Background of the Dispute

The case stems from ROSS Intelligence's efforts to create an AI-powered legal research platform to compete with Thomson Reuters' Westlaw service. When ROSS sought to develop its AI tool, it needed a substantial database of legal questions and answers for training data. After Thomson Reuters declined to license Westlaw's content to ROSS (citing their status as competitors), ROSS contracted with LegalEase, a legal research company, to obtain training data in the form of "Bulk Memos."

These Bulk Memos were compilations of legal questions with answers created by lawyers using Westlaw's headnotes as source material. LegalEase provided its lawyers with guidelines on creating these questions, specifically instructing them not to directly copy and paste Westlaw headnotes. Nevertheless, ROSS ultimately obtained and used approximately 25,000 Bulk Memos to train its AI search tool.

When Thomson Reuters discovered this arrangement, it filed suit alleging copyright infringement of its headnotes, which are copyrightable summaries of key points of law from judicial opinions. The case has significant implications for both the legal research market and the broader landscape of AI development.

In a significant decision at the intersection of artificial intelligence and copyright law, the U.S. District Court for the District of Delaware has now ruled largely in favor of Thomson Reuters. The court's opinion provides important guidance on the use of copyrighted materials to train AI systems and the application of fair use in the AI context.

Key Holdings

• The court granted partial summary judgment to Thomson Reuters on direct copyright infringement for 2,243 headnotes, finding that ROSS had copied Thomson Reuters' copyrighted content to train its AI legal research tool.
• The court rejected all of ROSS's defenses, including innocent infringement, copyright misuse, merger doctrine, scenes à faire, and notably, fair use.
• While some copyrighted materials may have expired or been untimely created (an issue left for the jury), the court found Thomson Reuters' headnotes and Key Number System met the threshold for copyright protection.

Fair Use Analysis

The court conducted a comprehensive analysis of all four statutory fair use factors, with factors one and four weighing most heavily in the analysis:

1. Purpose and Character of Use (Favored Thomson Reuters)
• Commercial Use: ROSS admitted its use was commercial
• Transformative Nature: Court found ROSS's use was not transformative because:
• Both companies used the content for legal research tools
• ROSS's AI was not generative AI but rather returned existing judicial opinions
• The "intermediate copying" defense failed because, unlike software compatibility cases, copying was not necessary to access underlying ideas
• Bad Faith: While considered, was not determinative given the commercial and non-transformative nature
2. Nature of the Copyrighted Work (Favored ROSS)
• Court found Westlaw's materials had more than minimal originality but were not highly creative
• Headnotes required editorial creativity and judgment, but less than creating work from scratch
• Key Number System was a factual compilation with limited creativity
• This factor "rarely played a significant role" in fair use analysis
3. Amount and Substantiality Used (Favored ROSS)
• Court focused on what was made accessible to the public, not the amount copied
• Since ROSS's output to end users did not include Westlaw’s headnotes, this factor favored ROSS
• Court rejected ROSS's argument about percentage of total headnotes taken, noting even small portions can be substantial if they constitute the "heart" of the work
4. Effect on Market (Favored Thomson Reuters)
• Considered the "single most important element of fair use"
• Court identified two affected markets:
• Current market for legal research platforms
• Potential derivative market for AI training data
• Rejected argument about public benefit in accessing law, noting:
• Legal opinions are already freely available
• Public interest in subject matter alone is insufficient
• Copyrights incentivize development of legal research tools

In balancing these factors, while factors two and three favored ROSS, the court gave greater weight to factors one and four, ultimately rejecting ROSS's fair use defense.

Implications for AI Development

This decision has significant implications for companies developing AI systems:

1. The opinion distinguishes between generative and non-generative AI, specifically noting that only non-generative AI was at issue here.
2. The court rejected the argument that "intermediate copying" for AI training constitutes fair use, distinguishing this case from software compatibility cases like Google v. Oracle, 593 U.S. 1 (2021).
3. Companies cannot avoid copyright infringement by merely transforming copyrighted content into training data if the end product serves a similar market purpose.

Key Takeaways

• Companies developing AI systems should carefully evaluate their AI training practices and data sources, and consider whether obtaining licenses is necessary.
• When using materials packaged by a third-party vendor to train AI systems, companies should conduct reasonable due diligence as to the origin of the data and information and implement appropriate safe guards to protect against unintentional copying.
• The decision suggests that using copyrighted materials to train AI systems that compete with the original work's market is likely to face significant legal scrutiny.
• The opinion leaves open questions about how this analysis might apply to training material that is not substantially similar to the original work, apply to generative AI systems, or uses that serve different market purposes.

Important Distinctions and Scope Limitations

This ruling's scope may be significantly limited due to key distinctions between ROSS's technology and current generative AI systems:

• Unlike generative AI systems, ROSS's platform did not create new content but instead returned existing judicial opinions in response to user queries—a process fundamentally similar to how Westlaw's own platform operates.
• This operational similarity was crucial to the court's analysis, particularly in rejecting the fair use defense, as both platforms served essentially the same market purpose.
• In contrast, pending generative AI cases involve systems that create new content, with defendants arguing there is a fundamental difference in how copyrighted works are used. These defendants contend their systems are not interested in the protected "expression" in copyrighted works but rather in the underlying data and patterns they embody.
• The transformative use analysis could differ significantly for generative AI cases. While Judge Bibas found ROSS's product was not transformative because it was commercial and functionally similar to Westlaw's platform, the analysis becomes more complex with large language models. These models transform copyrighted works into mathematical representations that enable the creation of entirely new works, potentially making the transformative use argument stronger than in cases where, like here, the AI tool simply retrieves and delivers existing content.
• Significantly, generative AI more directly promotes the Copyright Act's fundamental constitutional purpose. The U.S. Constitution, in Article I, Section 8, Clause 8, explicitly states that Congress shall have the power "To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries." By creating new works and advancing creative expression, generative AI arguably furthers this constitutional objective in a way that ROSS's mere repackaging of existing content does not.

These distinctions suggest that courts may need to develop different analytical frameworks when addressing copyright claims involving generative AI systems that transform and recombine source material rather than simply retrieving and displaying it.

Big Picture

This case is one of the first in a series of cases concerning allegations of copyright infringement arising from the training of AI. Many of these cases have been brought by book authors, alleging that their copyrighted works were used to train the defendants’ AI, which, according to the plaintiffs, constitute copyright infringement. The court’s decision in this case to strengthen copyright laws in favor of copyright holders has set a benchmark that all subsequent cases will be compared to. In other words, courts around the U.S. are likely to use the court’s finding of copyright infringement to gauge the merits of the plaintiff’s copyright infringement claims arising from AI training and decide accordingly. Many major players in the world of AI are involved in these cases that are still pending, and the AI landscape is likely to change very rapidly in the near future. Any company operating in the AI space or even tangentially connected in any way is going to need experts with extensive knowledge in the field of AI to respond quickly and strategically when a new holding comes down. Dentons US offers this expertise and is ready to provide it, along with a whole host of other services, to its clients.

Contact

For questions about how this opinion may affect your business or for assistance with IP or AI questions, please contact Victor Johnson (victor.johnson@dentons.com), James Tuck (james.tuck@dentons.com), Vivian Sandoval (vivian.sandoval@dentons.com) or anyone from the Dentons US Intellectual Property & Technology Group.