AI, Copyright & Fair Use: Key Lessons from Thomson Reuters v. Ross Intelligence

Background of the Case

The recent summary judgment in Thomson Reuters v. Ross Intelligence marks a pivotal moment in copyright law, particularly for artificial intelligence (AI). Thomson Reuters, the parent company of Westlaw, sued Ross Intelligence, an AI-driven legal research startup, for allegedly infringing on its proprietary headnotes and Key Number System. After being denied a license to Westlaw’s content, Ross allegedly obtained Westlaw’s editorial materials through a third-party contractor, LegalEase, which provided thousands of "Bulk Memos" derived from Westlaw headnotes.

In 2023, the court denied summary judgment, questioning the originality of Westlaw’s headnotes and leaving the fair use defense for a jury. However, in 2025, Judge Stephanos Bibas granted partial summary judgment to Thomson Reuters. The court ruled that Ross infringed 2,243 headnotes, rejected Ross’s fair use defense, and emphasized the originality of Thomson Reuters’ compilations.

Key Takeaways from the Judgment

1. Originality of Derivative Works

The court emphasized that the threshold for originality in copyright is "extremely low," requiring only a "minimal degree of creativity." The focus is on whether a work is original, not the effort put into it.

The court determined that Westlaw's headnotes exhibit sufficient creativity for copyright protection, likening their creation to sculpting raw marble into a protected work. This ruling reinforces the copyrightability of curated legal compilations.

2. Fair Use Defense Rejected

Ross argued that its AI model served a different purpose than Westlaw’s headnotes, making its use transformative. However, the court ruled that Ross merely restructured and repackaged the content for a competing legal research tool rather than creating something new. The ruling suggests AI developers must be cautious when training models on proprietary datasets, as intermediate copying, even without verbatim reproduction, may not qualify as fair use.

The court acknowledged that intermediate copying can sometimes be permitted under fair use. Unlike in Google v. Oracle (2021), the court emphasized that Ross’s commercial purpose and lack of necessity to copy (unlike code interoperability) weighed against fair use.

3. Market Harm

The court placed considerable weight on the fourth fair use factor: market impact. Because Ross sought to create a competing product that would undermine Westlaw’s market position, the court found its use of Thomson Reuters’ content unjustified. The ruling underscores that even if an AI model does not output copyrighted material verbatim, its training process can still be infringing if it diminishes the commercial viability of the original work.

The court weighed all four fair use factors:

1. Purpose and character of the use
2. Nature of the copyrighted work
3. Amount and substantiality of the portion used
4. Effect on the market of the original work

The first and fourth factors favored Thomson Reuters, ultimately leading to the ruling against Ross.

Critique of the Court's Ruling

1. Narrow View of Transformative Use

The court dismissed Ross’s AI training as non-transformative because it competes with Westlaw. However, AI tools parse data to generate insights, a fundamentally different purpose than presenting headnotes. By conflating end-use competition with process innovation, the decision stifles AI development across industries.

The court’s narrow interpretation of transformative use, especially in the context of AI development, is debatable. It distinguished this case from other intermediate copying cases involving computer code, arguing that "computer programs differ from books, films, and many other literary works" because they serve functional purposes. However, this distinction may not be appropriate in the context of AI, where vast amounts of data are needed for innovation.

2. Disproportionate Weight on Market Harm

While the court correctly identifies the four fair use factors, it places excessive weight on the commercial nature of Ross’s use. The court noted that the first and fourth factors weigh most heavily. This analysis gives little consideration to the public interest in AI tools that could democratize access to legal research. Although the court argued that legal opinions are freely available, it overlooked the value of headnotes in making legal research more accessible.

3. Overextension of Originality

The court’s analogy of headnotes to “sculpting marble” risks lowering the originality threshold. By equating verbatim excerpts from judicial opinions (public domain) to creative works, the ruling grants copyright holders undue control over factual compilations. This contradicts Feist Publications, Inc. v. Rural Telephone Service Co. (1991), which held that “sweat of the brow” alone does not justify copyright protection.

4. Adverse Effect on AI Innovation

By prioritizing potential markets (e.g., hypothetical AI-training licenses), the ruling forces developers to navigate a labyrinth of licensing agreements or risk litigation. This disadvantages startups that lack the resources to negotiate with established players like Thomson Reuters.

5. Lack of Consideration for AI Uniqueness

The court did not fully address the unique challenges posed by AI in copyright law. The court focused more on existing copyright doctrines, rather than trying to determine how new AI technologies fit into these frameworks. The court was very clear that it was only dealing with non-generative AI in this instance

Implications of the Court's Ruling

Implications for Copyright Holders

1. Stronger Protections for Derivative Works

The ruling affirms that even minimal creativity (e.g., curating headnotes or designing taxonomies like Westlaw’s Key Number System) qualifies for copyright protection. Plaintiffs can now pursue infringement claims even if copyrighted material is used indirectly (e.g., in AI training data), provided the use lacks a transformative purpose.

2. New Revenue Streams & Market Leverage

The court highlighted the potential market for AI training data, reinforcing the value of structured data as a commercial commodity. Rights holders may explore licensing agreements to monetize their content, strengthening their negotiating power for AI training datasets.

3. Building a Strong Case

Plaintiffs should focus on gathering evidence that establishes a causal connection between their copyrighted material and the AI developer's training data. The ruling suggests that commercial and non-transformative uses are more likely to be deemed infringement than fair use. Market harm, including lost licensing revenue or competition in AI-driven sectors, should be a focal point in litigation.

Implications for AI Companies and AI Developers

1. Data Is King, but Not Always Free

The ruling clarifies that copyrighted material cannot be freely used for commercial AI training. AI developers should carefully vet their data sources and consider licensing agreements to avoid costly litigation.

2. A Narrow View of Transformative Use

The court ruled that Ross's use was not transformative because it directly substituted Westlaw’s service. This narrow interpretation of transformative use could be challenging for AI developers who train models on copyrighted works without directly reproducing them in their output.

3. Innovation vs. Copyright

The ruling favors established copyright holders, narrowing the pathway for AI developers to rely on fair use in training models. This decision may push companies to secure licenses or develop entirely original datasets.

Expect increased investment in AI-generated datasets to bypass copyright risks. AI developers should explore public domain data, develop synthetic data models, or crowdsource annotations to reduce dependence on proprietary databases.

Conclusion

Thomson Reuters v. Ross Intelligence is a landmark case that highlights the tension between copyright law and the rapid advancement of AI. The judgment favors content creators and requires AI developers to tread carefully when leveraging third-party materials. As AI evolves, courts will grapple with nuanced questions about transformative use, market harm, and public interest. Tech companies must proactively audit their training data sources and consider licensing frameworks to mitigate legal risks.

This decision is unlikely to be the final word. As Judge Bibas noted, "A smart man knows when he is right; a wise man knows when he is wrong. Wisdom does not always find me, so I try to embrace it when it does—even if it comes late, as it did here." and the AI industry will undoubtedly seek further clarity through appeals and future litigation.

*This blog post is for informational purposes only and does not constitute legal advice.

*This blog post was written with the aid of generative AI tools.