Rapid fire fair use decisions suggest AI training is permissible -- sometimes

Bartz v. Anthropic

On June 23, 2025, the U.S. District Court for the Northern District of California issued a landmark ruling partially granting Anthropic PBC's motion for summary judgment in a putative class action brought by book authors. The court found that Anthropic's use of lawfully acquired copyrighted books to train its LLM constituted fair use, but denied judgment on claims involving books obtained from piracy sites, which will proceed to trial.

Anthropic, an artificial intelligence firm, compiled a large volume of copyrighted works into a digitized “central library” that was used in part to train its generative AI model, Claude. Some of these works were lawfully purchased from commercial distributors, while others were downloaded from piracy sites.  Importantly, the infringement allegations centered entirely on the use of copyrighted works for training; plaintiffs did not allege that Claude reproduced any copyrighted material in its outputs, and the record confirmed it did not.

In an opinion that analogized AI training to the way humans learn from the content we consume, Judge William Alsup held that Anthropic's use of lawfully acquired copyrighted works to train Claude was “quintessentially transformative” and thus fair. Relying heavily on the model's generation of new material and the lack of infringing expression in its output, the court concluded that “authors cannot rightly exclude anyone from using their works for training or learning as such.”

Two key points underpin the court's decision, with potential implications for the various other pending AI training cases.  First, the court noted that its holding applies where the works used to train were “purchased or otherwise accessed lawfully.”  In this way, the opinion suggests that fair use does not require model developers to purchase their training data so long as they access it lawfully, leaving open the possibility that content obtained through methods such as lawful web scraping could also qualify. Second, the court was clear that its holding only applies where the training process at issue does not generate infringing outputs or result in the provision of infringing copies to end users.  The court was explicit that, “if that were not so, this would be a different case.”

Kadrey v. Meta Platforms, Inc.

Days later, on June 25, 2025, another Northern District of California judge reached a similar conclusion, but on a different basis. Like Judge Alsup, Judge Vince Chhabria held that Meta's use of copyrighted books to train its AI model, Llama, was “highly transformative” because the copying served a different purpose: teaching an LLM to perform new generative functions.  But Judge Chhabria did not rest his fair use holding on that transformative nature, instead finding plaintiffs had failed to demonstrate harm to the market for plaintiffs' books by virtue of such training.

The case dealt with a similar set of facts as Anthropic. To train its LLM, Meta sourced extensive data from varying sources, including two book databases which housed the plaintiff's copyrighted books. Also like Anthropic, there was no evidence of Llama producing infringing output.  Meta “post-trained its models” to prevent them from outputting infringing material, and extensive testimony indicated that the Llama model could not, at present, produce “any significant percentage” of the plaintiffs' books.

Unlike the Anthropic court, however, Judge Chhabria focused his decision on the market harm element of the fair use analysis, criticizing Anthropic's reliance on the purpose and character of the use and emphasizing that transformative nature is not dispositive.  Instead, Judge Chhabria took the position that the fourth factor of the fair use analysis, the effect on the market, “is the most important.”

With respect to that market harm, the Court characterized plaintiffs' theories as arguing, first, “that Llama is capable of reproducing small snippets” from their works and, second, “that Meta, by using their works for training without permission, has diminished the authors' ability to license their works for the purpose of training large language models.”

Referring to these theories of market harm as “half-hearted,” the Court described both of these theories as “clear losers,” making sweeping findings that “Llama is not capable of generating enough text from the plaintiffs' books to matter, and the plaintiffs are not entitled to the market for licensing their works as AI training data.”

But the Court did not find the concept of market harm to be futile. To the contrary, the Court recognized that its own conclusion on market harm may “be in significant tension with reality,” and described another argument plaintiffs did not pursue, namely “that Meta has copied their works to create a product that will likely flood the market with similar works, causing market dilution” as “potentially winning.”

Ultimately, the court explained: “this ruling does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful.  It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”  Of course, the ruling did not find Meta's conduct unlawful, either.

Looking Ahead

These decisions mark the first time a federal court has addressed whether the use of copyrighted materials to train generative AI qualifies as fair use.  While both decisions ultimately decline to find such use infringing, they do so on different bases, with both recognizing the possibility of a contrary outcome under different circumstances.

As a result, and while these decisions signal that AI training can be considered transformative and may under certain circumstances be fair, the broader fair use debate – and the legal framework under which it will be decided – is far from resolved.

Authored by Lauren Cury, Anna Kurian Shaw, and Casey Magersupp.