The concept of transformative use has emerged as a pivotal issue in fair-use analysis, particularly in cases that involve training data for artificial intelligence (AI). At its core, the transformative-use inquiry asks whether the new work repurposes original material to serve a markedly different function or market than that of the copyrighted work.
In this regard, the revised summary judgment handed down today in Thomson Reuters v. Ross Intelligence provides a significant touchstone, as the U.S. District Court for the District of Delaware held that Ross Intelligence’s use of Westlaw headnotes to train its AI legal-research tool failed the transformation test. By repackaging the headnotes in a manner that directly replicated Westlaw’s legal-research service, the court found that the AI tool operated in the same market as Westlaw, thereby undermining any claim to transformative use.
This perspective resonates with the earlier 2nd U.S. Circuit Court of Appeals precedent set in American Geophysical Union v. Texaco, where the court rejected the argument that photocopying entire journal articles for research purposes constituted a transformative use. The Texaco decision elucidated that mere duplication—even if it facilitates broader research or innovation—does not suffice to transform the original work if the material is consumed in its original, unaltered form. Indeed, that case provides what may be an appropriate analogy to AI training, in which the systems are trained on the expressive content of protected works for that expressive content directly. That is, AIs “learn” from the copies they make in much the same way that humans “learn” about expressive works by directly perceiving them.
As I will discuss below, the Ross case pushes a similar line of reasoning with a different set of cases. Nonetheless, even with this potentially adverse precedent, there is a potential distinction to be made between Ross and its purpose-specific legal research tool, and general-purpose large language models (LLMs).
Transformative Use in Thomson Reuters v Ross
In Thomson Reuters v. Ross Intelligence, the court examined the methodology by which Ross Intelligence employed Westlaw headnotes to train its AI legal-research tool. The headnotes, which distill legal principles from judicial opinions, were repurposed by Ross to create an AI system that offered legal-research services. In doing so, the court found that Ross’s tool operated squarely within the same market as Westlaw, with its primary function being to facilitate legal research. This determination proved central to the court’s overall fair-use analysis.
Under the framework established in Andy Warhol Foundation v. Goldsmith, the court emphasized that, if a new use shares the same or highly similar purposes as the original work, it strongly suggests a lack of transformation. In the case at-bar, Ross’s reliance on Westlaw headnotes to develop a competing legal-research tool meant that the underlying purpose of the original expression remained unaltered. Rather than adding new meaning, commentary, or purpose, Ross’ use simply repackaged the headnotes in a format that was virtually identical to their original function—namely, assisting users in locating relevant legal cases. As a result, Judge Stephanos Bibas of 3rd U.S. Circuit Court of Appeals—sitting by designation at the district court—found that factor one (which considers the purpose and character of the use) weighed heavily against a finding of fair use.
Ross attempted to draw an analogy to the “intermediate copying” cases, such as those involving Sega and Sony v. Connectix, where courts permitted limited copying of functional code for the purposes of reverse engineering. The analogy that Ross was attempting to draw was to compare training an AI system on content and the conversion of protected works into numerical weights that represent the relationship of words, and the copying done by software developers trying to reverse engineer an application programming interface (API) in order to create compatible new works.
Judge Bibas, however, distinguished these precedents on the basis that they involved functional computer code—primarily comprising unprotected ideas—and were necessary for ensuring compatibility between systems. By contrast, the Westlaw headnotes constitute creative expression, and the copying at-issue was not essential to uncover any unprotected elements. This distinction further undercut the argument that Ross’ use could be deemed transformative.
The commercial nature of Ross’ use also contributed to the court’s adverse view. By seeking to profit from a product that competes directly with Westlaw, Ross’ use maintained the same “research” or “reference” purpose inherent in Westlaw’s headnotes. This commercial aim, coupled with the absence of any significant recontextualization of the original material, reinforced the conclusion that Ross’ copying failed to achieve the requisite transformation.
Ultimately, because Ross’s appropriation of the headnotes did not introduce any new meaning or commentary beyond their original function, the court held that the copying was insufficiently transformative to qualify for fair use. The convergence of an unaltered purpose and the direct competition with Westlaw’s established market for headnotes served as a pivotal basis for rejecting the fair-use defense in this case.
Contrasting Texaco with Ross
American Geophysical Union v. Texaco serves as a seminal illustration of why copying for research purposes does not automatically confer a transformative use. In that case, Texaco scientists photocopied entire journal articles, not to reinterpret or add new meaning to the works, but simply to enable company personnel to read and study the material in the same manner as originally published. The court held that this type of copying was not transformative, because it duplicated the works for their conventional reading purpose rather than reconfiguring or recontextualizing their expressive content.
This holding underscores a critical takeaway: the existence of a broader public benefit—whether it be research, training, or scientific advancement—does not, in itself, satisfy the requirement of transformation under copyright law. When the copied material is employed in a manner that “consumes” the original expression for the same fundamental function, the use can fail the first factor of the fair-use analysis. In Texaco, the mere replication of articles for the purpose of reading did nothing to alter or enhance the original creative expression, thereby falling short of the transformative threshold.
It could be argued that when an AI system “reads” copyrighted works—ingesting and processing the material solely for its expressive content—it performs an act analogous to the photocopying in Texaco. Without a novel or recontextualized purpose, such copying risks being deemed nontransformative, even if it ultimately benefits a broader public. Thus, just as Texaco’s use did not transform the original works, AI training that merely replicates the work’s inherent expressive value may similarly fail to meet the transformation requirement under fair use.
Both Texaco and Ross embody the principle that downstream benefits—such as advancing research or fostering AI innovation—do not in themselves confer a transformative use. In both cases, the courts rejected the notion that a broader public benefit can justify the reproduction of copyrighted works when the copying merely replicates the original work’s primary function.
In Texaco, the photocopying of entire journal articles supplanted the need to purchase additional subscriptions, effectively substituting for the original publications’ market function. Similarly, in Ross, using Westlaw headnotes to train an AI legal-research tool directly supplanted the traditional role of Westlaw—duplicating its core service of facilitating legal research.
In each case, the copying did not transform the original work. Instead, it duplicated the work’s market function, thereby reinforcing the argument that downstream benefits alone cannot satisfy the transformation requirement under the fair-use doctrine.
Where General LLMs Diverge: Future vs Original Markets
With all that said, it could very well be the case that general purpose LLMs can prevail in a fair-use claim—at least under the reasoning of Ross, with its stress on the downstream market effects. General LLMs differ markedly from the AI system at issue in Ross in that they do not necessarily target the same consumption market as the original works. In fact, Judge Bibas’ opinion made a point of distinguishing between “generative AI” and the AI that Ross was using:
Ross was using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with Westlaw. It is undisputed that Ross’s AI is not generative AI (AI that writes new content itself). Rather, when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written.
Whereas Ross’ legal-research tool was designed to serve an identical market—to provide legal professionals with access to the equivalent of Westlaw headnotes—general LLMs typically do not aim to replace the experience of, e.g., reading a novel for enjoyment or viewing original artworks. Instead, these models are frequently employed to generate new outputs, such as composing fresh text or producing novel images, thereby displacing future creative services, rather than substituting for the direct consumption of existing works.
This distinction underpins a more favorable transformative-use argument for general LLMs. When an AI’s purpose is to create outputs that differ substantially from the original function—such as generating a short story inspired by, but not a replication of, an existing novel—the use may be viewed as transformative. Such a process, which reinterprets the source material to produce an entirely new work, diverges from merely consuming the original copyrighted expression for its intended purpose.
In this respect, the transformation effected by a general LLM stands in contrast to the relatively more straightforward replication observed in Ross, suggesting that the use of copyrighted works in training may be more likely to qualify as fair use when it results in outputs that serve a distinctly different market function.
Nonetheless, the potential benefits for AI producers are tempered by the risk that, even if general LLMs do not supplant the original market for a given work, they may nevertheless encroach upon derivative or future licensing markets. The use of copyrighted material in training an LLM might diminish the market value of future licensing opportunities that rightsholders could otherwise develop.
Under the Ross framework, such market effects could tip the balance slightly in favor of the AI producer in a fair-use analysis. But given the multifactorial nature of the fair-use inquiry, this advantage is far from determinative, and must be weighed carefully against the other factors that protect the economic interests of the original creators.
The Ongoing Balancing Act
Fair use remains an inherently multifactorial inquiry, requiring a careful and nuanced evaluation of the purpose and character of the use; the nature of the copyrighted work; the amount and substantiality of the portion taken; and the effect on the potential market for the original. Each of these factors must be weighed together, and no single factor is determinative.
Even when a general-purpose LLM operates in a market distinct from that of the original work, courts may still conclude that the use is not transformative. This could happen, for example, if the copying in question fails to alter the underlying expressive content or if it imposes significant market harm. Indeed, despite the possibility that an LLM may not directly supplant the original work’s market, the risk remains that its deployment could adversely affect derivative or future licensing markets.
The critical fourth factor—one Judge Bibas found particularly important—is market harm. Future judgments that hinge on this factor may pivot significantly based on whether the original creators are able to secure licensing agreements for the use of their works in AI training. Consequently, even if the LLM’s outputs serve a different function from the original works, the potential for market substitution or harm cannot be dismissed.
The precedents established in Ross and Texaco demonstrate that copying for the same function may not easily be seen as transformative under the fair-use doctrine. While general LLMs might theoretically have a more favorable argument—given that they often generate novel outputs, rather than simply reproducing the original expression—the overall outcome remains uncertain.
The balancing test inherent in fair-use analysis continues to evolve, and future judicial rulings or legislative interventions may ultimately shift the contours of what is deemed transformative in the context of AI training.
