Protect your works from unauthorised use in AI datasets.
Calculate the quote
Avvocato Arlo Canella
Can the use of artificial intelligence in legal research violate copyright laws? The case of Thomson Reuters vs. Ross Intelligence has set a significant precedent in the United States. The court ruled that Ross Intelligence had unlawfully used Thomson Reuters’ legal summaries to train its AI, constituting direct copyright infringement. This decision, based on the fair use doctrine, stands in contrast to recent European rulings, such as the LAION case in Germany, where data collection for AI training was deemed lawful under the research exception provided by Directive (EU) 2019/790. But why did AI lose in the U.S. while it prevailed in Europe? What are the implications of this ruling for the future of AI and copyright?
The legal battle began in 2020 when Ross Intelligence, an emerging AI startup specializing in legal research, developed a natural language-based legal search engine. Its ambitious goal was to provide lawyers with a tool capable of analyzing judicial decisions and retrieving relevant citations, simplifying the legal research process.
However, the core of the dispute lies in the data used to train the AI. Ross Intelligence built its system using an essential element of any legal database: headnotes, which are concise summaries of judicial rulings highlighting key legal points. These headnotes, crafted by legal experts, are the backbone of Westlaw, the renowned legal research platform owned by Thomson Reuters.
Aware of the strategic value of these summaries, Ross Intelligence initially sought a license from Thomson Reuters, which was denied. Following this refusal, the startup adopted an alternative approach: it commissioned a third-party company to create approximately 25,000 “Bulk Memos”, documents containing legal questions paired with both correct and incorrect answers. However, during the trial, it emerged that these memos had been generated using Westlaw’s headnotes, often replicating their language and structure verbatim.
Faced with this evidence, Thomson Reuters filed a copyright infringement lawsuit, arguing that Ross Intelligence had exploited its copyrighted materials to train a competing AI without authorization or compensation for the editorial work behind Westlaw’s legal summaries.
After an initial dismissal in 2023, the United States District Court for the District of Delaware re-evaluated the case and, on February 11, 2025, issued one of the first rulings directly addressing the use of copyrighted content in AI training.
Judge Stephanos Bibas of the U.S. Court of Appeals for the Third Circuit, sitting as a designated judge in the district court, ruled in Thomson Reuters Enterprise Centre GmbH & West Publishing Corporation v. Ross Intelligence Inc. (Civil Action No. 20-613-LPS) that Ross Intelligence had infringed copyright by using Westlaw’s editorial materials without authorization to train its AI-based legal search engine.
This ruling is set to create a significant precedent—not only for its direct implications but also because the judge applied the four-factor fair use test, a fundamental principle of U.S. copyright law. But what is fair use, and how did the court interpret it in this case?
The fair use doctrine is one of the most distinctive elements of U.S. copyright law. Codified under Section 107 of Title 17 of the U.S. Code, it allows the use of copyrighted works under specific circumstances without the copyright holder’s consent, as long as the use is justified by certain purposes and does not undermine the market for the original work.
This principle seeks to balance copyright protection with the need to ensure freedom of expression, innovation, and scientific progress. Many digital technologies have relied on fair use to develop without immediately facing copyright violations. However, as demonstrated by the Thomson Reuters v. Ross Intelligence case, the application of this exception is not automatic — U.S. courts evaluate fair use on a case-by-case basis using a four-factor test:
1. Purpose and character of the use
The court examines whether the use is commercial or serves educational, informational, or critical purposes. Non-commercial and transformative uses are more likely to qualify as fair use.
2. Nature of the copyrighted work
Creative works (e.g., novels, films, or artworks) receive stronger copyright protection, while factual works (e.g., technical manuals or judicial decisions) may have **weaker protection.
3. Amount and substantiality of the portion used
Even small excerpts can violate copyright if they capture the essence of the original work.
4. Effect on the market for the original work
This is often the most critical factor — if unauthorized use competes with or diminishes the market value of the original work, it is unlikely to be considered fair use.
Tech companies have frequently invoked fair use to justify using copyrighted content for AI training. However, this ruling is one of the first to explicitly reject this defense, clearly stating that using copyrighted content for training a commercial AI is not automatically protected under fair use.
Judge Stephanos Bibas rigorously applied the four-factor fair use test.
The first factor (purpose and character of use) weighed heavily against Ross Intelligence. The Supreme Court’s decision in Campbell v. Acuff-Rose Music, Inc. (510 U.S. 569, 1994) established that a use is more likely to be fair use if it is highly transformative. However, the court found that Ross Intelligence’s use was not transformative — it did not alter the headnotes or use them for a new purpose but instead integrated them into its AI training data to develop a competing legal research tool.
The second factor (nature of the work) also favored Thomson Reuters. Although judicial opinions are public domain, Westlaw’s headnotes are editorial summaries requiring legal expertise, making them copyrightable works.
The third factor (amount and substantiality used) further weighed against Ross Intelligence. The startup argued that it used only a small fraction of Westlaw’s database, but the court ruled that what mattered was not the quantity but the significance of the extracted content.
The fourth and most critical factor (market impact) was decisive. The court concluded that Ross Intelligence’s AI product directly competed with Westlaw, diminishing the commercial value of the original work.
This ruling challenges a common assumption in AI development: legal summaries are not mere raw data — they are protected intellectual property.
The Thomson Reuters v. Ross Intelligence ruling has set a crucial precedent: training AI with copyrighted content is not automatically permitted under fair use. The court emphasized that legal summaries are original editorial works, not just raw data.
This decision is particularly relevant because it does not involve a generative AI model but rather an AI-powered legal search engine. The debate around copyright and AI has primarily focused on generative models like those from OpenAI, Google, and Meta, but this case highlights that even non-generative AI systems can infringe copyright.
While fair use restricted AI usage in legal research in the U.S., a parallel case in Europe had a different outcome. In September 2024, the Hamburg Regional Court ruled in favor of LAION, stating that data collection for AI training was lawful under the European Union’s research exception (Directive (EU) 2019/790).
However, a major legal loophole remains: LAION’s datasets were later used by Stability AI, whose commercial AI models (such as Stable Diffusion) power major tech companies. This raises a critical question:
Is Europe’s research exception strong enough to protect copyright holders from commercial AI exploitation?
Avvocato Arlo Canella
