Fair Use in AI Copyright Litigation: A Surprising Turn in Thomson Reuters v. Ross

From the pages of The New York Times to the…general counsel’s office of The New York Times, AI copyright litigation is all the rage. Possible questions include the philosophical—e.g. “Could an AI agent hold a copyright?”—but the most asked are the pecuniary—e.g. “Do the creators of AI tools owe me, a copyright holder, money?”

Unsurprising for the newness of the technology, the legal landscape in this area defies confident prediction. Federal district courts across the land are reading and re-reading two recent and potentially relevant Supreme Court cases, Google LLC v. Oracle America, Inc. and Andy Warhol Foundation for Visual Arts, Inc. v. Goldsmith, to predict how the Supreme Court will eventually weigh in on the AI copyright disputes.

Notable defendants include a veritable “who’s who” of developers of AI tools.

• Github faces contentions in Doe v. Github that it breached software licenses of anonymous open-source contributors in training its GitHub Copilot.
• Google is alleged to have infringed copyrights by scraping web material on which to train Gemini in class actions consolidated as In re Google Generative AI Copyright Litigation.
• NVIDIA is undergoing discovery in Nazemian and Dubus v. NVIDIA Corporation with plaintiffs alleging copyright infringement in the training of Nemo Megatron-GPT.
• OpenAI faces numerous lawsuits relating to the development and use of its ChatGPT models.
• Meta has been sued in Kadrey v. Meta by multiple authors who allege that Llama was impermissibly and inappropriately trained on their writing.
• And of course, Microsoft and OpenAI are alleged in The New York Times Company v. Microsoft to have infringed the paper’s copyrights in training ChatGPT and Copilot.

In these cases and others like them, defendants may rely on the “fair use” defense. Fair use analysis calls for balancing the factors of the purpose of the use, the nature of the copyrighted work, the amount of the work used, and effect of the use on the market for the work. Traditionally, uses which are “transformative”—e.g. uses “altering the original with new expression, meaning, or message“ per the Supreme Court in Campbell v. Acuff-Rose—are more likely to be considered fair.

Against this backdrop, this week brought a development in what is perhaps the first-filed AI copyright case: Thomson Reuters v. Ross. In 2020 Thomson Reuters, proprietor of legal-research database Westlaw, sued Ross in the District of Delaware alleging that Ross’s training of AI models on Westlaw’s “headnotes” to construct an AI-powered competitor to Westlaw constituted copyright infringement.

Among other things, Ross raises a fair use defense. Ross’s fair use argument was bolstered by the contention that, because Westlaw’s headnotes are primarily language pulled from non-copyrighted judicial opinions, they are not sufficiently original as to merit copyright protection.

For a time, Judge Stephanos Bibas gave credence to these arguments. In 2023, Judge Bibas denied motions for summary judgment and ruled that issues of fair use and copyright infringement, as well as the underlying issue of headnote originality, would have to be decided by a jury at trial.

However, in 2024, Judge Bibas indicated that he was reconsidering his own motion. And earlier this week, what was telegraphed came to pass: Judge Bibas issued a revised summary judgment order for Thomson Reuters finding no fair use and direct copyright infringement of a large portion of Westlaw’s headnotes.

In Judge Bibas’s telling, his mind was changed on originality as he came to see the Westlaw headnotes as copyrightable “sculpture” carved from the “raw marble” of judicial opinions.

Reconsidering fair use as well, Judge Bibas found it particularly weighty that Ross’s use of the Westlaw headnotes—to train a competing platform competing with Westlaw in the marketplace—was commercial in purpose and likely to negatively impact the market for Westlaw’s headnotes.

Finding Ross to have committed copyright infringement on summary judgment is a surprising result for reasons well beyond the self-reversal. Copyright infringement could be alleged to occur at different stages in the deployment of AI tools, including: scraping content to obtain training data, converting training data into tokens, adjusting model weights, and providing output. The alleged infringement in this case differs from that in certain other AI copyright suits. Here the infringing act is not outputting Westlaw headnotes, but is instead manually copying data used for training Ross’s model in its product, which provides answers to user-generated questions in a manner analogous to a legal scholar skilled at using Westlaw headnotes to point inquirers to relevant cases. That Digital Millennium Copyright Act claims have been dropped in many suits for lack of direct reproduction heightens the salience of Judge Bibas’s finding of infringement through provision of training data.

Judge Bibas’s opinion is surely of interest to those invested in analogous litigation elsewhere. For instance, music publishers in litigation with Anthropic filed a motion seeking to have the ruling considered as a supplemental authority in Concord Music Group, Inc. v. Anthropic PBC less than 24 hours after it was handed down.

To litigators representing AI companies who felt a sudden chill in the air on February 11, Judge Bibas offers a light blanket. The opinion notes with intentionality that Ross’s tool was “non-generative” AI intended to function as a direct competitor to Westlaw and was therefore substantially non-transformative in fair use analysis. This provides a clear distinction for many cases. And of course, a summary judgment order in Delaware is far from nationwide binding precedent.

The intersection of AI and copyright remains a hot topic with much at stake. While the latest ruling in Thomson Reuters v. Ross far from resolves the morass, stakeholders should take note whether it becomes mainstream or an outlier. Certainty in AI copyright litigation may remain chimeric until the technology is old hat and a case or two make their way to the Supreme Court.

Perhaps you’ll read about it in The New York Times.