U.S. District Judge Stephanos Bibas ruled in Delaware on February 11, 2025, that Ross Intelligence could not claim fair use for its use of copyrighted Westlaw editorial material to develop a competing AI-powered legal-research tool. The decision in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. concerned legal headnotes, not a general prohibition on training artificial intelligence with copyrighted content.
That distinction matters as the ruling circulates under the broader claim that training AI on someone else’s editorial work is not fair use. The court rejected that defense on the facts before it, but copyright law still requires courts to examine the particular material, purpose and market involved.
What Ross used—and why it mattered
Westlaw, owned by Thomson Reuters, provides access to court decisions alongside editorial tools that help lawyers find and understand relevant cases. Those additions include headnotes: short summaries of legal points prepared through editorial work.
The dispute was not about whether a company could read judicial opinions or analyze the law. Judicial opinions themselves are generally outside copyright protection in the United States. The central issue was the protectable expression in Westlaw’s editorial additions.
Ross was developing a legal-research service that used artificial intelligence to help answer users’ legal questions. After unsuccessfully seeking a license from Thomson Reuters, Ross obtained training material through a contractor, LegalEase Solutions, which produced documents known as bulk memos.
Those memos contained legal questions and answers derived from Westlaw headnotes. The court found infringement involving 2,243 headnotes and rejected Ross’s argument that its use qualified as fair use.
The distinction between underlying legal information and an editor’s expression is essential. Copyright does not give a publisher ownership of legal principles. It can, however, protect original language used to summarize or explain those principles.
Why the fair-use defense failed
Fair use permits some uses of copyrighted material without permission. Courts evaluate four statutory factors rather than applying a universal exemption for research, software development or machine learning:
- The purpose and character of the use, including its commercial nature.
- The nature of the copyrighted work.
- The amount and substantiality of the material used.
- The effect on the work’s actual or potential market.
Bibas concluded that the first and fourth factors favored Thomson Reuters, while the second and third favored Ross. The analysis was not a simple two-to-two tie: courts weigh the factors together, and the judge treated market effect as particularly important.
A competing product, not an unrelated use
On the first factor, the court focused on Ross’s commercial purpose and the relationship between the two products. Ross used the material to build a legal-research service that would compete with Westlaw.
The use was not sufficiently transformative merely because copyrighted text became an input in a technical development process. In the court’s analysis, the headnotes helped create a product serving a closely related legal-research purpose.
That reasoning challenges a broad assumption sometimes made about AI: that processing a work through an algorithm automatically changes its purpose enough to establish fair use. The ruling shows why the intended function of the resulting product still matters.
The market extended beyond visible copying
The fourth factor addressed competition with Westlaw and the potential market for licensing its material as AI training data. The court considered both relevant to the commercial consequences of Ross’s use.
This makes the case significant even though the dispute was not centered on a chatbot reproducing headnotes to users. Copyright questions can arise from copying during development, not only from text displayed in a finished product.
At the same time, the decision does not establish that every claimed training-data licensing market defeats fair use. That assessment remains tied to the evidence and circumstances of each dispute.
What the ruling does not decide
Bibas expressly distinguished the technology before him from generative AI. Ross’s system was designed for legal research, rather than to generate new expressive material in the manner associated with modern conversational models.
That limits how confidently the ruling can be applied to disputes involving chatbots, image generators or other systems trained on different collections for different purposes. Similar legal questions may arise, but the factual comparisons are not interchangeable.
The ruling also did not hold that every element of a legal publisher’s product is protected. Copyrightability and fair use are separate questions: a claimant must establish relevant rights before a court decides whether an otherwise infringing use is excused.
Nor should a federal district court decision be described as a nationwide appellate rule settling AI copyright law. Its reasoning can influence other courts without automatically determining their conclusions.
The practical stakes for publishers and developers
For publishers, the decision illustrates how original editorial work can retain legal significance when used as development material rather than republished directly. Summaries and annotations are not necessarily free inputs simply because they describe publicly accessible facts or documents.
For developers, it underscores the importance of knowing where training material comes from, what rights attach to it and how the finished service relates to the source’s market. Obtaining data through a contractor does not eliminate those questions.
The central takeaway is narrower—and more useful—than the viral headline: in this case, using protected editorial material to build a competing legal-research system did not qualify as fair use. Whether another AI training project does will require its own legal and factual analysis.
