Federal Court Rules Against Ross Intelligence in Pivotal AI Copyright Dispute with Thomson Reuters

Federal Court Rules Against Ross Intelligence in Pivotal AI Copyright Dispute with Thomson Reuters
Photo by Steve Johnson on Unsplash

In a landmark ruling that reshapes the contours of copyright law in the age of artificial intelligence, the U.S. District Court for the District of Delaware granted partial summary judgment in favor of Thomson Reuters, holding that Ross Intelligence infringed upon the copyrights in Westlaw’s headnotes and organizational system when it used them to train its AI-powered legal research tool.

The decision, issued February 11 by U.S. Circuit Judge Stephanos Bibas sitting by designation, revises a prior 2023 ruling and delivers a significant win to traditional legal publishers grappling with the rise of generative and search-based AI in the legal industry.

The case, Thomson Reuters Enterprise Centre GmbH et al. v. Ross Intelligence Inc., centers on the use of copyrighted editorial content from Westlaw—a subscription-based legal research platform—by Ross Intelligence, a startup aiming to disrupt the market with an AI-powered search engine.

After Thomson Reuters denied Ross’s request to license Westlaw content, Ross partnered with LegalEase, a third-party contractor, to obtain “Bulk Memos”—compilations of legal questions and answers developed using Westlaw’s headnotes as source material. Ross then used these memos to train its AI model. In 2020, Thomson Reuters filed suit, alleging direct and indirect copyright infringement.

In 2023, Judge Bibas had declined to resolve the copyright claims at summary judgment. But ahead of a planned 2024 trial, he revisited the case and ultimately revised his earlier view, concluding that Ross’s conduct went beyond what copyright law permits.

Core Holding: Ross Copied Thousands of Westlaw Headnotes

The court held that Ross infringed the copyrights of 2,243 individual headnotes from Westlaw, finding that these editorial summaries constitute original, protectable expressions under the Copyright Act. Although judicial opinions themselves are not copyrightable, Westlaw’s headnotes, created through editorial selection and synthesis, met the low threshold for originality established in Feist Publications v. Rural Telephone Service Co.

Judge Bibas emphasized that even headnotes that quote opinions verbatim can still qualify for protection due to the editorial judgment involved in selecting what language to excerpt. “A headnote is a short, key point of law chiseled out of a lengthy judicial opinion,” Bibas wrote, analogizing the process to a sculptor choosing what to carve from marble.

The court also found that Westlaw’s Key Number System—a taxonomy for organizing legal topics—was sufficiently original to merit copyright protection.

Fair Use Defense Rejected

Ross had asserted fair use as an affirmative defense, arguing that its copying was both transformative and necessary to create a novel AI tool. The court disagreed.

Under the four-factor fair use test, Judge Bibas found that the first and fourth factors—purpose and character of the use, and market effect—weighed heavily against Ross. He reasoned that Ross's use was commercial and not transformative, as it simply used Westlaw's headnotes to train a competing product.

“Ross took the headnotes to make it easier to develop a competing legal research tool,” Bibas wrote, adding that the public has no right to Thomson Reuters’s “parsing of the law.”

While Ross’s use was intermediate—it did not display the headnotes to end users—the court distinguished the case from precedents involving functional computer code, such as Google v. Oracle, noting that copying was not necessary to access underlying legal ideas.

Factors two and three—nature and amount of the work used—tilted slightly in Ross’s favor but were ultimately outweighed. Notably, the court emphasized that even if Ross’s AI never surfaced Westlaw headnotes to users, the act of using them for model training still implicated significant market harm.

Additional Defenses Fall Flat

The court also rejected Ross’s alternative defenses, including innocent infringement, copyright misuse, merger, and scènes à faire. The merger doctrine—which excludes protection when ideas and expression are inseparable—was deemed inapplicable because many editorial choices exist in summarizing legal holdings.

Similarly, the scènes à faire defense, which denies protection to standard elements dictated by subject matter, was found irrelevant. “Nothing about a judicial opinion requires it to be slimmed down to Thomson Reuters’s headnotes or categorized by key numbers,” the court held.

This ruling carries weighty implications for developers of AI tools, especially those using third-party content to train large language models or domain-specific AI systems. The opinion suggests that using editorial legal content—even at intermediate stages of model training—can constitute copyright infringement if done without permission.

Crucially, Judge Bibas limited his ruling to non-generative AI. Still, the logic of the decision could apply broadly to training datasets, putting increased pressure on AI developers to secure licenses or use public domain content.

The court left unresolved certain factual questions, including whether some of the infringed headnotes had fallen into the public domain. Those issues will proceed to trial. However, the legal foundation has now been firmly laid in Thomson Reuters’s favor.

Looking Ahead

Ross Intelligence, which has been defunct since 2021, may appeal the decision. Meanwhile, legal publishers and AI companies are likely to scrutinize their own data pipelines in light of this opinion.

In an era when AI tools increasingly shape legal research and practice, this case may prove to be a bellwether for how copyright law governs the frontier of algorithmic innovation.