Early Battle in the AI Wars – Thomson Reuters v Ross Intelligence

On 11 February 2025, Judge Stephanos Bibas, of the US District Court for the District of Delaware, handed down the first major fair use ruling amidst a wave of federal litigation challenging the unauthorised use of copyright material for purposes of developing artificial intelligence (AI) models. In a 180-degree turn from an earlier ruling in the same case, Thomson Reuters v Ross Intelligence, Inc, Judge Bibas rejected the defendant’s fair use arguments on summary judgment, handing plaintiff Thomson Reuters a victory in a closely watched case (that victory may prove fleeting as the Third Circuit granted an interlocutory appeal from defendant Ross Intelligence).

Thomson Reuters: Only One of Many Challenges to the Creation and Deployment of AI Models

Since the Thomson Reuters case started in 2020, plaintiffs have filed many high-profile lawsuits targeting the companies behind major AI large language models (LLMs). These issues have become a cross-industry concern, with plaintiffs emerging from the literary, music, journalism and publishing worlds. Plaintiffs include celebrities such as Sarah Silverman, Junot Diaz, Ta-Nehisi Coates and George RR Martin, as well as major media companies such as the New York Times and music industry powerhouses such as Universal Music Publishing Group and Concord Music Publishing Group.

Across the various lawsuits, plaintiffs argue (among many other arguments) that the defendant companies reproduced plaintiffs’ copyrighted works, without compensation, to train AI models. In turn, the defendants have asserted that their alleged conduct is protected by copyright law’s fair use doctrine.

For all sides, and perhaps for the world at large, the stakes could not be higher. For the plaintiffs, these suits could fundamentally alter the scope of protection for their copyrights and the degree of compensation for reproductions and derivative use of their works. An adverse ruling could completely reshape the compensation models for use of copyrighted material. For the defendants, these cases potentially strike at the very core of their business: the ability to develop and train AI models on vast datasets, often comprised of copyrighted material. For copyright practitioners, these cases will test and possibly even redefine the boundaries of the fair use doctrine. Finally, if these lawsuits prevail, they could reshape the trajectory of what may be the defining technological breakthrough of our era – AI – whose models are already deeply woven into the decision-making fabric of business, government, healthcare, education and daily life.

The Legal Landscape

Copyright infringement arises when a defendant copies original, protectable expression without authorisation. However, the “fair use” of a copyrighted work is not an infringement of a copyright. The fair use doctrine, under Section 107 of the Copyright Act, sets out six non-exclusive categories of favoured use:

• criticism;
• comment;
• news reporting;
• teaching;
• scholarship; and
• research.

Four factors are also listed to consider when assessing whether a particular use constitutes fair use, including:

• the purpose and character of the use – ie, whether the use has a different purpose or conveys a new meaning or message compared with the original, or whether it is merely a substitute for the original (a test that has been widely reduced to the shorthand of whether it is “transformative”);
• the nature of the copyrighted work;
• the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
• the effect of the use on the market for or value of the copyrighted work.

Many of the present copyright cases will ultimately turn on whether the defendant companies’ practices constitute transformative use under the first factor and the extent of market harm cognisable under the fourth factor. In Campbell v Acuff-Rose Music, the court defined the vital first factor inquiry as “whether the new work merely ‘supersede[s] the object’ of the original creation, or instead adds something new, with a further purpose or different character, altering the first with new expression, meaning or message.”

Most of the AI-focused cases have yet to make it out of discovery. A few, however, have reached the summary judgment stage and await a decision. See Kadrey v Meta, No 23-cv-03417-VC (TSH) (ND Cal, filed 7 July 2023).  Until decisions are arrived at in these cases, Judge Bibas’s two decisions in Thomson Reuters stand as two of the most substantive fair use assessments to date in this area.

Thomson Reuters and Judge Bibas’s Initial Summary Judgment Decision

In Thomson Reuters, plaintiff Thomson Reuters – the company behind the Westlaw legal research database – argued that the defendant illegally trained a competing, AI legal search tool on source material pulled from Westlaw’s proprietary headnotes. Defendant Ross Intelligence – a legal software and AI start-up that shut down active operations in January 2021 – argued that their use of the headnotes entailed fair use.

In the initial 2023 ruling, Judge Bibas largely denied Thomson Reuters’s motion for summary judgment on the issues of copyright infringement and fair use. Under the first fair use factor, Ross attempted to argue that any alleged copying occurred at an intermediate stage and was thus permissible under existing case law. Specifically, Ross noted that Westlaw’s headnotes do not appear in Ross’s final, consumer-facing product. Rather, Ross turned the headnotes into data “about the relationships among legal words” and then fed that information into their AI. Thomson Reuters argued that Ross plainly copied its attorney-editors’ legal synthesis to use in Ross’s own, competing legal-research platform.

The court ultimately left this first factor to the jury to determine whether Ross’s AI only used the headnotes to learn how to produce judicial opinions, or whether – as Thomson Reuters argued – Ross used “untransformed text” from the headnotes to “replicate and reproduce the creative drafting done by Westlaw’s attorney-editors”. Judge Bibas viewed this determination as a material question of fact for the jury to decide. The opinion also left the critical determination of the fourth factor market impact assessment largely to the jury, given factual disputes about whether Ross’s product was a market substitute for Westlaw and whether Thomson Reuters intended to use its headnotes as training data for its own potential AI tools.

Judge Bibas’s Reversal of the Ruling

Judge Bibas abruptly reversed his own ruling in a February 2025 updated summary judgment decision. Judge Bibas’s revised 2025 decision ultimately granted summary judgment for Thomson Reuters based on the court’s revised fair use analysis. This ruling emphasised the Supreme Court’s holding in Warhol that “[i]f an original work and a secondary use share the same or highly similar purposes, and the secondary use is of a commercial nature, the first factor is likely to weigh against fair use, absent some other justification for copying”. The 2025 opinion focused more on both the purpose and character of Ross’s use and its commercial nature. The 2025 ruling also emphasised the market harm factor as “undoubtedly the single most important element of fair use”.

In the 2025 first factor analysis, Judge Bibas distinguished copying that was “necessary” for competitors to innovate, and, quoting Warhol, copying not “reasonably necessary to achieve the user’s new purpose”. Judge Bibas found that Ross’s intermediate copying did not depend on a need to copy to reach “underlying ideas” and thus was not reasonably necessary.

Judge Bibas also reasoned that if Ross used copyrighted materials for similar, commercial purposes such use would not qualify as fair use. The court ultimately found that Ross’s use was not transformative “because it does not have a ‘further purpose or different character’” from the plaintiff’s uses. Instead, Ross used the plaintiff’s “headnotes” to create a legal research tool that directly competed with Westlaw. Such use was commercial as the company stood to profit from the exploited copyrighted works without paying the “customary price”.

The court then held that the second and third factors weighed in favour of the defendant. In addressing factor two – the nature of the original work – the judge determined that Thomson Reuters’s material was ultimately not especially creative. While there was no factual dispute that the headnotes contained creative elements, the headnotes themselves were far from the kinds of creative works that historically garner greater protection.

Judge Bibas maintained his 2023 position that the second factor went to the defendant Ross, but included a citation noting that this factor “rarely played a significant role in the determination of a fair use dispute”.

On factor three – the amount of the work used and how substantial a part it was relative to the whole – Judge Bibas’s 2023 opinion left the ruling open-ended but leaning in Ross’s favour. The 2025 opinion more decisively handed this factor to Ross, largely because Ross’s output to an end user did not include an actual Westlaw headnote. In other words, because Ross did not make the plaintiff’s headnotes available to the public, Ross did not ultimately reproduce the critical, underlying copyrighted material. In what may prove an important caveat for analogous cases, the Judge reiterated that, while the alleged copier must not take the “heart” of the work, ultimately “[t]he percentage of a total work copied is neither necessary nor sufficient to decide factor three”.

Ultimately, the fourth factor – the likely effect of the copying on the market for the original – swayed the case in the plaintiff’s favour in Judge Bibas’s view as the defendant “meant to compete with [plaintiff] by developing a market substitute”. The 2025 ruling emphasised that “it does not matter whether [plaintiff] has used the data to train its own legal search tools; the effect on a potential market for AI training data is enough”. On this point, the court found that Ross’s use harmed both current and potential derivative markets, particularly considering the burgeoning markets for AI training data. The court found that Ross meant to compete with Westlaw “by developing a market substitute”. It did not matter that Westlaw had not begun training its own legal search tools; rather, it was enough that Ross’s use might have an adverse effect on a potential new market.

In briefly commenting on the issue of the “‘public benefit [...] the copying will likely produce”, the Judge found that, while there is a public interest in accessing the law, there is no public right to Westlaw’s “parsing of the law”. The court determined that there was nothing Thomson Reuters created that Ross could not have created for itself or hired others to create without infringing copyrights. Weighing all factors, the court granted summary judgment for Thomson Reuters on fair use.

Charting the Decision’s Likely Impact

As several of the high-profile generative AI cases move towards summary judgment stages, plaintiffs and defendants will surely parse Judge Bibas’s 2025 opinion for persuasive elements to support their respective positions. In finding for Thomson Reuters, Judge Bibas signalled that the fair use defence may ultimately not turn on sweeping pronouncements regarding the scope of the fair use doctrine but may turn on highly individualised facts (including assessments of potential market harm) in each and every case.

Still, defendants may be quick to distinguish Thomson Reuters by noting that the decision does not concern generative AI. In his 2025 opinion, Judge Bibas clarified that “only non-generative AI is before me today”. The judge defined “generative AI” as “AI that writes new content itself”, distinct from an AI tool such as the one Ross built which “spits back relevant judicial opinions that have already been written”. In this context, defendants could argue that their systems produce an entirely different, sufficiently transformative product, unlike the kinds of tools at issue in Thomson Reuters. Defendants will argue that their LLMs do not simply copy and regurgitate material but rather transform material into a completely new and interactive end result. LLMs use the underlying works to create an entirely transformed product that bears little resemblance to any of the original training-set data.

In addressing the second factor, plaintiffs in other cases may seek to distinguish Ross by pointing out that courts have historically given greater protection to creative works. While the material in Thomson Reuters entailed less creative legal writing, the works at issue in the generative AI cases include novels, stories and poetry by some of the most renowned fiction writers. Defendants, in turn, may respond that Judge Bibas correctly observed that factor two has “rarely played a significant role in the determination of a fair use dispute”.

On factor three, plaintiffs in the other cases may claim that the copying at issue in the generative AI cases covers not just the “heart” of their works but the entire work full stop. Defendants in turn will likely respond that the factor three analysis properly focuses not necessarily on the quantity of the copying but on whether the amount taken was reasonably necessary to achieve the intended transformative purpose. From this, they will be able to argue with some force that copying the entirety of the works was necessary for purposes of improving the reliability and function of the resulting models.

Ultimately, the generative AI cases may turn on how courts evaluate the fourth fair use factor and the weight they assign to it. Judge Bibas’s analysis of this factor in Ross, if upheld on appeal, is likely to become a touchstone for plaintiffs in other AI-related copyright disputes. In assessing whether Ross’s use of Westlaw headnotes harmed Thomson Reuters’s market, Judge Bibas embraced not only a traditional market substitution theory but also came within a hair of adopting a market dilution theory of harm that was later echoed in the Copyright Office’s May 2025 report on copyright and generative AI training. Specifically, Judge Bibas arguably extends traditional fourth factor market harm analysis by finding cognisable market harm arising from Ross’s intention to compete with Westlaw by developing a “market substitute”. Judge Bibas emphasised that permitting AI developers to freely copy headnotes to train competing tools could disincentivise licensing and impair Westlaw’s ability to monetise its content. If this reasoning is adopted at the appellate level, it could significantly expand the scope of cognisable market harm under factor four, encompassing not just lost sales but broader threats to emerging licensing ecosystems. As courts confront the novel challenges posed by generative AI, the recognition of market dilution as a form of harm could become a pivotal element in the fair use equation.

Subsequent plaintiff-friendly decisions may push defendants back into settlement and licensing conversations. Rulings that lean in favour of the defendants could shape entire infrastructures of copyright protection and compensation. The wheels are now well in motion for what may be the most consequential decisions in contemporary copyright law. Stay tuned.