Brett Carmody Discusses Application of Fair Use for AI Development in Law360 Article
In order to develop artificial intelligence large language models, which produce a written response when given a prompt by a user, companies have acquired numerous copies of written works to use in training the AI, raising questions in copyright law.
Once authors learned about the use of their books, they began bringing litigation against AI developers for using copies of the books without the authors' permission.
To date, AI developers largely have not contested the assertion that making and using the copies of the books is, technically, infringement; instead, they have argued that they are not liable because the fair use defense applies.
Although the cases have been pending for a few years now, case decisions have been rare on whether the use of copyrighted books for training generative AI is indeed fair use.
Two recent decisions from the U.S. District Court for the Northern District of California — Bartz v. Anthropic PBC and Kadrey v. Meta Platforms Inc. — conclude that training AI is fair use, but they include some meaningful differences in reasoning that could affect pending and future cases.
And despite a finding of fair use in one case, Bartz, the decision was followed by a $1.5 billion class settlement, which the court preliminarily approved on Sept. 25.
Fair Use
Copyright law protects creative, original works that are fixed in a tangible medium. Copyright protection provides specific exclusive rights to the copyright owner, including the right to make copies of the work.
Fair use is a defense under copyright law under which a person who infringes on a copyrighted work is not liable if the infringing use is considered fair.
In determining whether the defendant's use is fair, the court considers and weighs four factors: (1) the purpose and character of the defendant's use; (2) the nature of the copyrighted work; (3) the amount and substantiality of the copyrighted work used by the defendant; and (4) the effect of the use on the market for or value of the copyrighted work.
Although all four factors are always considered, courts generally give the first and fourth factors the most weight.
Lawsuits Against AI Developers for Copyright Infringement
So far in the copyright litigation against AI developers, the main theory that has survived the motion to dismiss stage is that the creation and use of copies of copyrighted works to train the AI model is infringement. Courts have also indicated that, if an AI model produces content that is the same or substantially similar to a copyrighted work, that also may constitute infringement.
On June 23 and June 25, respectively, the Northern District of California issued decisions on whether fair use applies in these circumstances in Bartz v. Anthropic PBC and Kadrey v. Meta Platforms Inc.
In both cases, the defendant AI developer acquired copies of the plaintiff authors' books and then made additional copies, without the authors' permission, to use for training AI models.
In Bartz, the defendant Anthropic acquired many of its initial copies of the books from "pirate / shadow libraries," which are digital libraries of works that were acquired without authorization, but also purchased some works from distributors and retailers. In Kadrey, defendant Meta acquired all of its initial copies of the relevant books from unauthorized libraries.
In both cases, the defendants moved for summary judgment, arguing that their use of the unauthorized copies of the plaintiffs' books was fair use.
Bartz v. Anthropic PBC
In addressing fair use in Bartz, U.S. District Judge William Alsup determined that there were three discrete uses or types of copies of the books to be separately analyzed: (1) the copies used to train the AI model; (2) the purchased print copies used to create a digital copy of the work;[1] and (3) the pirated copies used to build a general purpose library of works for Anthropic to keep indefinitely.[2]
The bulk of the court's analysis focused on the first factor under fair use — the purpose and character of Anthropic's use — and analogized the training of AI to a human learning to read and write, finding that such use is highly transformative and therefore favors a finding of fair use.
In addressing the effect of the use on the market and value for the books, the court dismissed the argument that AI models' creation of works that will compete with the authors' books, while not being substantially similar to those books, will affect the books' value, reasoning that the Copyright Act does not protect authors against this type of competition.
Thus, the court concluded that the factors supported a finding of fair use for the copies used to train the AI model.
Importantly, the court's conclusion appears to be based on the assumption that the initial copy acquired by Anthropic was authorized, e.g., purchased from a retailer, and so the court's finding of fair use was arguably limited to unauthorized intermediate copies that were created from the initial authorized copy.
The court determined that it did not need to address whether training AI from copies that were initially pirated would be fair use, in light of Anthropic's general purpose library built from pirated copies. However, the court seemed skeptical that a defendant could show fair use of books, even for a transformative use, if the initial copies were pirated.
As for the copies used to create the library, the court concluded that pirating copies to create a research library for general purposes was not transformative, particularly where the books were available for purchase or loan.
The court also found that the effect on the market for the books was significant, as the pirated copies for the library displaced demand for the books, copy for copy. As a result, the court concluded that the use of pirated copies for the library did not constitute fair use.
On Sept. 5, the parties in Bartz filed a motion for preliminary approval of a class settlement for $1.5 billion, which is expected to equate to $3,000 per work infringed. On Sept. 25, the court granted preliminary approval of the settlement.
Kadrey v. Meta Platforms Inc.
On June 25, two days after the decision in Bartz, U.S. District Judge Vince
Chhabria issued his order in Kadrey on Meta's motion for summary judgment regarding fair use.[3]
At the outset, Judge Chhabria agreed with Judge Alsup in finding that the use of copyrighted works to train AI models is highly transformative, so that the first factor supports fair use.
However, where Judge Alsup apparently limited that finding to intermediate copies created from an initial copy that was properly purchased — and seemed skeptical that fair use could be established when the initial copies were pirated — Judge Chhabria concluded that the fact that Meta pirated the initial copies was not material in this case.
Judge Chhabria reasoned that construing Meta's download of the initial copy as pirating would be begging the question, because the purpose of the fair use analysis is to determine whether a given act of copying is actually unlawful, and the initial downloading of the books must still be considered in light of its ultimate highly transformative purpose: training Llama, which is Meta's family of large language models.
Accordingly, Judge Chhabria found that using the books for training AI was transformative and supported fair use, regardless of whether the initial copies were unauthorized.
Turning to the fourth factor, which Judge Chhabria described as the most important, he explained that there are at least three arguments that a defendant's copying for purposes of training generative AI models harmed the market for the plaintiffs' works:
- The AI model produces identical or substantially similar works;
- Unauthorized copying for AI training harms the market for licensing the plaintiffs' works; and
- Even if the works produced by the AI model are not identical or substantially similar to the plaintiffs' works, they are similar enough that they will compete with the plaintiffs' works and indirectly substitute for them, like a reader looking for a romance novel choosing an AI-generated book instead of the plaintiff's book.
The court quickly rejected the first two arguments, concluding that the evidence showed that Meta's AI model was not capable of producing any meaningful amount of the plaintiffs' works, and the loss of fees from licensing for the defendant's particular use is not a recognized harm for fair use analysis.
The court seemed to find the third argument the most persuasive, despite ultimately ruling against the plaintiff. Judge Chhabria emphasized how an AI model would be able to produce numerous books in significantly less time and using a fraction of the creativity, which could address the same topics as the plaintiffs' books or be of the same genre, thereby diluting the market and indirectly competing with the plaintiffs' works.
The court disagreed that the only competition considered under fair use should be works that actually infringe on the plaintiffs' work; AI-generated works of the same genre that a reader may choose to read instead of the plaintiffs' books still indirectly compete.
The court acknowledged that this type of indirect substitution and level of market dilution is often immaterial in other cases, which involve only a single competing work that would not have the substantial impact required under the fair use analysis.
But in this case, the risk of market dilution is highly relevant, and the fact that it has not been a significant consideration in other cases does not mean it is unimportant.
However, the court concluded that the plaintiffs in Kadrey had failed to submit any evidence to support a finding that their particular works had suffered from market dilution.
The plaintiffs did not produce evidence that the AI model could produce book-length and book-style output, that any such books compete with any of the plaintiffs' works, or that sales of the plaintiffs' works will be affected.
The court opined that, in cases like this, it seems that the plaintiffs will often win, assuming better-developed records on the market effects of the defendant's use. Unfortunately for the plaintiffs in Kadrey, though, the lack of evidence on market impact meant that Meta had established its fair use defense.
Summary and Looking Forward
Following the decisions on fair use in Bartz and Kadrey, there appears to be solid support for concluding that the use of copyrighted works to train AI models is highly transformative and weighs strongly in favor of fair use.
However, while both decisions agree on this point in the case of intermediate copies created from purchased — or otherwise authorized — initial copies of works, there is disagreement on whether that same conclusion applies if the initial copies were acquired without authorization.
The Bartz and Kadrey decisions also disagreed on the impact on the market and value of the works. While the Bartz decision dismissed the risk of market dilution by AI-generated works, the Kadrey decision suggested that the indirect substitution from such widespread dilution would likely result in plaintiffs prevailing in other cases.
As a result, despite the same ultimate conclusion — that training AI is fair use — defendants in pending and future AI copyright cases may still face uncertainty in establishing the defense.
For those that pirated their initial copies of the works, they will want to point to the Kadrey decision and how it downplayed that issue in the analysis, while asking the court to ignore as dicta the discussion of how most such cases should not be fair use.
Plaintiffs, on the other hand, may appreciate the dicta in the Bartz decision questioning whether pirating initial copies of copyrighted works could ever be fair use, but presumably would rely primarily on the Kadrey dicta indicating that, if the plaintiffs can show the impact that AI-generated works have had on the market for their works, that consideration should usually preclude the fair use defense.
Indeed, the $1.5 billion settlement in Bartz indicates that, even with a partial fair use finding, there can still be significant exposure for defendants.
- The court concluded that creating a digital copy from a purchased print copy is fair use.
- Bartz v. Anthropic PBC, No. C 24-05417 WHA, 2025 WL 1741691, at *7-14 (N.D. Cal. June 23, 2025).
- Kadrey v. Meta Platforms Inc., No. 23-cv-03417-VC, 2025 WL 1752484 (N.D. Cal. June 25, 2025).
