The application and use of Artificial Intelligence (AI) in modern society continues to expand, and with this growth, copyright owners are increasingly eager to protect their intellectual property from AI developers seeking to train AI models on copyrighted material. In Thomson Reuters Enterprise Centre GmbH & West Publishing Corp. v. Ross Intelligence Inc.,1 Judge Stephanos Bibas of the U.S. Court of Appeals for the Third Circuit, sitting by designation in the U.S. District Court for the District of Delaware, recently issued the first decision rejecting a fair use defense for using copyrighted material to train non-generative AI. While the application of this decision is limited to this context, it serves as a further indicator of how courts may approach copyright infringement claims and fair use defenses in the broader context of AI.
The case involved Thomson Reuters’ Westlaw legal research tool and Ross Intelligence, a startup offering a competing product that leverages AI for legal research. Ross initially sought to license Westlaw’s data to train its AI but was refused by Thomson Reuters. As a result, Ross used attorney-created “Bulk Memos,” which were compilations of legal questions with good and bad answers from Westlaw headnotes, to train its AI. In essence, “Ross built its competing product using Bulk Memos, which in turn were built from Westlaw headnotes.”2 Thomson Reuters filed a lawsuit in 2020, alleging that Ross had infringed its copyright by using thousands of Westlaw headnotes and other materials as training data for Ross’ AI system.
In September 2023, the court rejected both parties’ initial requests for summary judgment, claiming the court lacked sufficient grounds to rule on direct copyright infringement and fair use. In October 2024, however, the court allowed the parties to file renewed motions for summary judgment. Judge Bibas revisited his decision after “studying the case materials more closely and realizing that [his] prior summary-judgment ruling had not gone far enough.”3
The court first addressed Thomson Reuters’ direct copyright infringement claim. Ross argued the Westlaw headnotes failed to meet the originality threshold. The court disagreed, stating, “Identifying which words matter and chiseling away the surrounding mass expresses the editor’s idea about what the important point of the law from the opinion is.”4 Therefore, Thomson Reuters’ “editorial expression ha[d] enough ‘creative spark’ to be original.”5
Next, the court determined whether Ross had copied the original elements found in the Westlaw headnotes. This was done by considering if Thomson Reuters had shown Ross actually copied Westlaw’s headnotes, and whether there was substantial similarity between the Westlaw headnotes and Ross’ corresponding Bulk Memos. The court found Ross copied 2,243 headnotes as the language of the Bulk Memos “look[ed] more like a headnote than…the underlying judicial opinion.” Interestingly, Judge Bibas stated that “[a]s a lawyer and judge, [he was] well positioned” to determine whether there was substantial similarity between the works.6 In finding substantial similarity, the court stated the Bulk Memos’ “language very closely tracks the language of the [headnotes] but not the language of the case opinion[s].”7
Turning to Ross’ fair use defense, the court addressed each of the four factors of that defense: (1) the use’s purpose and character, (2) the copyrighted work’s nature, (3) the amount and substantiality of the use, and (4) the use’s effect on the copyrighted work’s value or potential market.8
The court relied on the U.S. Supreme Court’s Andy Warhol Foundation v. Lynn Goldsmith decision in analyzing the first factor.9 Beyond the established commercial nature of Ross’ use, the court noted that Ross’ AI was “not transformative because it does not have a further purpose or different character” than Thomson Reuters’ Westlaw.10 This element favored Thomson Reuters, as “Ross was using Thomson Reuters’ headnotes as AI data to create a legal research tool to compete with Westlaw.”11
In analyzing the second fair use factor, Judge Bibas found that the Westlaw headnotes had “more than the minimal spark of originality required for copyright validity…[but] the material is not that creative.”12 The lack of inherent creativity in the Westlaw headnotes favored Ross. The court also noted that this factor “has rarely played a significant role in the determination of a fair use dispute.”13
With regard to the third factor, the court evaluated “the amount and substantiality of what is thereby made accessible to a public for which it may serve as a competing substitute.”14 The court determined that the third factor also favored Ross, as the users of Ross’ AI were never exposed to the Westlaw headnotes.
The fourth, and most significant, fair use factor favored Thomson Reuters. The court found that Ross’ use was “meant to compete with Westlaw by developing a market alternative.”15 Ross’ AI also threatened Thomson Reuters’ potential market for selling Westlaw headnotes as data to train other legal AIs.
In the court’s analysis, two factors favored each party. However, since the first and fourth factors – the most influential – favored Thomson Reuters, the court rejected Ross’ fair use defense. While the fair use analysis was the most significant aspect of the ruling, the court also dismissed Ross’ affirmative defenses of innocent infringement, copyright misuse, merger, and scenes à faire.
The lasting effect of this decision may be limited since the district court’s decision is likely to be appealed. Even still, it is confined to the realm of non-generative AI. However, Thomson Reuters is instructive for both copyright owners and those involved in the development of AI.
For copyright owners, this decision provides guidance in enforcing their copyrights where the AI tool in question is non-generative. It may also bolster infringement arguments in generative AI cases where the output is minimally transformative. In addition, the decision provides strategic guidance in how to tackle a fair use defense asserted by an AI developer.
For AI developers, the decision reinforces the need to conduct due diligence in selecting material to train or improve AI tools, including selecting works in the public domain and obtaining appropriate licenses for copyrighted works where available. When purchasing training data from third parties, AI developers should ensure that indemnification is included in the terms of purchase. With respect to the output of AI tools, especially those that are non-generative, developers should consider minimizing the amount of copyright material made publicly available.
This blog post was drafted by John Allen, an intellectual property attorney in the Salt Lake City office of Spencer Fane and Jeff Ratinoff, an intellectual property attorney in the San Jose, California, office. For more information, visit www.spencerfane.com.
—
1 No. 1:20-cv-613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025).
2 Id. at *1
3 Id. at *1.
4 Id. at *3.
5 Id.
6 Id. at *5.
7 Id. at *6.
8 Id. at *7.
9 598 U.S. 508 (2023).
10 2025 WL 458520, at *7.
11 Id.
12 Id. at *9.
13 Id.
14 Id.
15 Id. at *10.
Click here to subscribe to Spencer Fane communications to ensure you receive timely updates like this directly in your inbox.
