AI, Copyright, and Fair Use
AI, Copyright, and Fair Use
Generative AI raises many legal issues. One recent one is the issue as to whether copyrighted works be used to train AI systems without permission. Currently, U.S. law is developing through U.S. Copyright Office guidance and court decisions, but it appears resolution on this matter are dependent on the facts of each case.
How Generative AI learns
Developers train AI models on collections of examples, which may include text, images, or other digital material. For a language model, text is divided into small units called tokens. The model practices predicting what token is likely to come next, compares its predictions with the examples, and adjusts its internal numerical settings. Repeating this process helps the model recognize patterns and generate responses to new prompts.
A trained model is not simply a searchable copy of its training collection, but it may sometimes reproduce or closely resemble material it encountered. Collecting and preparing training materials can involve copying books, articles, images, and other works—some of which may be protected by copyright.
Copyright and the Fair Use Doctrine
Federal copyright protects original works of authorship. The Fair Use Doctrine is a legal limitation on copyright owners’ exclusive rights; this includes such purposes as criticism, comment, news reporting, teaching, scholarship, and research that are outside copyright protection. However, this is not absolute. Courts still weigh four factors: the purpose and character of the use, the nature of the work, the amount used, and the effect on the work’s market or value. Notably, no one factor automatically resolved this issue. See 17 U.S.C. § 107.
AI developers may argue that training changes a work’s purpose and is therefore transformative. Copyright owners may argue that using their works without permission harms existing or potential markets, including markets for licensing works as training data. Courts assess those arguments under the facts and evidence in each case.
Court Decisions In Bartz v. Anthropic PBC, a federal district court ruled that Anthropic’s use of books to train its language model was fair use under the circumstances before it. The court separately ruled that keeping pirated digital books in a general library was not fair use. Claims concerning pirated books were later resolved through a court-approved settlement; the settlement did not establish a general rule that AI training is fair use. In Kadrey v. Meta Platforms, Inc., another federal district court ruled for Meta on the training claims based on the evidence presented. The judge stressed that different evidence, particularly evidence of harm to authors’ markets, could lead to a different result. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., the Third Circuit held that Ross’s copying of Westlaw editorial headnotes to train a competing legal research product was not fair use. The court emphasized the similarity in purpose and the potential harm to Thomson Reuters’s market. This ruling concerned specific works and uses; it does not resolve every dispute over AI training.
These decisions show that courts are applying existing copyright law to different training methods, sources, and markets. No single ruling answers every question about AI and copyright.
This article is for educational purposes only and is not legal advice. Legal developments may change.
Further reading · U.S. Copyright Office: Copyright and Artificial Intelligence · Copyright Office Part 3 report: Generative AI Training (PDF)
Case citations: Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025); Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025); Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153, 2026 WL 2932557 (3d Cir. Sept. 29, 2026).
