No fair use in U.S. copyright for training an AI model?

On Feb. 11, 2025, the Delaware court ruled that using copyrighted material to train an AI model does not constitute "fair use" (Case 1:20-cv-00613, ECF 770).

The bottom line

The plaintiff was Thomson Reuters, owner of the legal platform Westlaw. Ross Intelligence - a new player in the legal search technology market - wanted to train its AI system with legal questions and answers, and to do so bought some 25,000 "Bulk Memos" from the company LegalEase. These memos were based on headnotes from Westlaw, which contain summaries of legal rulings.

The court had to consider three main questions:

1. Are the headnotes original and therefore copyrighted?
2. Was there infringement?
3. Can Ross invoke the fair use defense?

Original work and infringement

The court confirmed that the headnotes are original. The threshold for originality is very low in the US: a "creative spark" is sufficient. The editorial work in preparing a headnote - summarizing and synthesizing court decisions - was considered sufficiently creative.

The court then found that there was actual copy and substantial similarity for at least 2,243 of the 2,830 headnotes examined. The content of the memos differed little from the Westlaw versions.

No fair use

Fair use is an important open standard in U.S. copyright law under which a particular use of material can be defended without permission.

However, the court rejected the fair use claim. Key considerations:

• Commercial character: The use was commercial and gave Ross a direct competitive advantage....
• Non-transformative use: The headnotes were not reworked into something new, but merely served as training material.
• Market damage: The use caused Westlaw financial harm by reducing licensing revenue.

And Europe?

The European (Belgian) copyright has no such open defense standard and only an exhaustive list of strictly delineated copyright exceptions. One of the exceptions to copyright that can be invoked for training AI models is that of TDM.

Along these lines, the Hamburg court ruled in a Sept. 27, 2024 (Laion) that training an AI model for scientific purposes fell under the TDM exception. For other purposes (such as commercial purposes), rights holders can, however, prevent such use via an opt-out.

So the irony is that, despite a broader fair use exception than the European exceptions, U.S. copyright seems stricter for training AI models than European copyright.