copyrighted material for AI training. Judge Bibas is right to call this market obvious, as more and more copyright owners are striking deals with AI companies to license works for training purposes.
The fact that there is an established and growing market for use of copyrighted works for AI training—including the works of the plaintiffs in many of the ongoing lawsuits—is something that alleged infringers are going to have a hard time overcoming. The opinion quotes the Supreme Court’s language in Harper & Row v. Nation Enterprises to recognize that factor four “is undoubtedly the single most important element of fair use.” With such weight rightly afforded to factor four, the obvious market (and potential market) harm to copyright owners when their works are used without permission to train large language models, image and music generators, and motion picture models could swing ongoing lawsuits against generative AI companies in favor of copyright owners.
Conclusion
Judge Bibas’s opinion is the first decision in a case that addresses whether AI training is fair use, and it unequivocally holds that it is not. It’s important to recognize that fair use is decided on a case-by-case basis, and this case involves a unique set of facts and a specific technology. That said, it’s hard to see how parts of the decision’s fair use analysis, especially with regard to factor four analyzing market harm, won’t influence the dozens of other ongoing AI infringement cases. At the very least, this decision, combined with preliminary orders favoring copyright owners in other cases, should have AI companies rethinking their reliance on fair use.
Why This Matters for Independent Publishers
This landmark decision carries significant implications for independent publishers, many of whom are concerned about how their content is being used to train AI systems—often without permission or compensation. Here’s why this case is worth paying attention to:
- Legal Precedent Is Taking Shape: As one of the first rulings to directly address whether AI training qualifies as fair use, this case begins to establish a legal framework that future courts may follow. Independent publishers now have a stronger footing to argue that unauthorized use of their content for AI training is infringement, not fair use.
- A Path Toward Licensing Revenue: The decision acknowledges the existence of a legitimate and growing market for licensing content to AI companies. This opens the door for publishers to potentially monetize their works in new ways, assuming they retain the appropriate rights.
- Know Your Rights, Protect Your IP: Many indie publishers may not have considered how their catalogues could be used in the AI ecosystem. This case underscores the importance of understanding copyright ownership, reviewing publishing agreements, and considering whether to opt into or out of AI-related licensing arrangements.
Kevin Madigan is the senior vice president of policy and government affairs at the Copyright Alliance. Madigan joined the Copyright Alliance in early 2020 after four years at the Center for the Protection of Intellectual Property (CPIP) at George Mason University’s Antonin Scalia Law School.