On February 11, 2025, the U.S. District Court for the District of Delaware granted summary judgment in favor of Thomson Reuters in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.[1] holding that use of copyrighted works to train AI models does not fall under the ‘fair use’ exception.
A. Background and Facts of the Case.
Thomson Reuters (“Reuters”) owns Westlaw, one of the biggest legal research platforms. Westlaw not only reproduces copies of judgments and statutes but also provides editorial content including ‘headnotes’ that summarize the key aspects of a case. Westlaw also organizes its content by using a proprietary classification system that organizes content by topic and sub-topic and assigns each case, topic and sub-topic with a unique number. Westlaw’s proprietary numerical taxonomy is called ‘the Key Number System’.
Ross Intelligence (“Ross”) developed and launched a competing legal research platform that uses artificial intelligence (“AI”) to assist in the search process. Ross needed a legal database to train its AI tool and therefore initially requested a license from Reuters to use the content on Westlaw for training purposes. Because Ross was a direct competitor, Reuters refused to grant a license. Ross then reached out to LegalEase to get the necessary training data in the form of ‘Bulk Memos’ that are essentially a compilation of lawyer-prepared legal questions and answers. When engaging lawyers to create the Bulk Memos, LegalEase provided those lawyers with Westlaw headnotes as a guide but had asked the lawyers to not reproduce the same in the Bulk Memos. As a result, many of the Bulk Memos were restatements of Westlaw headnotes or were otherwise created by referencing Westlaw content. Once Ross had the necessary training data, it was able to train its AI tool and come out with a competing legal research platform.
In May 2020, Reuters sued Ross alleging that Ross used its copyrighted work to train Ross’ competing legal research platform.
B. Summary Judgment and Rejection of ‘Fair Use’ Defense.
In 2023, the U.S. District Court for the District of Delaware denied Reuters’ motion of summary judgment on most issues.[2] However, upon a more detailed review of the materials submitted by the parties, Judge Bibas recognized that his 2023 summary judgment ruling did not fully address the issues. As a result, he permitted the parties to renew their summary judgment motions. Reuters moved for summary judgment on copyright infringement and rejection of related defenses, while Ross sought summary judgment asserting non-infringement and a ruling in favor of fair use. Upon reviewing the renewed motions, on February 11, 2025, Judge Bibas, granted Reuters’ motion for partial summary judgment on direct copyright infringement and rejected the related defenses (including ‘fair use’). While the 23 page order delves into a number of copyright issues, however, for the purposes of this Article, we are only focusing on the Court’s rejection of the ‘fair use’ doctrine.
The Court considered the following four factors before rejecting the fair use doctrine and ruling in favor of Reuters:
- Purpose and character of the use, including whether it is commercial or nonprofit: The Court sided with Reuters on this factor since Ross’ use of Reuters’ copyrighted work was for a ‘commercial purpose’. It essentially used Reuters’ copyrighted work to create a competing product in the market. In addition, the Court found that Ross’ use was not transformative. In other words, Ross did not use the copyrighted work for any additional purpose or different character. Ross’ use of Reuters’ copyrighted work was for the same purpose that Reuters used it for which was to improve the legal research database. Additionally, the Court rejected Ross’ argument related to intermediate copying. Ross argued that the headnotes were never reproduced to the end-users and were merely an intermediate step to train its AI tool. Further, Ross asserted that such intermediate copying was protected by fair use. However, the Court rejected this argument for two reasons (a) Ross is copying written words whereas courts have allowed intermediate copying in the case of computer code[3]; and (b) Ross’ copying was not necessary to innovate or create a new purpose. Courts have allowed intermediate copying in computer programming on the grounds that it was necessary for competitors to innovate (for instance, intermediate copying was done to make the two software products compatible). Here, Ross has merely copied the headnotes to create a competing product.
- Nature of the copyrighted work: The second factor focuses on the level of originality and creativity in the copyrighted work. Even though Reuters’ copyrighted work has the required minimum level of creativity needed for copyright protection, it is not that creative. Therefore, the Court ruled in favor of Ross on factor two.
- How much of the work was used: The third factor focuses on the quantity and quality of the copyrighted work used by the alleged infringer. The test is whether the alleged infringer took the ‘heart’ of the work. According to the Court, factor three also favored Ross since Ross’ output to the end user did not include the headnotes. In other words, the headnotes were only used to train the AI tool and were not reproduced to the end user.
- Effect on the copyrighted work’s value or market: The fourth factor focuses on the likely effect of alleged infringer’s copying on the market for the original. The Court found that this factor favored Reuters as Ross’ legal research platform is not serving a different purpose from Westlaw. In fact, it is a market substitute for Westlaw. The Court further found that the alleged copying was significantly hampering Reuters’ market for legal research platforms, and therefore, could not be protected under the fair use doctrine. Finally, the Court noted that Ross could have easily created the materials required to train the AI model either by itself or with LegalEase’s assistance.
Even though both parties had two factors in their favor, the Court relied on prior Supreme Court precedent,[4] which placed the most importance on the fourth factor. Therefore, the Court rejected Ross’ fair use defense argument and ruled in favor of Reuters. In other words, the Court held that, at least in this particular case, fair use is not a viable defense to copyright infringement when the copyrighted work of others is used to train AI without permission.
C. Impact on AI Providers and Copyright Holders
This summary judgment ruling by Judge Bibas is a major victory for copyright holders, reinforcing their control over proprietary content and opening potential revenue streams through licensing agreements with those that are looking to use such proprietary content to train their AI tools. On the other hand, this ruling could represent a significant roadblock for AI development. AI providers must now be mindful of the kind of data sets they are using to train their AI models. As a best practice, AI providers should either rely on non-copyrighted works or obtain proper licenses to avoid any legal liability. This summary judgment ruling underscores the value of original and creative works and signals a more restrictive approach to AI training that could shape the industry standards moving forward.
However, this summary judgment ruling is not a definitive precedent, and certainly not for generative AI. Judge Bibas explicitly stated that Ross’ legal research database is not a generative AI tool, rather it just provides judicial opinions based on users’ search queries. Even though the reasoning in this decision can influence the ongoing litigation in the generative AI space, it is certainly possible for a court to be more receptive to a fair use defense in the case of generative AI tools, where there is potentially a better argument that such use is ‘transformative’ (factor one). Further, the opinion discussed above is a summary judgment ruling and could easily be reversed or modified on appeal depending on how Ross decides to proceed with this case.
Although the ruling sent shockwaves through the AI industry, it offers only a preliminary indication of the legal landscape rather than a definitive roadmap for future cases. Please reach out to a member of Benesch’s Artificial Intelligence Industry Group with any questions.
Daniel S. Marks is a Partner in the firm’s Intellectual Property and Artificial Intelligence Groups. He may be reached at dmarks@beneschlaw.com or 216.363.6101.
Shivdutt Trivedi is an Associate in the firm’s Intellectual Property and Artificial Intelligence Groups. He may be reached at strivedi@beneschlaw.com or 646.593.7053.
[1] 2025 WL 458520 (D. Del. Feb. 11, 2025).
[2] 694 F. Supp. 3d 467 (D. Del. 2023).
[3] Sony Comput. Ent., Inc. v. Connectizx Corp., 203 F.3d 596 (9th Cir. 2000); Sega Enters. Ltd. v. Accolade, Inc., 977 F. 2d 1510 (9th Cir. 1992).
[4] Harper & Row, Publishers, Inc. v. Nation Enters., 471 U.S. 539 (1985).