AI Training and Copyright: Third Circuit Rejects Fair Use in Thomson Reuters v. Ross

The decision of the US federal court in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. is one of the most significant decisions to date at the intersection of copyright law and the development of artificial intelligence systems.

The court affirmed the lower court’s judgment in favor of Thomson Reuters and held that Ross’s use of materials derived from Westlaw headnotes to develop an AI powered legal search tool was not protected by the fair use doctrine.

At the same time, the decision should be read carefully. It does not hold that every use of copyrighted material for AI training is unlawful. The case concerned a very specific set of facts: the use of Westlaw’s editorial legal content to develop a competing legal search tool, in a market where the content and its licensing have clear commercial value.

Westlaw headnotes and Ross: what did Ross’s system learn from Westlaw?

Westlaw is one of the leading legal research databases in the United States. In addition to judicial opinions themselves, which are not protected by copyright, Westlaw includes editorial content prepared by legal editors. A central part of that editorial layer is the headnotes: short summaries that identify and formulate legal points drawn from court decisions.

The court held that these headnotes were not merely copies of the underlying judicial opinions. Although they were based on public legal materials, they involved sufficient selection, wording, summarization and legal editing to qualify for copyright protection.

The dispute focused on thousands of Westlaw headnotes and on whether Ross’s use of those headnotes, or materials derived from them, to develop its system amounted to infringement that was not protected by fair use.

Ross sought to develop an AI powered legal search tool. The tool was intended to help users locate relevant legal information. In that sense, it operated in a market very close to Westlaw’s own market. This fact was central to the legal analysis: the case was not about an abstract technological use of protected material, but about the use of such material to develop a competing product in the legal research market.

The legal question: was this fair use?

The fair use doctrine permits, in certain circumstances, the use of a copyrighted work without the permission of the rights holder. The analysis considers several factors, including the purpose of the use, the nature of the work, the amount used, and the effect of the use on the market for the work or for potential markets.
In this case, the question was not whether artificial intelligence is desirable or whether technological innovation should be encouraged. The question was narrower: can the use of protected editorial content to build a commercial legal search tool that competes with the rights holder be considered fair use?

The court treated the case as a relatively conventional copyright dispute, rather than as a broad ruling on the future of legal AI.

The US court’s decision

The US court confirmed that Westlaw’s legal headnotes are protected by copyright. It accepted the view that even where the starting point is a judicial opinion that is not itself protected, the legal editing, selection and wording of the headnote may be sufficiently original to attract protection.

The court also held that Ross’s use was not fair use. One of the central considerations was the commercial nature of the use and the fact that Ross’s system was intended to compete with Thomson Reuters’ product.

In this context, the court also considered whether the use was “transformative”. In simple terms, this asks whether Ross used the protected material for a genuinely new and different purpose, with a different character, rather than using it to create a product serving a similar function. The closer the use is to the original purpose of the protected work, and the more the new product operates in the same market and competes with the rights holder, the harder it is to argue that the use is sufficiently transformative to justify use without a licence.

Another important consideration was the potential effect on Thomson Reuters’ market. The court did not look only at whether Ross copied text and displayed it to users. It also considered whether the use harmed Thomson Reuters’ commercial market, including a potential market for licensing content for AI development and training.

This point is especially important for companies that own professional databases, editorial content, or commercially valuable data.

What questions remain open?

It is equally important to understand what the decision does not decide. The decision should not be read as holding that every use of copyrighted works for AI development or training is necessarily copyright infringement.

The case concerned a legal search tool that competed directly with the rights holder’s product. It did not concern a general purpose generative AI model trained on broad and diverse datasets. For that reason, the implications of the decision for general purpose generative AI models, or for uses that are not directly competitive, will need to be assessed on the facts of each case.

Ross’s case was relatively focused: the use of editorial legal content to develop a commercial legal product operating in a market very close to that of the rights holder.

Implications for development, licensing and risk management

The decision highlights several practical points for companies developing AI systems, particularly in professional sectors that rely on databases, editorial content, or commercially valuable information.

• First, the source of training or development materials matters. When an AI system is developed using professional content, summaries, classifications, analysis, or other edited content, it is not enough to ask whether the information is “available” or whether part of it is based on public material. It is necessary to consider whether there is a layer of selection, arrangement, wording or editing that may be protected by copyright.
• Second, it matters whether the new product competes with the rights holder’s product. The use of protected material to develop a product that operates in the same market, serves the same purposes and targets the same customers will face a more difficult path to fair use.
• Third, the existence of an actual or potential licensing market may be significant. Where the rights holder can show that there is commercial value in licensing the content for AI development or training, unauthorized use of that content may be treated as harming that market.
• Finally, the mere use of AI technology does not make the use transformative. Courts will examine the actual purpose of the use, the nature of the new product, the proximity between that product and the original product, and the commercial impact on the rights holder. From a practical perspective, the decision reinforces the importance of early copyright diligence, documenting data sources, assessing licensing options and evaluating legal risk before using existing content to develop AI systems.

Conclusion

The decision in Thomson Reuters v. Ross is an important milestone in the development of copyright law in the age of artificial intelligence. It does not close the broader debate over when AI training or AI development based on copyrighted works may qualify as fair use.

It does, however, mark an important boundary: where a company uses protected editorial content to build a commercial product that competes with the rights holder, and especially where there is or may be a licensing market for that content, the fair use defense may become significantly more difficult.