Introduction: A Landmark Decision, Still Unfolding
The intersection of artificial intelligence and copyright law continues to be a legal gray area, and it has only grown more layered since courts first began weighing in. On February 11, 2025, Judge Stephanos Bibas of the U.S. Court of Appeals for the Third Circuit (sitting by designation in the U.S. District Court for the District of Delaware) issued the first U.S. federal court decision addressing fair use in the context of AI training data. In Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., 765 F. Supp. 3d 382 (D. Del. 2025), the court ruled that Ross Intelligence’s use of Thomson Reuters’s Westlaw headnotes to train a nongenerative AI tool was not protected by fair use.
That ruling is not the end of the story. The Third Circuit granted Ross permission to pursue an interlocutory appeal, heard oral argument on June 11, 2026, and has not yet issued a decision as of this writing. District court proceedings remain stayed pending that appeal. Any assessment of Thomson Reuters v. Ross today has to account for the fact that its final scope is still being decided, and reporting from the argument suggests at least some appellate skepticism toward Ross’s transformative-use theory, though the outcome remains genuinely open.
At the time of the district court’s ruling, the decision sparked widespread media coverage, with many commentators suggesting it set a precedent for the flood of AI-related copyright lawsuits then pending. However, despite its significance, the Thomson Reuters v. Ross ruling has had a narrower applicability than some headlines suggested, and rulings issued in other cases just months later reinforced that point directly. Key differences, such as the nature of Ross’s product, access to the copyrighted works, and the type of AI involved, limited the breadth of the case’s influence even before the appeal was granted.
This article revisits those distinctions, examines how they have already played out in other courts, and offers practical takeaways for creators and companies navigating the evolving legal landscape of AI and copyright.
1. Competing Products Make All the Difference
Why Ross’s AI Tool Was Seen as a Direct Market Substitute
A central reason Ross lost on fair use’s fourth factor, effect on the market, was that its AI tool directly competed with Thomson Reuters’s Westlaw platform. Ross used Westlaw’s copyrighted headnotes to develop a similar legal research product, creating a clear market substitute. Even if Ross’s AI didn’t replicate Westlaw’s user interface, the underlying use harmed Thomson Reuters’s potential market, including the market for licensing training data.
By contrast, many AI copyright cases involve generative AI systems trained on publicly available online articles, where the content owners aren’t developing competing AI tools. Without direct market competition, courts may reach different conclusions on fair use’s market-effect factor—and as discussed below, they already have.
2. Paywalls Matter: Access Restrictions Shaped the Court’s View
Why Westlaw’s Paywall Strengthened Thomson Reuters’s Case
Westlaw’s headnotes aren’t publicly accessible; they’re behind a paywall available only to subscribers. Although copyright protection doesn’t require restricted access, Thomson Reuters’s paywall and user agreements signaled its intent to control how the headnotes were used. Ross’s unauthorized access and use of this content, possibly in violation of Westlaw’s terms of service, further undermined its fair use argument.
This contrasts with other AI lawsuits that involve freely accessible online materials. When content is publicly available, especially without paywalls or usage restrictions, courts might weigh fair use factors differently, particularly if the AI tool’s use is more transformative. As a later section of this article shows, access alone is not the whole picture. How the underlying copies were obtained matters just as much.
3. Generative vs. Nongenerative AI: Why That Distinction Matters
Ross’s Tool Didn’t “Transform” the Original Work
One of the most critical factors in the court’s decision was the type of AI at issue. Fair use’s first factor looks to the purpose and character of the use of the copyrighted material, particularly whether the use was “transformative.” Ross’s technology was nongenerative. It used the headnotes to enhance a legal research tool that performed the same basic function as Westlaw: providing legal information. Because the use was not transformative, the court found against Ross on fair use’s first factor.
Many AI copyright cases, however, involve generative AI models that synthesize training data into entirely new outputs (such as original text, images, or code) that differ from the purposes of the original works. When the authors first began to draft this article, courts had yet to address how these transformative uses fit into the fair use framework. That has since changed.
4. The Generative AI Test Has Already Arrived: Bartz and Kadrey
This article’s central prediction, that generative AI training might fare differently than Ross’s nongenerative tool, did not stay hypothetical for long. Within weeks of the Thomson Reuters decision, two Northern District of California judges ruled on closely related questions involving large language models trained on copyrighted books.
In Bartz v. Anthropic, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), the court held that training Anthropic’s models on lawfully acquired copyrighted books was fair use, finding the training process highly transformative. The same court reached the opposite conclusion for a different category of inputs: books Anthropic had obtained from pirate libraries and retained in a general-purpose corpus, which the court found infringing regardless of how the resulting models were later trained. Just days later, in Kadrey v. Meta (788 F. Supp. 3d 1026 (N.D. Cal. 2025), a different judge reached a broadly similar split on comparable facts involving Meta’s Llama models.
Bartz went on to settle for $1.5 billion, the largest copyright class action settlement on record, with final court approval entered in July 2026. Kadrey remains active on the claims the court did not resolve on summary judgment, including issues tied to how the underlying training copies were acquired.
Together, these rulings sharpen, rather than simply confirm, the distinction this article draws. The line that matters most is not only generative versus nongenerative AI. It is also lawful acquisition versus unlawful acquisition of the underlying training data. A company that trains a generative model on copyrighted works it never had the right to possess in the first place may find fair use unavailable regardless of how transformative the resulting model is. That has direct, practical implications for how companies vet the provenance of training data, and the tools built on it, not just how they classify the type of AI involved.
5. Protecting Your Creative Work While Legal Uncertainty Persists
What Content Owners Can Do Now, Without Waiting for Courts to Decide
While the legal landscape evolves, creators and businesses don’t have to stand still. There are proactive steps that can be taken to protect intellectual property in the face of AI-related challenges.
- Use strong contracts: Clear licensing terms and usage restrictions can prevent unauthorized use of your content, even for AI training purposes.
- Diversify your IP protections: Patents, trademarks, and trade secrets can complement copyright protection, covering aspects that copyright doesn’t.
- Develop licensing strategies: Creating licensing models for AI training data can turn potential risks into revenue opportunities while maintaining control over how your content is used.
- Scrutinize provenance, not just purpose: As Bartz and Kadrey show, courts are drawing a sharp line between lawfully acquired and pirated training data. Companies licensing or deploying AI tools should ask how the underlying training data was sourced, not only what the AI does with it.
Conclusion: One Decision, Fewer Open Questions Than a Year Ago
The Thomson Reuters v. Ross ruling marked a pivotal moment in the evolving relationship between copyright law and artificial intelligence, but it was never a catchall precedent, and the year since has proven that out. Bartz v. Anthropic and Kadrey v. Meta already show that generative AI cases can and do come out differently, particularly where training data was lawfully acquired. At the same time, Thomson Reuters v. Ross itself remains unresolved on appeal, and the Third Circuit’s ruling, whenever it is issued, will be the first federal appellate word on fair use and AI training of any kind.
For content creators, technology companies, and legal professionals, the key takeaway is unchanged in substance but sharper in focus: Don’t assume any single decision, including this one, answers every question about AI and copyright. Stay informed, be proactive about how training data is sourced and licensed, and prepare for a legal landscape that is still being actively written, one appellate ruling at a time.
The views expressed here are the authors’ own and do not necessarily reflect the positions of their respective employers or firms.
