Calling a use of copyrighted material “AI training” does not automatically make it lawful. A new US appeals ruling has made that point in a closely watched dispute—but reading it as a verdict on every generative AI model would go too far.
In an opinion made public on 30 September 2026, the US Court of Appeals for the Third Circuit affirmed Thomson Reuters’ copyright victory over Ross Intelligence concerning 2,243 Westlaw headnotes used in training a competing legal-search system.
The material was more than raw legal text
Headnotes are editorial summaries of legal points in court opinions. The distinction matters: the case concerns copyright in that authored material, not private ownership of the underlying law. The court found the headnotes sufficiently original for protection and rejected Ross’s fair-use defence.
Reuters’ 30 September report explains the significance of the unsealed reasoning. It adds detail to the appeals outcome, rather than announcing a new copyright statute for AI.
Copyright distinguishes facts from protected expression. A training pipeline does not itself answer whether a particular use is permitted.
Fair use depends on the particular use
The US Copyright Office’s explanation sets out four factors: the purpose of the use, the nature of the protected work, the amount taken and the effect on its market. Courts consider these together and evaluate cases individually. A commercial use is not automatically unlawful, and a research label is not automatically decisive.
The appeals court emphasised that Ross used the headnotes to create a commercial substitute for Westlaw serving a closely related purpose. It considered the use minimally transformative. “Transformative” refers to a sufficiently different purpose or character, not merely switching to a newer technical process.
Ross’s system was not generative AI. It retrieved passages from existing judicial opinions rather than creating new expression. The court explicitly distinguished that situation from arguments being made in other litigation about generative models. That is a substantial limit on what this decision establishes.
The larger dispute remains open
The policy tension is real. Publishers and other creators want control over commercially valuable work; technology developers want workable ways to build useful systems from information. A world in which every permission is expensive could favour companies with large licensing budgets. A world in which no permission is needed could undermine the people producing the material.
Those are competing policy concerns, not findings that this court resolved for the entire AI industry. Different training sources, product functions and market effects can produce different legal questions.
Our assessment is that this ruling makes sweeping slogans less useful. The relevant questions are what was copied, what rights attach to it, why it was used and what product the copying helped create.
The decision is consequential precisely because it is concrete. It should sharpen the debate over AI training without being stretched into a universal answer.
Featured image: A library interior, shown as representative research context; it does not depict Westlaw or Ross Intelligence. Photo: Thomas Bormans / Unsplash.
Leave a Reply