Third Circuit affirms Westlaw AI-training win, with reasoning sealed

The story

The U.S. Court of Appeals for the Third Circuit has upheld Thomson Reuters' copyright victory against Ross Intelligence in a dispute over material used to develop an AI-powered legal-research system. Reuters reported on September 29 that the Philadelphia-based court rejected Ross's fair-use position, making this the first U.S. appellate result in a copyright case centered on AI training. The most important qualification is unusually large: the appellate court's reasoning remains sealed, so the legal path to the result is not yet public.

The dispute began with Westlaw, Thomson Reuters' legal-research platform. Court opinions themselves are public law, but Westlaw adds editor-created headnotes that summarize legal propositions and connect readers to relevant passages. Thomson Reuters alleged that Ross obtained thousands of those headnotes through an intermediary and used them as training material for a system designed to return relevant judicial opinions in response to natural-language questions. Ross argued that this use was fair because its system did not display the headnotes to users and served a different function.

In February 2025, Judge Stephanos Bibas of the U.S. District Court for the District of Delaware rejected that defense on summary judgment. His official opinion concluded that Ross used the headnotes to make development of a competing legal-research tool easier and that the use was not transformative. The court placed particular weight on the purpose of the use and its effect on an existing or potential market for the copied material. It did not rule that the underlying judicial opinions belonged to Thomson Reuters; the copyright claim concerned the company's editorial headnotes and related selection work.

The Third Circuit has now affirmed the result, but a sealed opinion is a poor foundation for broad doctrine. Until a public version appears, readers cannot know whether the panel adopted all of the district court's analysis, relied on a narrower ground or modified parts of the fair-use framework. The judgment matters to the parties immediately. Its value as precedent for developers, publishers and lower courts will depend on what the eventual public text actually says.

The factual boundaries matter as much as the procedural ones. Ross's product was a retrieval-oriented legal-search engine, not a generative model that produces new text, images, audio or code. Reuters, whose parent company is Thomson Reuters, noted that distinction in its report. Ross also sought to compete in the same legal-research market from which the source material came. Those facts may sharpen the market-harm argument in ways that do not map neatly onto every foundation model or every training dataset.

Even so, the outcome enters an industry debate with enormous economic stakes. AI developers often argue that training transforms source material into statistical capabilities rather than substituting for the works themselves. Copyright owners counter that unlicensed ingestion can reproduce protected expression, create competing products and undermine licensing markets. Fair use is fact-specific, balancing purpose and character, the nature of the work, the amount used and market effects. One appellate decision can influence that debate without resolving it across different technologies and markets.

The ruling also highlights a design choice for AI businesses: data provenance can become a product risk, not merely a compliance task. A company building on editorial summaries, annotated databases or other value-added collections must distinguish public-domain inputs from protected layers created around them. Contracts, licenses, source tracking and auditable training records can affect whether a system survives litigation long enough to reach the market. Ross shut down its platform in 2021, citing the cost of the lawsuit, illustrating how uncertainty itself can be commercially decisive.

INNOVOX analysis: the strongest immediate signal is about direct competition, not AI in the abstract. A developer that uses a publisher's curated material to accelerate a substitute product faces a harder fair-use story than one using heterogeneous works for a capability far removed from the source market. But the sealed reasoning prevents a reliable statement about where the Third Circuit drew that line. Treating the judgment as either a blanket ban on unlicensed AI training or a narrow anomaly would move beyond the evidence now available.

What to watch next is the public record. An unsealed or redacted opinion should reveal how the panel treated transformation, intermediate copying and potential licensing markets. Ross could seek rehearing or ask the Supreme Court to review the case, while other courts may decide whether the reasoning applies to generative systems. For innovators and rights holders, the practical lesson is already clearer: document the origin and legal status of training inputs, evaluate whether the resulting product competes with the source, and do not assume that hidden intermediate use is automatically fair.

INNOVOX analysis

The result strengthens the negotiating position of publishers whose editorial content is used to develop a directly competing product. Its wider force depends on the still-sealed reasoning: market substitution, the creative choices embedded in Westlaw headnotes and Ross's non-generative system may make this dispute materially different from other AI cases.

What to watch

Watch for an unsealed or redacted Third Circuit opinion, any petition for rehearing or Supreme Court review, and how other courts distinguish this competitive legal-search system from generative models trained on broader corpora.