Not So Fast: Some Oddities in the Anthropic Fair Use Opinion

First published on 6/26/25 on The Illusion of More

Headlines flood the feeds announcing that a California District Court sided with AI developer Anthropic, finding that LLM training with unlicensed works is fair use. While the headlines are true, I wouldn’t read the conclusions as gospel just yet. In the big picture, we are going to see a variety of fair use opinions in the more than 40 copyright cases against AI developers, and different facts with different legal arguments—and in different circuits—are likely to yield a mosaic of results. And odds are, there will be a question or two that goes to the Supreme Court. But in the meantime, in ruling on fair use for the defendant in Bartz v. Anthropic, Judge Alsup’s opinion includes some oddities that make it ripe for significant challenges in the Ninth Circuit. Here are the basics:

Anthropic acquired and ingested books in two ways to train the LLM it calls Claude. Initially, it scraped books from pirate libraries—an act the court held to be infringing, and which some legal experts have noted could result in a devastating damage award in a trial on that issue. But more significant as potential guidance to AI training overall, the court held that Anthropic engaged in fair use when it purchased printed books and “destructively scanned” those books to make digital copies it used to train Claude. The author plaintiffs did not allege that the defendant output any infringing copies of books that were fed into the model.

With that, the following oddball aspects of the decision stand out to me. The discussion is not exhaustive, and others’ opinions will vary:

Transformativeness as Near Dicta

In his opinion, Judge Alsup characterizes the purpose of the use as “spectacularly” and “quintessentially” transformative and “among the most transformative we will see in our lifetimes.” Yet, in the 12-page section weighing factor one, the opinion barely discusses a rationale for finding the purpose of the use transformative under any authorities. For instance, the opinion notes that like Google Books, Claude also does not make infringing outputs available, but the opinion does not articulate a cultural, social, intellectual, or other stated purpose of Claude as a rationale for finding it transformative. This apparently self-evident conclusion leaves one to wonder whether the novelty of the technology itself distracted the court from a proper analysis under the fair use doctrine. As Terry Hart writes on Copyhype:

In this context, a use is not transformative merely because it produces something new or technologically sophisticated. Rather, as established through decades of judicial interpretation, a transformative use is one that relates back to the original work by creating new information and insights about that work. Generative AI does not do this. It instead reappropriates the expressive content of the work to enable the generation of synthetic expressive content completely divorced from the original work.

Contradicting Warhol

The court’s apparently tautological holding that Claude is “transformative” on the sole basis that it is a high-tech achievement, contradicts recent fair use precedent, including the Supreme Court’s ruling in AWF v. Warhol. As Hart notes, and as discussed in other posts, making “something new,” no matter how cool, is not a sufficient basis for finding that a use is transformative. Because Judge Alsup does not persuasively explain why Anthropic’s purpose is transformative (let alone “spectacularly” so) in context to copyright law, we might expect this principle to be vigorously argued on appeal.

Odd Citing to First Sale Doctrine

The opinion finds it compelling that Anthropic legally obtained physical copies of books that were destroyed in the process of scanning them to make physical copies. Judge Alsup cites details like bindings being ripped off to emphasize the point that only one copy remained of each book and that the process of converting physical books into digital books served a transformative purpose…

Anthropic purchased its print copies fair and square. With each purchase came entitlement for Anthropic to “dispose[ ]” each copy as it saw fit. 17 U.S.C. § 109(a). So, Anthropic was entitled to keep the copies in its central library for all the ordinary uses.

The statutory first sale rule categorically allows you or me to “sell or otherwise dispose of” a copy of a lawfully obtained book by giving it away, reselling it, burning it, or using it as a doorstop—but it absolutely does not allow copying it for the sake of convenience, storage, or any other purpose. That the court cites §109 as permitting mass copying in a commercial venture like AI training is perplexing. Moreover, the opinion contradicts case law. As the SDNY stated in Hachette v. Internet Archive “IA points to no case authorizing the first recipient of a book to reproduce the entire book without permission….”

Equally perplexing, the court appears to find that the unlicensed copying serves a transformative purpose separate from the overall purpose of LLM training…

Anthropic argues that the central library use was part and parcel of the LLM training use and therefore transformative. This order disagrees. However, this order holds that the mere conversion of a print book to a digital file to save space and enable searchability was transformative for that reason alone.

I cannot think of any authority to support that finding. “Saving space” and “searchability” served internal, operational purposes for Anthropic—purposes that cannot stand alone, but which can only tilt factor one toward fair use if the overall purpose of those operations is transformative. Yet, Judge Alsup appears to say that the overall purpose (despite being “quintessentially” transformative) is immaterial, thereby implying that merely converting print books to digital copies is transformative on its own. If that is what the opinion is saying, it contradicts every court’s rejection of such claims and endorses violation of the derivative works right of authors.

Shoehorning Infringement Considerations into Fair Use

In this and any copyright infringement case, the plaintiff(s) must argue that the protectable “expression” has been copied by the defendant. Here, the court finds the following:

Yes, Claude has outputted grammar, composition, and style that the underlying LLM distilled from thousands of works. But if someone were to read all the modern-day classics because of their exceptional expression, memorize them, and then emulate a blend of their best writing, would that violate the Copyright Act? Of course not.

True but problematic. First, the court should decide whether protected expression was unlawfully copied and then, if so, whether the copying is exempted by fair use. Here, the court appears to collapse these considerations into the factor one analysis by inaptly shifting attention to Claude’s outputs, which are not at issue in this case. Taken in combination with the fact that the opinion offers scant discussion for finding transformativeness overall, this citation to what copyright does not protect (§102) tucked into the fair use analysis suggests that the court is overly sympathetic to the idea that the purpose of generative AI is analogous to human authorship. This is an error, both as a matter of cultural interest and law.

Misreading the Purpose of Copyright Law

Further validating my concern that the court is somewhat distracted by the shiny new technology, the opinion states the following:

…Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works. This is not the kind of competitive or creative displacement that concerns the Copyright Act. The Act seeks to advance original works of authorship, not to protect authors against competition. [emphasis added]

As discussed in other posts, that view misunderstands the nature of GAI by assuming it “advances original works of authorship” in competition with “authors.” This is unfounded. The court has absolutely no basis to assume that merely because Claude is allegedly capable of outputting works that mimic “good writing,” that those outputs will be works of “authorship” as a matter of law. On the contrary, the better Claude is at “writing” without a writer, the more it will necessarily output non-human expression that is not protected by the Copyright Act.

It cannot be reasonable for the courts to find that the purpose of our copyright law is to breed mass production of uncopyrightable works or to incentivize the poor author to pose as a great author through the mask of technology. That is a recipe not only for widespread market dilution but for cultural and intellectual dilution in direct conflict with the constitutional rationale for intellectual property law.

Regarding market dilution, Judge Chhabria in his recent finding for Meta in the Kadrey case, criticized Judge Alsup’s inapt comparison to “training schoolchildren” and understating the potential market harm of GAI. Market dilution is a novel consideration discussed in the Copyright Office’s third report on AI, and that will be the subject of a new post. In the meantime, buckle up. The AI fair use ride is just getting started.