New Ruling on AI and Copyright Infringement Provides Insight into Courts’ View of Using Protected Works to Train an AI System and the Applicability of the Fair Use Defense to Non-Generative AI.
Thomson Reuters enter. Ctr. GMBH v. Ross Intel. Inc., No. 1:20-cv-613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025)
The United States District Court for the District of Delaware has recently revised a previous order and granted summary judgment in favor of the plaintiff, Thomson Reuters (“Reuters”), while rejecting all defenses raised by defendant Ross Intelligence (“Ross”).
Ross sought to create a legal-research search engine utilizing artificial intelligence (“AI”) and to train the AI through “good” and “bad” answers to legal questions. Ross initially sought a license to use the content from Reuters’s Westlaw product, but Reuters declined. Ross then made a deal with LegalEase to purchase 25,000 “Bulk Memos” that LegalEase created by guiding lawyers to create the “good” and “bad” answers. LegalEase trained those lawyers to do so using Westlaw headnotes. Simply put, “Ross built its competing product using Bulk Memos, which were built from Westlaw headnotes. When Thomson Reuters found out, it sued Ross for copyright infringement.”
I. Direct Copyright Infringement
A claim for direct copyright infringement requires showing: (1) a valid copyright and (2) actual, substantially similar copying of a protectable element of the copyrighted work. Although Judge Bibas had initially denied summary judgment because he viewed the question of originality—an aspect of copyright validity—as a fact question for the jury, he had a “change of heart” and determined Westlaw’s headnotes and Key Number System were original, valid works.
The court found evidence of actual copying by comparing the similarity of the Bulk Memo questions to Westlaw’s headnotes and the noncopyrightable judicial opinions to which the headnotes pertained. Certain Bulk Memo questions appeared more like headnotes than the underlying judicial opinion. Judge Bibas concluded Ross’s access to the headnotes in conjunction with the probative similarity between the works added up to evidence of actual copying.
Taking into consideration that the less protectable expression a work contains, the more similar the allegedly infringing work must be to it, the court found substantial similarities between the Bulk Memos and 2,243 headnotes. He granted summary judgment on direct copyright infringement on those headnotes, because the language in the Bulk Memos very closely tracked the language of the headnotes, but not the language of the judicial opinion. The rest of the allegedly infringed headnotes survived summary judgment, meaning the trier of fact will determine whether evidence supports an infringement claim.
II. Defenses to Copyright Infringement
Ross raised several defenses, the majority of which Judge Bibas quickly dismissed as inapplicable (innocent infringement, copyright misuse, and the merger defense). The Court most closely analyzed Ross’s assertion of the fair-use defense, which requires consideration of the following factors: (1) the use’s purpose and character, including whether it is commercial or nonprofit (2) the copyrighted work’s nature; (3) how much of the work was used and how substantial a part was relative to the copyrighted work’s whole; and (4) how Ross’s use affected the copyrighted work’s value or potential market.
The Court found the first factor favored Reuters, because Ross’s use was commercial and not “transformative” (i.e. did not have a further purpose or different character from the original). Ross argued that their copying occurred at an intermediate stage of the process, not in the final product. That is, the headnotes were converted into numerical data, which was then fed into its AI system. This sort of “intermediate” copying has been held permissible under the first fair-use factor in the context of computer programming. Judge Bibas reasoned that computer programming is inapt to copying headnotes because computer programming is more functional than other copyrighted works and is only permissible when the intermediate copying is necessary for competitors to innovate. Because Ross did not need to infringe to innovate, and because the broad purpose and character of Ross’s use was to develop a competing legal research tool, the use was not transformative. Judge Bibas explicitly stated this analysis only applied to non-generative AI.
The second and third factors favored Ross, because while the headnotes satisfied the originality requirement, the material “is not that creative,” and Ross’s output did not make headnotes available to the public.
The final factor is the most important to the analysis, and the court determined it favored Reuters, because Ross meant to create a market substitute to compete with Reuters. Further, Judge Bibas explained that “it does not matter whether Thompson Reuters has used the data to train its own legal search tools; the effect on a potential market for AI training data is enough.” Because of the importance of the fourth factor, the court granted summary judgment to Reuters on the fair-use defense.
In sum, the court granted partial summary judgment on 2,243 headnotes, subject to the fact finder determining whether all included headnotes have currently valid copyright registrations. The court also granted summary judgment to Reuters against all of Ross’s defenses. This case clarifies that materials used to train non-generative AI may be subject to copyright infringement claims, especially when the material is used to create a competing market substitute.
