On February 11, 2025, a U.S. District Court in Delaware issued a decision of first impression finding that a technology startup’s use of copyrighted materials to train its artificial intelligence platform infringed the copyright holder’s rights and did not constitute fair use.[1]  The Delaware District Court is the first to decide on a defendant’s fair use defense to an infringement claim for unauthorized use of copyrighted works to train an artificial intelligence (“AI”) tool.  As previously reported on this blog, U.S. District Courts in New York and California are also facing, and have yet to decide, this hotly contested issue.[2]

Thomson Reuters v. Ross Intelligence

Nearly five years ago, on May 6, 2020, Thomson Reuters Enterprise Centre GmbH (“Thomson Reuters”), owner of the prominent legal research platform Westlaw, sued technology startup Ross Intelligence Inc. (“Ross”) for copyright infringement based on Ross’s use of Thomson Reuters’s copyrighted materials to train AI powering Ross’s competing legal research platform.[3]  On September 25, 2023, Judge Bibas of the District Court for the District of Delaware denied Thomson Reuters’s motion for summary judgment on its copyright claims, finding that questions of fact required a jury trial.[4] The parties continued to litigate the case and prepare for trial.  On October 1, 2024, Thomson Reuters renewed its motion for summary judgment.[5]  On February 11, 2025, Judge Bibas issued an order granting partial summary judgment to Thomson Reuters, finding Ross’s use of Thomson Reuters’s copyrighted materials to train Ross’s AI platform infringed Thomson Reuters’s rights and did not constitute fair use.  

Ross’s Use of Thomson Reuters’s Copyrighted Materials to Train its AI

Thomson Reuters’s  legal research platform Westlaw contains legal sources, such as cases, statutes, law journals and treatises, as well as Thomson Reuters’s own editorial content and annotations, including headnotes that summarize key points of law and case holdings.[6]  Ross, a startup competitor of Thomson Reuters, developed “a legal-research search engine that uses artificial intelligence.”[7]  In developing that search engine, Ross worked with LegalEase, a former Westlaw client, to develop approximately 25,000 “Bulk Memos” to train Ross’s AI search tool.[8]  In doing so, Ross and LegalEase altered the form of Westlaw’s headnotes, by turning statements into questions, rather than directly copying them.[9]  The headnotes at issue are sealed, but Judge Bibas provided the following example to demonstrate how minimally Ross changed Westlaw’s headnotes for use in its AI training memos[10]:

Thomson Reuters’s Copyright Claim

Thomson Reuters claimed that Ross directly infringed its copyrights, (i) in Westlaw’s legal headnotes specifically, and (ii) in Westlaw’s Key Number System as a whole.  The Delaware District Court initially found, in its September 2023 opinion denying summary judgment to Thomson Reuters, that the question of whether Thomson Reuters’s legal headnotes and Key Number System were “original” enough to be copyrightable must be tried to a jury.  However, in its subsequent February 2025 opinion granting summary judgment to Thomson Reuters, the Delaware District Court reexamined the originality standard, finding that “the originality threshold is ‘extremely low,’ requiring only ‘some minimal degree of creativity . . . some creative spark.’”[11]  The Delaware District Court determined that each Westlaw headnote could be “an individual, copyrightable work,” comparing the editorial judgment required—even if taken verbatim from an opinion—to the work of a sculptor, because “[i]dentifying which words matter and chiseling away the surrounding mass expresses the editor’s idea about what the important point of law from the opinion is.”[12]  The court also found Westlaw’s Key Number System sufficiently original to be copyrightable.[13]

Judge Bibas’s decision only applies to a portion of the headnotes Ross allegedly infringed, so there may still be a trial on the rest of the headnotes (which require a jury to determine the validity of the copyright and whether there was copying), whether Ross copied the Key Number System and whether Ross accessed and used certain judicial opinions containing Thomson Reuters’s editorial decisions.

Fair Use

The Delaware District Court rejected Ross’s defense that it did not violate the Copyright Act because its use of Thomson Reuters’s copyrighted materials to train its AI legal research system was “fair use.”  In doing so, Judge Bibas conducted the four-part fair use analysis contained in the Copyright Act.

Factor One:  The first fair use factor examines the purpose and character of the work, and whether the allegedly infringing use is “transformative.”  The court found in Thomson Reuters’s favor on this factor, determining thatRoss’s use of Thomson Reuters’s materials (i) was commercial, which weighs against fair use,[14] and (ii) was not transformative, because “it does not have a ‘further purpose or different character,’” and created “a tool to compete with Westlaw.”[15]  Ross argued that so-called “intermediate copying”—permitted in computer coding cases like Google LLC v. Oracle Am., Inc., 593 U.S. 1 (2021), in which the Supreme Court held that it was “fair use” for Google to incorporate copyrighted code where it was necessary in developing a new product—allowed it to “turn[] the headnotes into numerical data about the relationships among legal words to feed into its AI.”[16] The Court rejected Ross’s argument, finding that its copying was not “reasonably necessary to achieve [its] new purpose.”[17] 

Factor Two:  The second fair use factor examines the originality of the copyrighted work.  The court found that—while Westlaw’s work “has more than the minimal spark of originality for copyright validity”—it was “not that creative,” so it decided this factor in favor of Ross.[18]

Factor Three:  The third fair use factor relates to the proportionate use of the copyrighted work.  The court found in favor of Ross on this factor, as “Ross did not make West headnotes available to the public.”[19]

Factor Four:  The fourth fair use factor—the “single most important element of fair use”—is the likely effect of the copying on the market, including “potential derivative ones.”[20]  The court found it beyond dispute that Ross “meant to compete with Westlaw by developing a market substitute.”[21]  Further, the court found no public right to Thompson Reuters’s analysis, despite the fact that legal opinions themselves are available,[22] and a value in developing effective legal-research tools, warranting copyright protection.[23]  Thus, the court found in Thomson Reuters’s favor on this factor.

The Implications of the Delaware District Court’s Fair Use Decision

Use of copyrighted materials to train AI tools is a hot topic, with potentially far-reaching implications.  As reported previously on this blog, in the first quarter of 2025, the Copyright Office intends to issue a report addressing “the legal issues related to the ingestion of copyrighted works to train AI, including licensing considerations and the allocation of potential liability.”    Meanwhile, as noted above, the issue is the subject of litigation in high-stakes disputes in U.S. District Courts in New York and California.  For example, as reported previously on this blog, in May 2024, a pair of voice actors brought copyright infringement claims based on the unauthorized use of their voices to train generative AI in a suit pending before Judge J. Paul Oetken in the District Court for the Southern District of New York.[24]  The list of cases continues to grow—on February 13, 2025, Advance Magazine Publishers Inc. (popularly known as Condé Nast) and a group of other publishers filed suit in the Southern District of New York against Cohere Inc. for its dissemination of the publishers’ copyrighted works and its use of those works to train its large-language model AI systems.[25]   

In Thomson Reuters v. Ross, the District Court in Delaware cautioned that “Ross’s AI is not generative AI (AI that writes new content itself)” and Judge Bibas “note[d] for readers that only non-generative AI [was] before [him].”[26]  However, the significance of that warning is unclear.  In Concord Music Group v. Anthropic PBC, pending in the District Court for the Northern District of California, which involves certain music publishers’ claims that generative AI tool Anthropic violated their copyrights in materials used to train it, the music publishers immediately sought leave of Court to submit Judge Bibas’s Thomson Reuters v. Ross decision as supplemental authority to support their pending motion for a preliminary injunction.[27]   Anthropic objected, citing Judge Bibas’s language cautioning that he did not consider generative AI in issuing that decision.[28]

* * * *

We will continue to monitor for further developments the cutting-edge territory of AI and its intersection with existing copyright and other legal frameworks at the state and federal level.

[1] Memorandum Opinion, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc., 1:20-cv-00613-UNA (D. Del. Feb. 11, 2025) (ECF No. 770) (the “Opinion”).

[2] See, e.g., The Intercept Media, Inc. v. OpenAI, Inc., No. 1:24-cv-01515-JSR (S.D.N.Y.) (where the court recently permitted discovery to move forward), and Concord Music Grp., Inc. v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal.).

[3] Complaint, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc., 1:20-cv-00613-UNA (D. Del. May 6, 2020) (ECF No. 1).  Plaintiffs also brought a tortious interference with contract claim, alleging that Ross worked with LegalEase, a former Thomson Reuters customer.  Id. at 15.

[4] See Thomson Reuters Enterprise Ctr. CmbH v. Ross Intell. Inc., 694 F. Supp. 3d 467, 478-87 (D. Del. 2023).

[5] Renewed Motion for Partial Summary Judgment, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc., 1:20-cv-00613-UNA (D. Del. Oct. 1, 2020) (ECF No. 674).

[6] Opinion at 2.

[7] Opinion at 3.

[8] Opinion at 3.

[9] Opinion at 3.

[10] Opinion at 3.

[11] Opinion at 6 (quoting Feist Publications, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 345 (1991)).

[12] Opinion at 7.

[13] Opinion at 8.

[14] Opinion at 16-17.

[15] Opinion at 17.

[16] Opinion at 17-18.

[17] Opinion at 19 (quoting Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 532 (2023)).

[18] Opinion at 20.

[19] Opinion at 21.

[20] Opinion at 21-22.

[21] Opinion at 22.

[22] Opinion at 22-23.

[23] Opinion at 22-23.

[24] Lehrman et al v. Lovo, Inc., No 1:24-cv-03770 (S.D.N.Y. filed May 15, 2024); see also, e.g., The Intercept Media, Inc. v. OpenAI, Inc., No. 1:24-cv-01515-JSR (S.D.N.Y.) (where the Court recently permitted discovery to move forward), The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y.), and Concord Music Grp., Inc. v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal.).

[25] Complaint, Advance Local Media LLC v. Cohere Inc., No. 1:25-cv-01305 (S.D.N.Y. Feb. 13, 2025) (ECF No. 1).

[26] Opinion at 17, 19.

[27] Motion for Leave to File Notice of Supplemental Authority, Concord Music Group v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal. Feb. 11, 2025) (ECF No.295).

[28] Opposition to Motion for Leave to File Notice of Supplemental Authority, Concord Music Group v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal. Feb. 11, 2025) (ECF No. 296).