This is an informal case summary prepared for the purposes of facilitating exchange during the 2026 WIPO IP Judges Forum.
Session 3: Copyright and AI Training
United States District Court for the Northern District of California [2025]: Kadrey v. Meta Platforms, Inc., 788 F.Supp.3d 1026
Date of judgment: June 25, 2025
Issuing authority: United States District Court for the Northern District of California
Level of the issuing authority: First Instance
Type of procedure: Judicial (Civil)
Subject matter: Copyright and Related Rights (Neighboring Rights)
Plaintiff/Appellant: Richard Kadrey and others (thirteen authors total)
Defendant/Respondent: Meta Platforms, Inc.
Keywords: Artificial intelligence (AI); Fair use; Large-language-model training; Transformative use; Market dilution; Licensing market; Copyright reproduction; Summary judgment
Basic facts: Thirteen authors, principally writers of fiction, nonfiction, memoir, and plays, brought a putative class action against Meta Platforms, Inc, alleging that Meta reproduced their copyrighted books without permission in order to train its Llama family of large language models (LLMs).
Meta initially explored obtaining licenses for books to be used as training material. Its efforts encountered practical and legal obstacles: publishers did not necessarily hold the relevant AI-training rights; those rights could be held by individual authors; rights could be territorially fragmented; no established collective-licensing mechanism existed for the use; some publishers did not respond; and only one publisher made a pricing proposal. Meta initially downloaded the Library Genesis (LibGen) dataset in October 2022 to assess its potential value for Llama training. After licensing discussions had not produced a workable arrangement, Meta decided in spring 2023 to use LibGen materials for training and discontinued its licensing efforts after determining that LibGen contained most of the works available from certain publishers with which Meta had been negotiating.
Meta also downloaded material from Anna’s Archive, a compilation drawing on shadow libraries, in early 2024. Meta acquired the datasets using BitTorrent. The parties disputed whether, and to what extent, Meta also uploaded or redistributed copyrighted material through the BitTorrent process. The court did not resolve that distinct alleged distribution claim, which was not the subject of the cross-motions for summary judgment.
The parties cross-moved for partial summary judgment on fair use. The plaintiffs argued that Meta’s copying could not realistically be fair use; Meta argued that the copying of the thirteen named plaintiffs’ books for Llama training was fair use as a matter of law.
The plaintiffs advanced two principal theories of market harm. First, they argued that Llama could reproduce portions of their books. The record, however, showed that even adversarial prompting could not elicit more than approximately 50 words and punctuation marks from any plaintiff’s work, and plaintiffs’ own expert accepted that Llama could not reproduce a significant percentage of any of the books. Second, the plaintiffs argued that unlicensed training impaired their ability to license their works for use as AI-training data. They also advanced a broader theory, developed only minimally on the evidence, that Llama would dilute the market for their works by enabling the production of large quantities of competing works with similar subject matter or genre.
Held: The court denied the plaintiffs’ motion for partial summary judgment and granted Meta’s cross-motion for partial summary judgment on fair use. On the record before it, Meta’s reproduction of the thirteen named plaintiffs’ books for use in training the Llama models constituted fair use under section 107 of the Copyright Act.
The first factor, concerning the purpose and character of the use, strongly favored Meta because Llama training was highly transformative. The second factor, concerning the expressive nature of the plaintiffs’ books, favored the plaintiffs but carried limited weight. The third factor favored Meta because copying entire books was reasonable in relation to the transformative purpose of training a high-quality LLM. The fourth factor, which the court considered “undoubtedly the single most important element of fair use,” favored Meta on the evidentiary record presented.
The ruling was narrow. It resolved the fair-use defense to the reproduction claim as it concerned the thirteen named plaintiffs’ works. It did not decide whether Meta had distributed copyrighted works through BitTorrent leeching or seeding, did not adjudicate the claims of the proposed class as a whole, and did not establish that Meta’s use of all copyrighted works to train Llama was generally lawful.
Relevant holdings in relation to copyright and AI training: The court began by identifying the broad question raised by the case: whether it is unlawful to use copyright-protected material to train generative AI models without authorization or payment. It indicated that, in most cases, the answer would likely be yes. Copyright law is intended to preserve incentives for human authorship, and a fair-use defense is unlikely to succeed where unlicensed training significantly diminishes rightsholders’ ability to derive economic value from their works.
The court nevertheless stressed that the question had to be decided on the actual record, rather than on generalized assumptions about the potential effects of generative AI. Fair use is a flexible and holistic inquiry, not a mechanical tally of four factors. Its central concern is whether the secondary use is likely to serve as, or facilitate, a market substitute for the copyrighted work and thereby undermine incentives to create. Fair use is an affirmative defence, and the party asserting it bears the burden of establishing the defense as a whole rather than necessarily prevailing independently under every factor.
Purpose and character of the use
The court held that Meta’s use was highly transformative. Meta copied the books to train models capable of generating diverse text and performing a wide range of functions, rather than to enable users to read the books for their original purposes of entertainment, education, or information.
The court rejected the analogy between model training and a person reading a book. An LLM does not read in the ordinary human sense: it processes text through repeated prediction tasks, including removing words, predicting them from context, and updating its internal statistical representations. Nor was Llama equivalent to a professor providing a book to a single student. Meta had created a tool available to a broad public that could potentially generate expression on a vast scale.
The plaintiffs’ argument that Llama could mimic their literary styles did not alter the analysis. Copyright protects expression, not style. Moreover, the evidence did not show that Llama could reproduce substantial portions of the plaintiffs’ works. Even under adversarial prompting, the plaintiffs had not established that Llama could produce more than about 50 words and punctuation marks from any particular book.
The court considered the plaintiffs’ contention that Meta’s downloading from shadow libraries had to be assessed wholly separately from the subsequent training of Llama. It rejected that contention on the record before it. While acknowledging that the downloading was a different use from the copying done in the course of training, the court held that the purpose of the downloads should be assessed in light of their ultimate use in Llama training, which it found highly transformative. This approach differs from that taken in Bartz v Anthropic PBC, where the court had treated the acquisition and retention of pirated library copies as a use distinct from specific LLM training, although the Kadrey court did not expressly address Bartz on this point; its only reference to Bartz concerns the treatment of market harm.
However, the court accepted that Meta’s downloading from shadow libraries, including its use of BitTorrent, could bear on the character of its conduct in two ways: as evidence of bad faith, and if it had benefited the operators of the shadow libraries and thereby supported and perpetuated their unauthorized copying and distribution. Bad faith, even if relevant, did not “move the needle” given the rest of the summary-judgment record, and the plaintiffs had produced no evidence that Meta’s torrenting had in fact benefited the shadow libraries. The separate distribution claim remained unresolved.
The fact that Meta was a commercial enterprise was relevant, but did not outweigh the strongly transformative character of the training use. Nor did the existence of downloaded books not ultimately used in training defeat fair use. The plaintiffs had offered no evidence that Meta had in fact downloaded copies that were never used for training, and fair use does not require a secondary user to make the lowest conceivable number of copies.
Nature of the works
The second factor favored the plaintiffs. Their books were highly expressive works, and model training relied on creative elements of expression, including word choice, word order, grammar, syntax, coherent structure, and style.
The court therefore rejected Meta’s reliance on intermediate-copying authorities such as Sega Enterprises Ltd v Accolade, Inc and Sony Computer Entertainment, Inc v Connectix Corp. Those cases concerned copying software to access unprotected functional elements or interfaces. By contrast, the quality of an LLM’s outputs depends in material part upon the creative and expressive qualities of its training material.
The court also distinguished Authors Guild v Google, Inc. Google Books’ search database was substantially content-agnostic: it enabled search and location of terms regardless of a particular work’s literary or expressive quality. LLM training differs because the quality of the model’s output depends on the quality of the language and expression present in the training corpus.
The factor carried limited weight, however, particularly because the works were published.
Amount and substantiality
The third factor favored Meta. Although Meta copied the plaintiffs’ books in their entirety, full copying was reasonable in relation to the highly transformative purpose of training a high-quality LLM. The court regarded the factor as substantially overlapping with the first factor: the permissible amount of copying depends upon the secondary use’s purpose and character.
The court did not require Meta to demonstrate that every individual book, or every copy, was strictly indispensable. It was enough that using complete works was reasonably related to the development of an LLM capable of producing the asserted transformative functions.
Market effect
The court described the fourth factor as “undoubtedly the single most important element of fair use.” It disagreed with the implication in Bartz that a sufficiently transformative purpose could make displacement caused by generative AI irrelevant. A use may be highly transformative and still fail as fair use if it substantially harms the market for the original works or materially diminishes the incentive to create them.
The court identified three possible forms of market harm from generative-AI training.
First, regurgitation or direct substitution. This theory failed because the evidence did not show that Llama could reproduce enough of any plaintiff’s book to enable users to access or read the book through the model, or to obtain a meaningful substitute for it. The isolated fragments that the model might generate did not amount to market substitution.
Second, loss of a licensing market for AI training. This theory also failed. The court held that rightsholders could not establish cognizable market harm merely by identifying a potential market for licensing works for a use the court had found transformative. Nor could the asserted loss of individual book sales to Meta itself determine the analysis. The fact that Meta might otherwise have purchased particular books did not establish market harm for purposes of fair use, because the acquisition was assessed in relation to the ultimate transformative use.
Third, market dilution through non-infringing competition. The court regarded this as the most plausible theory. Generative AI could enable the production of enormous quantities of works similar in genre, subject matter, or audience appeal to human-created works, at a fraction of the time and creative effort normally required. Such output might reduce demand for the original works and depress the economic incentives that copyright seeks to preserve, even if the AI outputs did not infringe by reproducing protected expression.
The plaintiffs, however, did not sufficiently develop this theory. They provided little evidence concerning Llama’s current or expected outputs, how those outputs would compete with their particular works, or how they would dilute the market for those works. The theory therefore did not create a genuine dispute of material fact sufficient to defeat Meta’s motion for summary judgment.
The public benefits associated with Llama’s capacity to perform diverse functions provided some additional support for fair use on this record. The court did not suggest that public utility would override substantial market harm; rather, in the absence of adequately supported market harm and in light of the highly transformative use, those benefits modestly supported Meta.
Finally, the court rejected the argument that a finding against fair use would necessarily halt generative-AI development. A failed fair-use defense would ordinarily mean that a developer must obtain licenses or pay for the relevant uses. The court suggested that licensing markets could emerge, with publishers negotiating the relevant subsidiary rights with authors.
Relevant legislation: United States Code, Title 17 – Copyrights (US455)