First Decision Finding Copyright Infringement in Generative AI

The first major decision on copyright liability in AI, Thomson Reuters v. Ross Intelligence, provides critical insight into how copyright laws will be applied in the context of generative AI training. On February 11 2025, Judge Bibas of the Third Circuit for the District of Delaware ruled that Ross had infringed Thomson Reuters’s copyright in Westlaw’s legal headnotes that served to train Ross’ AI legal research engine.

This bulletin delves into these key findings and offers a comparative analysis from a Canadian perspective. Those operating in the AI space should take note of this decision and its impact on using copyrighted materials to train AI systems.

Note that on March 18, 2025, Ross filed a motion for certification for interlocutory appeal (see here).

Summary of the Case

Thomson Reuters, who owns the legal research platform Westlaw, accused Ross, an AI-driven legal research company, of using its headnotes to train its AI search tool without permission. The headnotes, which summarize key points of law from judicial opinions, are integral to Westlaw's legal research system. Ross had sought to license these headnotes but was refused by Thomson Reuters, leading Ross to acquire similar data through a third party, LegalEase. LegalEase had created legal memos based on Westlaw’s headnotes and sold those memos to Ross who used them to train its AI.

In a reversal of its previous decision, the Court granted partial summary judgment to Thomson Reuters on copyright infringement and rejected Ross’ fair use defense, finding that Ross had infringed Thomson Reuters’ copyright in 2,243 headnotes (of a total 2,830 headnotes). The Court found that the LegalEase memos were substantially similar to 2,243 of the Westlaw headnotes, and since Ross had copied the memos as is, Ross was found to have infringed Westlaw’s copyright in those headnotes. The remaining headnotes will go to trial. Summary judgment was granted because, as the Court stated, “actual copying [was] so obvious that no reasonable jury could find otherwise.”

Ross argued that this copying should not be held against them because the headnotes did not appear in the content generated by the AI for end users. They claimed that the intermediate use of the headnotes for training purposes was not the same as directly reproducing them in their product. However, the court rejected this argument, emphasizing that the unauthorized copying of a substantial portion of the headnotes for commercial purposes, even at an intermediate stage, still constituted copyright infringement.

In evaluating Ross’ fair use defense, the Court considered the four statutory factors under American copyright law and found that upon weighing these factors, Ross' use did not qualify as fair use:

1. Use’s Purpose and Character: Ross’use was commercial and not transformative because Ross’ AI tool served a similar (and competing) purpose to Westlaw, adding no significant new expression, meaning, or value. The Court distinguished Ross’ “intermediate” use against previous decisions, such as Google v. Oracle, where intermediate copying was permitted under the first fair use factor, because those cases were about copying computer code, not literary works, and in Google, the copying was necessary for innovation. The Court emphasized that this case was more akin to the US Supreme Court decision in Warhol Foundation v. Goldsmith, where the commercial and competing nature of the infringing use was determinative. This factor favored Thomson Reuters.
2. Copyrighted Work’s Nature: The Court considered that each headnote was an individual copyrighted work. However, although Westlaw’s material had more than the minimal spark of originality required for copyright validity, the Court found the headnotes were ultimately not that creative. The Court highlighted that when the copyright is “thin” (or less expressive), the more exact the copy must be. This factor favored Ross.
3. Amount of Use: This factor looks at the quantity and quality of the infringing use. Ross copied a substantial portion of headnotes (2,243 in total), which the Court considered excessive and which weighed against fair use. Nonetheless, because Ross did not make the Westlaw headnotes available to the public (as they were only used in an intermediate step), this factor ultimately favoured Ross.
4. Effect on the Market: The Court concluded that Ross' use negatively impacted the potential market for the Westlaw headnotes. Furthermore, Ross could have trained its AI without infringing Thomson Reuters’ copyright since Ross could have created headnotes itself or hired LegalEase to do so. This factor favored Thomson Reuters. According to the Court, this factor was “undoubtedly the single most important element of [the] fair use [analysis].”

This decision underscores the limitations of the fair use defense in cases of direct competition and commercial exploitation, especially in an AI context.

How This Decision Might Have Differed in Canada

If the Thomson Reuters decision were adjudicated under Canadian law, the analysis would have differed due to the distinct fair dealing framework in Canada, but the outcome would have likely been the same.

Canadian courts apply a two-step test for fair dealing. First, determining if the dealing falls under one of the enumerated purposes in the Copyright Act: research, private study, criticism, review, or news reporting. Then, assessing the fairness of the dealing using six factors: the purpose of the dealing, the character of the dealing, the amount of the dealing, alternatives to the dealing, the nature of the work, and the effect of the dealing on the work.

Interestingly, one of Canada’s landmark copyright and fair dealing decisions, CCH Canadian v. Law Society of Upper Canada, deals with the use of headnotes and case summaries. In CCH, the Supreme Court found that photocopying headnotes and case summaries for research and private study purposes was fair dealing.

In this case, Ross could argue that its use of Westlaw’s headnotes for training its AI tool fell under the purpose of research. While the fairness analysis would be different, it would likely lead to the same conclusion:

1. Purpose of the Dealing: While Ross could argue the purpose was for the development of a legal research tool (i.e. for the allowable purpose of “research”), the commercial nature of the use would weigh against fairness.
2. Character of the Dealing: Under Canadian law, this factor looks at how the works are dealt with, for example are the copied works widely distributed, is a single copy made, are there industry practices at play, etc. Here, the systematic and extensive copying of headnotes might be seen as less fair, especially as the copied materials were used to create a competing product. However, the fact that the headnotes were ultimately not circulated to the public might mitigate this factor.
3. Amount of the Dealing: Similar to the third factor under the American fair use analysis, this factor looks at the quantity and quality of the infringing use. In Ross’ case, the significant number of headnotes copied (2,243) would likely be seen as excessive, particularly given that the copying involved key elements of the original works.
4. Alternatives to the Dealing: In Canada, whether there are non-infringing alternatives is a standalone factor, whereas in the US, this is weaved into the fourth factor. Here, Ross could have developed its AI tool without copying the headnotes, since it could have hired LegalEase to develop headnotes without copying Thomson Reuters’ works. This would have weighed against a finding of fairness.
5. Nature of the Work: The headnotes, while factual, involve editorial judgment and creativity, which Canadian courts are likely to protect, similar to the US. Indeed, the Canadian landmark copyright decision, CCH, involved similar legal headnotes material.
6. Effect of the Dealing on the Work: The negative impact on Thomson Reuters' market, given that Ross' product directly competed with Westlaw, would strongly weigh against fairness, as it did in the US fair use analysis.

While the Canadian framework does not explicitly consider “transformativeness,” the focus on the fairness of the dealing, including the commercial nature and market impact, would likely lead to a similar conclusion. Thus, despite differences in the analytical approach, the outcome, i.e. a finding against Ross, would likely remain the same due to the substantial and competitive nature of the copying.

Canadian Cases to Look Out For

In Canada, two significant cases to watch regarding copyright infringement and generative AI are CanLII v. Caseway AI (B.C. Supreme Court) and Toronto Star, Metroland Media, Postmedia, The Globe and Mail, The Canadian Press, and CBC v. OpenAI (Ontario’s Superior Court).

The CanLII v. Caseway AI case involves the Canadian Legal Information Institute's claim that Caseway AI used its legal databases to train AI models without permission, raising questions about the copyright protection of legal databases and the applicability of fair dealing.

The media organizations' case against OpenAI centers on allegations that OpenAI used their copyrighted content to train its AI models, potentially infringing on their copyright. This case will examine whether OpenAI's use of news articles and reports can be justified under Canadian fair dealing provisions.

Both cases are poised to set important precedents for how Canadian courts handle copyright issues in the context of AI and will provide valuable guidance for the technology sector.