In February 2025, the U.S. District Court for the District of Delaware issued a landmark legal decision regarding the intersection of artificial intelligence and copyright. The case, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., involved two key players in the legal and technology worlds and set a precedent that sets the tone for future litigation related to AI training with copyrighted content.
Background and context of the case
In May 2020, Thomson Reuters sued Ross Intelligence for using Westlaw’s editorial content, specifically its headnotes—legal summaries created by human editors and organized using the “Key Number” system—without permission to train an AI-powered legal search engine.
Ross initially attempted to obtain a license, but when Thomson Reuters refused, he acquired approximately 25,000 “Bulk Memos” through the firm LegalEase. These attorney-created memos accurately reflected the content of Westlaw’s headnotes and were used to train its AI.
In September 2023, Judge Bibas (a Third Circuit Court appointee) partially denied the summary motions, leaving the matter to a jury. However, he reviewed his own decision in the months leading up to trial and allowed the summary motions to be renewed.
El fallo del 11 de febrero de 2025
On February 11, 2025, Judge Bibas issued a judgment that:
It granted partial summary judgment to Thomson Reuters, finding that the headnotes were original copyrighted works and that Ross directly copied at least 2,243 of them.
It rejected all of Ross’s arguments, including fair use, finding them flawed as a matter of law.
This ruling is considered the first case in the US where a court explicitly concludes that the use of copyrighted content to train an AI is not covered by fair use.
Analysis of the Fair Use Factors
The court evaluated the four legal factors and concluded as follows:
Purpose and nature of use:
Commercial, as Ross was developing a rival product to Westlaw.
Not transformative, as it replicated the target function: legal search.
Nature of copyrighted material:
Although the headnotes contained a lower level of creativity, they were still original and protected. This factor partially favored Ross, but it wasn’t decisive.
Quantity and substantiality of use:
Ross copied extensively, even though the results did not contain the headnotes directly.
However, the use was extensive and relevant to his function, thus unfavorable to fair use.
Market Impact:
Creation of a direct replacement for Westlaw.
It affected the potential monetization of AI training licenses. This was the most critical and decisive factor.
Implicaciones para el sector IA
This ruling opens a new chapter in jurisprudence:
Precedent for other litigation: Similar cases brought by authors, entities, and technology media against OpenAI, Meta, Microsoft, etc., can cite this recent ruling.
Distinction between generative and non-generative AI: Although the ruling involves non-generative AI, its logic extends to the realm of large language models (LLMs).
Need for explicit licenses: Even “intermediate” or auxiliary content used for training may require prior consent.
Concerns on Wall Street: The ruling is already prompting warnings about potential legal risks for investors in AI companies.
What innovative companies and developers can learn from AI
Review the licenses and permissions for all data used to train models.
Assess legal risks before incorporating proprietary or competitive content.
Explore licensing agreements with entities that own valuable databases.
Implement audit policies on the provenance of training data.
