What happened
The Third Circuit held that Thomson Reuters’ Westlaw headnotes are protectable by copyright and that ROSS’s use of those headnotes to train a non‑ legal‑search engine was not fair use. The court found the use commercial, minimally transformative, and harmful to Westlaw’s licensing market, and sent the case back to the District of Delaware for trial on damages.
In Thomson Reuters v. ROSS, the Third Circuit affirmed a partial summary‑judgment finding that ROSS infringed Thomson Reuters’ copyright by copying roughly 25,000 Westlaw‑written headnotes into internal memoranda used to train its AI legal‑search platform. The court applied the four fair‑use factors, weighing three against ROSS: the commercial, minimally transformative nature of the use; the amount of protected material copied; and the negative impact on Westlaw’s licensing market.
The court emphasized that the headnotes contain original, creative expression beyond mere , making them fully copyrightable. ROSS’s argument that copying was necessary to reach unprotected judicial opinions was rejected; the court noted that the opinions were freely available and that copying the headnotes was a matter of convenience, not necessity.
A footnote distinguished the case from generative‑AI models, referencing DOJ concerns in the OpenAI and Bartz v. Anthropic matters. The court suggested that large language models that generate original text might be viewed as more transformative, but warned that the licensing‑market factor would still be relevant.
The case now returns to the District of Delaware for trial on the remaining claims, including damages and tortious interference.
Source details: natlawreview.com ↗
Why it matters
The decision clarifies that copying copyrighted legal annotations to train AI—even when the AI does not generate original text—does not automatically qualify as transformative fair use. It signals heightened risk for AI developers who rely on proprietary data, underscores the importance of licensing agreements, and hints that generative‑AI models may be judged differently under the first fair‑use factor. The ruling also reinforces the emerging market for licensing headnotes as AI training data, affecting both content owners and AI firms.
The ruling provides one of the first appellate-level analyses of AI training data under U.S. copyright law, offering a concrete precedent for future disputes involving proprietary content.
By rejecting the fair‑use defense for a non‑ tool, the decision may push AI developers to secure licenses for copyrighted datasets or redesign training pipelines to rely on public domain material.
Content owners, especially legal publishers, gain a judicial endorsement of their ability to monetize annotations and other editorial enhancements as a distinct licensing market for AI training.
The footnote’s distinction between non‑generative and introduces a nuanced legal question: whether highly transformative generative models will receive more favorable fair‑use treatment, a question that upcoming generative‑AI cases will likely address.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
Future appellate rulings on generative‑AI training, how courts treat the licensing market factor for highly transformative models, and whether AI developers will shift to exclusively unprotected data sources or negotiate broader licenses.
Appellate outcomes in pending generative‑AI copyright cases, such as the OpenAI matter in the Southern District of New York, which could set a different standard for transformative use.
Potential legislative or regulatory responses that may clarify licensing obligations for AI training data, especially in the legal and scholarly domains.
Industry shifts toward building AI systems on openly licensed or public‑domain corpora to mitigate litigation risk.
Monitoring how Westlaw and other legal information providers expand licensing programs for AI developers.