Thomson Reuters v ROSS: What It Means for Copyright Law and How AI Developers Should Move Forward

Introduction

Artificial Intelligence (AI) is growing at an incredibly rapid pace, with new developments emerging almost weekly. While the hype for this ground-breaking technology continues to grow, so are the number of legal disputes surrounding its use. On February 11, 2025, the Delaware District Court delivered a partial summary judgement for the case Thomson Reuters v ROSS, and the judicial decisions made within may serve as impactful precedents for future cases to follow. 

Thomson Reuters v ROSS: What Happened?

ROSS, a legal research startup, wanted to utilise machine learning and AI for its search engine. To train its AI, ROSS initially tried to license Reuters’ legal content (with Reuters serving as the owner of the legal research platform Westlaw). When this effort failed, ROSS hired a 3rd party legal research company to create training materials instead. These materials consisted of memos with legal questions and answers. Reuters subsequently filed a copyright infringement complaint, claiming that the questions used by ROSS were essentially their headnotes with question marks at the end. 

The Court ultimately granted partial summary judgement in favour of Reuters, finding that ROSS’s use of Reuters’ materials constituted copyright infringement and rejected its fair use defence. To understand the Court’s reasoning, we must explore the elements of copyright infringement and the legal doctrine of fair use.

Copyright Infringement

Judge Bibas laid out the elements of copyright infringement as follows. 

The plaintiff must prove that they:

(1)   own a valid copyright, and

(2)   the defendant copied protectable elements of the copyrighted work.

To establish (2), the plaintiff must show that:

(a)   there has been an act of actual copying by the defendant and

(b)  the defendant’s work is substantially similar to the plaintiff’s copyrighted work.

The Court ruled that Reuters’ owned a valid copyright and that their headnotes are individual, copyrightable works. Furthermore, to evaluate “copying”, the Court compared the memo questions to the headnotes. In addition to Ross’s expert conceding that the memo questions closely resemble the headnotes, the Court ruled that there was actual copying of Reuters’ headnotes by ROSS, and that there is substantial similarity. Having established that there is direct copyright infringement, the Court went on to examine the potential defenses ROSS may have.

Fair Use

With roots in common law and the goal to promote freedom of expression, fair use serves as a US doctrine that permits limited use of copyrighted content when it’s done for a “transformative” purpose. Since fair use is an affirmative defence, the defendant bears the burden of proof, and 4 factors are usually considered to evaluate whether this defence is established:

(1)   the purpose and character of the use,

(2)   the nature of the copyrighted work,

(3)   the amount and substantiality of the part of copyrighted work that was used, and

(4)   the effect of the use on the value of the copyrighted work or potential market.

After balancing the four factors, considering that (2) is of less significance and (4) is of the most, partial summary judgement for direct copyright infringement was granted in favor of Reuters.

Moving Forward: What Happens Next?

This ruling is significant for AI developers since it outlines the importance of licensing arrangements in training AI models. The court’s decision clarifies that creating an AI model is not considered transformative and fair use, since its ultimate purpose is to compete with the original works’ owners. Therefore, AI developers must now be selective with what data they use to train their AI – given that they may be roped into legal trouble if they utilize copyrighted material without prior permission (which they can obtain through licensing).

However, the current case does not serve as an absolute precedent set for all AI-related cases – especially since Ross’s current platform does not constitute a form of generative AI. As AI continues to evolve, the courts will have to make rulings on increasingly difficult judicial ambiguities. While the ultimate goal is to transparently dictate the limits of AI use, it is equally crucial for the community to debate what we can do versus what we should do with AI. Only in an environment that promotes ethical and responsible innovation can AI truly thrive whilst ultimately not compromising societal values.

References and Further Reading

“Fair Use Index” (U.S. Copyright Office)

“Fair Warning: AI’s First Copyright Fair Use Ruling, Thomson Reuters v. ROSS” (Hunton Andrews Kurth LLP)

“Measuring Fair Use: The Four Factors” (Stanford Libraries)

Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025)

“Thomson Reuters v. Ross Intelligence: A Landmark Case on AI Training and Copyright” (Munck Wilson Mandala)

“What is Fair Use?” (Stanford Libraries)