The Limits of Precedent: Decoding the Thomson Reuters v. Ross Intelligence Ruling

The legal landscape surrounding Artificial Intelligence (AI) and intellectual property has just witnessed a significant, albeit highly specialized, shift. Yesterday, the United States Third Circuit Court of Appeals officially unsealed its opinion in Thomson Reuters v. Ross Intelligence, a protracted legal battle that has been closely watched by copyright holders and technology developers alike.

The court affirmed the district court’s earlier ruling, declaring that Ross Intelligence’s unauthorized use of Thomson Reuters’ proprietary legal headnotes to train its AI system did not constitute “fair use.” While the headlines have been quick to brand this a “landmark victory” for copyright holders in the age of AI, a deeper examination of the court’s reasoning suggests that the decision is more of a narrow carve-out than a sweeping mandate for the future of generative AI.

As the legal system grapples with the tension between technological innovation and the protection of creative labor, this ruling provides a critical, if limited, data point. To understand why this case matters—and why its impact on giants like OpenAI or Anthropic may be minimal—one must look beyond the verdict and into the specific, fact-dependent circumstances that defined the litigation.

Chronology of a Legal Conflict

The dispute between the legal information behemoth Thomson Reuters (owner of the Westlaw platform) and the now-defunct AI startup Ross Intelligence began in earnest in December 2020. However, the seeds of the conflict were sown years prior.

• The Origins (Late 2010s): Ross Intelligence sought to disrupt the legal research market by developing an AI-driven platform capable of competing directly with Westlaw. Recognizing that their AI needed high-quality training data, Ross approached Thomson Reuters for a licensing agreement. When Thomson Reuters declined, Ross allegedly pivoted to a strategy involving a third-party partner, LegalEASE.
• The Alleged Infringement: Ross instructed LegalEASE to generate approximately 25,000 "bulk memos" to train their model. While they claimed to have advised these lawyers not to copy-paste content from Westlaw, Thomson Reuters asserted that 2,243 of their proprietary "headnotes"—the highly valued, human-written summaries of legal points within court cases—had been misappropriated.
• The Lawsuit: Thomson Reuters filed suit in a Delaware federal court in 2020, alleging willful copyright infringement. Although Ross Intelligence ceased operations shortly after the suit was filed, the litigation persisted as a matter of principle and potential damages.
• The Procedural Seesaw: In September 2023, the presiding judge initially determined that the question of fair use was a matter for a jury, declining to rule on summary judgment. However, in a surprising reversal in February 2025, the judge ruled in favor of Thomson Reuters as a matter of law.
• The Appellate Decision: Ross Intelligence appealed the district court’s finding to the Third Circuit. Yesterday’s unsealing of the appellate opinion confirmed the lower court’s stance, solidifying the defeat for the defunct startup.

Analyzing the Four-Factor Fair Use Test

The core of the dispute rested on the "Fair Use" doctrine, a legal defense that allows for the limited use of copyrighted material without permission. Courts typically evaluate this through four specific factors. In the case of Thomson Reuters v. Ross Intelligence, the court’s analysis favored the plaintiff, largely due to the "substitutive" nature of the defendant’s actions.

1. Purpose and Character of the Use

The court found that Ross Intelligence’s use was "minimally transformative" and heavily commercial. Because the goal was to create a direct competitor to Westlaw, the AI did not create something fundamentally new or distinct enough to warrant a transformative defense. It was, in the court’s view, a replacement for the original product.

2. Nature of the Copyrighted Work

This was the only factor that provided any comfort to the defense, as the court acknowledged that the headnotes were factual in nature and had been previously published. However, this minor advantage was insufficient to overcome the weight of the other factors.

3. Amount and Substantiality of the Portion Used

The court noted that Ross Intelligence had copied the entirety of the headnotes in question. The fact that this represented only a small percentage of the total Westlaw database was irrelevant; the law focuses on the nature of the specific copyrighted work taken, not the ratio of the training set size.

4. Effect Upon the Potential Market

This was arguably the most damning factor for Ross. Because the AI was designed to provide the same functionality as Westlaw, the court determined that the model functioned as a direct market substitute. By effectively "stealing" the market value of the human-authored summaries, Ross significantly harmed the economic potential of Thomson Reuters’ intellectual property.

The "Substitutive Competition" Distinction

The most important takeaway for industry observers is the court’s nuanced differentiation between what Ross Intelligence did and what modern Generative AI models, such as ChatGPT or Claude, are doing.

Ross Intelligence was not building a general-purpose language model; it was building a vertical tool designed to replicate the output of a specific database. The court identified this as "substitutive competition." In contrast, recent rulings—such as those involving Anthropic—have leaned toward finding that using copyrighted content to train models that generate "original expression" may qualify as fair use.

This distinction is crucial. If a company uses data to create an output that essentially mimics the source material, they are vulnerable to copyright claims. If a company uses data to train a model that creates entirely new, non-substitutive content, the legal landscape is significantly murkier and arguably more favorable to the AI companies. The Thomson Reuters ruling reinforces the idea that the "market harm" factor will continue to be the primary battleground in future AI litigation.

Implications for Creators and AI Companies

For content creators, this ruling is a victory, but one that should be viewed through a sober lens. It confirms that "training data" is not a magic shield against copyright infringement if the resulting AI is designed to replace the original source.

However, the ruling also hints at an emerging reality: the rise of a licensing market. The court noted that by bypassing licensing, companies risk usurping a legitimate market for content. This suggests that the most likely long-term outcome for the AI industry will be a shift toward contractual licensing models rather than a total reliance on "fair use" as a catch-all defense.

What to Expect Next

The legal battles involving the New York Times, various authors, and record labels against major AI developers are still pending. These cases involve generative models that are vastly different from the specialized, retrieval-based tool built by Ross Intelligence.

1. Fact-Dependent Outcomes: Future courts will likely continue to parse the specific "purpose" of the AI. General-purpose models will face different scrutiny than industry-specific tools.
2. The Licensing Trend: As the "market harm" argument gains traction, AI companies may find that proactively licensing data is a cheaper insurance policy than risking years of litigation.
3. Long-term Uncertainty: We are in the early stages of a judicial cycle that could last decades. This case is but one chapter in a much larger, ongoing dialogue between the law and technology.

Final Thoughts: A Measured Victory

While creators and publishers may feel heartened by the Third Circuit’s decision, it is imperative to resist the urge to declare this a "silver bullet" for the copyright community. The Thomson Reuters case was built on a unique set of facts: a defunct company, a direct market competitor, and a specific, human-authored dataset.

The broader AI landscape, populated by models that digest the sum of human knowledge to produce emergent content, remains a distinct legal territory. For now, the takeaway is clear: if you are building an AI that aims to directly cannibalize the business model of the data you train on, the courts are likely to stand with the original rights holders. Beyond that, the future of AI and copyright remains an open, evolving, and deeply complex question.

As the legal community continues to digest this opinion, one thing is certain: the era of "move fast and break things" in the data training space is coming to a close, replaced by a new era of rigorous judicial oversight and, eventually, a standardized framework for digital intellectual property.