Companies must be mindful of the ultimate purpose of new artificial intelligence tools to avoid running into copyright infringement issues during the training process. If widely adopted, the Thomson Reuters v. Ross Intelligence decision suggests “intermediate copying” cases are unlikely to provide a strong defense when the final output of a tool mirrors the products it was trained on. Accordingly, the key question is likely to what extent the AI system is competing with the underlying copyrighted work. The further away the system is, the more likely it is to be protected under the fair-use doctrine.
The Thomson Reuters case is a novel application of the “fair use” defense of copyright law to an AI system. The fair-use doctrine allows limited use of copyrighted materials without the copyright owner’s permission, based on a series of factors that balance the rights of copyright holders with the public interest. While there are several copyright cases involving AI moving through the courts, Thomson Reuters is one of the first substantive decisions to consider whether the use of copyrighted materials to train an AI constitutes copyright infringement or whether that copying is protected by the fair-use defense.
If widely adopted, this decision will likely have critical effects on the analysis of other AI systems, including generative AI systems. In particular, the court’s interpretation of the “intermediate copying” case law and analysis of whether the output of an AI system is a “market substitute” for the copyrighted material are likely equally applicable to generative AI systems.
Generative AI companies are typically accused of copyright infringement for two primary reasons—using copyrighted materials to train the AI system without the copyright holder’s permission or generating a derivative work or close copy of copyrighted materials as an output.
A generative AI system that can create new works is referred to here as being a different type of AI system from a non-generative AI system.
In Thomson Reuters, the training data used by a non-generative AI system was the only basis for infringement. While expressly limited to non-generative AI aspects of the defendant’s system, the fair-use analysis in Thomson Reuters likely has equal application to generative AI cases. However, generative AI systems have a variety of unique aspects that might distinguish them from the same outcome.
Intermediate Copying Defense
“Intermediate copying” refers to duplicating a work during the development of a different product. In Thomson Reuters, Ross Intelligence argued that its intermediate copying of turning the copied works into numerical data to train the non-generative AI system was fair use under copyright law. However, the court disagreed, concluding that Ross’s copying was unlike the cited fair-use case law, which were all software cases in which the copying was necessary for competitors to innovate.
If widely adopted, this analysis likely applies equally to generative AI systems. While the training of non-generative AI systems differs from generative AI systems, the core question will be whether the copying was necessary to innovate, or whether it was just a shortcut.
Like in Thomson Reuters, while taking copyrighted materials might “make it easier to develop a competing legal research tool” or other product, it isn’t necessary for a generative AI system to do so.
Consider a generative AI system for creating new music: Using copyrighted music to train a system to create new music might be easier, be more convenient, or provide output more likely to be popular. But there is no shortage of public domain music that could be used instead. Therefore, it is likely not “necessary” to use that copyrighted music to train the system merely to create new music—especially music in the style of an artist whose copyrighted music the system was trained on.
One caveat to consider is that the Thomson Reuters court distinguished the computer programs of the earlier intermediate copying cases from other kinds of text, so future cases dealing with software may have to take that distinction into account.
Market Substitutes
In analyzing the four fair-use factors, the Thomson Reuters court clearly emphasized one of them—the effect of the use on the potential market for, or value of, the copyrighted work.
Not all AI systems create the same outputs. This is where the “market substitute” criterion becomes critical.
In Thomson Reuters, the court concluded that Ross “meant to compete with Westlaw by developing a market substitute.” In a similar vein, if a generative AI system is trained on copyrighted materials it intends to ultimately compete against (such as an artist’s copyrighted music to create new music in the style of that artist), the system operator will likely have a weaker fair-use argument because it is likely to create a market substitute to those works used for training.
But generative AI systems vary dramatically in terms of application. If the Thomson Reuters analytical framework is widely adopted, the more a generative AI system is like a tool that helps users innovate or create—without competing with the works used to train the system—the stronger the fair-use argument will likely be. Similarly, the further the generative AI system’s outputs are from the underlying copyrighted materials, the stronger the fair-use argument will likely be because those new types of works are less likely to affect the market for the training-set works.
The case is Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., D. Del., No. 1:20-cv-613-SB, decided 2/11/25.
This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law and Bloomberg Tax, or its owners.
Author Information
Jennifer Lantz is a partner in Duane Morris’ Intellectual Property Practice Group in its Silicon Valley office.
Jeremy Elman is a partner in Duane Morris’ Intellectual Property Practice Group in its Silicon Valley office.
Max DiBaise is an associate in Duane Morris’ Intellectual Property Practice Group in its Silicon Valley office.
Write for Us: Author Guidelines