AI Companies Prevail in Path-Breaking Decisions on Fair Use

Last week, artificial intelligence companies won two significant copyright infringement lawsuits brought by copyright holders, marking an important milestone in the development of the law around AI. These decisions – Bartz v. Anthropic and Kadrey v. Meta (decided on June 23 and 25, 2025, respectively), along with a February 2025 decision in Thomson Reuters v. ROSS Intelligence – suggest that AI companies have plausible defenses to the intellectual property claims that have dogged them since generative AI technologies became widely available several years ago. Whether AI companies can, in all cases, successfully assert that their use of copyrighted content is “fair” will depend on their circumstances and further development of the law by the courts and Congress.

A large number of cases pending in federal courts claim that AI companies violated the Copyright Act when they trained their AI systems on allegedly copyrighted datasets. Although case-specific allegations differ in some respects, at the core these cases allege that, in training their AI systems, companies copied materials protected by the Copyright Act – such as books, visual arts, and newspaper articles – and are thus liable for copyright infringement. Commentators have suggested that any one such copyright infringement case could result in billions of dollars in potential liability to AI companies.

In response to copyright infringement claims, AI companies raised several arguments and defenses. Common across all cases, defendants assert that they are not liable for infringement because any use is fair use. 

When asserting fair use, defendants essentially argue that even where there is infringement, there is no liability. While a copyright gives the author a virtual monopoly on their work, the fair use doctrine recognizes that this monopoly can impede and stifle creativity. Therefore, the Copyright Act and courts allow the use of a copyrighted work — even if the use constitutes infringement — when that use is considered “fair.” Courts applying this doctrine have, for example, allowed parodies of songs to be made, the use of copyrighted materials for news reports or educational purposes, and the copying of software code to develop mobile phone operating systems or computer games.

In the context of AI training cases, three courts now have decided whether the “fair use” doctrine applies.  A district court in Delaware concluded that the fair use doctrine did not excuse the alleged copyright infringement.  Last week, two courts in the Northern District of California concluded otherwise, for different reasons.

Doctrinally, fair use jurisprudence considers four factors — no one of which is dispositive: 

1. Factor one contemplates whether the copying was intended for commercial use and whether the end product was “transformative,” meaning the purpose and character of the use was not the same as the copyrighted material. If commercial or not transformative, then this factor weighs against applying the defense.
2. Factor two contemplates whether the work at issue is more factual or a work of fiction. The more factual a work, the less copyright protection it deserves.
3. Factor three contemplates how much of the work was infringed. The more that was used, the more this factor weighs against fair use.
4. Factor four contemplates whether the end product created by using the copyrighted material hurts the market for the copyrighted materials that were infringed.

In the Delaware case of Thomson Reuters v. ROSS Intelligence, Judge Stephanos Bibas held in February 2025 that there was no fair use in the incorporation of copyrighted materials for AI training on the basis that: (i) under the first factor, the original materials used that were alleged to infringe were subject to full copyright protection, and (ii) the end product that resulted from the alleged infringement was a product that competed directly against the copyright holder’s original product, thereby failing the fourth factor.

In contrast, both Northern District of California Judges William Alsup in Bartz v. Anthropic and Vince Chhabria in Kadrey v. Meta held that fair use excused the alleged infringement. While both courts came to the same conclusion, they varied in their reasoning of how the fair use doctrine’s four-factor test applied and with significant caveats. In this differentiation, the courts demonstrated that how and why the fair use doctrine applies remains uncertain. Appellate courts or Congress may provide further clarification and direction.

In Judge Alsup’s Anthropic decision, the court found that using copyrighted books to train Claude, Anthropic’s core software product, was fair use. There, plaintiffs alleged Anthropic downloaded pirated copies of their work and then used their stolen works to train a large language model to generate new text.

In evaluating factor one, the court concluded that using books to train Claude was akin to a human reading books to learn to read and write.  In the court’s view, nothing can be wrong with a human reading the classics to emulate their style.  Accordingly, Claude’s output did not infringe, and therefore, Anthropic used books to create something completely transformative.  In evaluating factor two, the court concluded that the books were entitled to the highest level of copyright protection.  This factor weighed against fair use. In evaluating factor three, the court concluded that Anthropic used as much of the original works as was necessary to obtain the result.  This weighed in favor of fair use.  Finally, in evaluating factor four, the court decided that there was no harm to the market for human-generated books.  The court also rejected the notion that it should consider whether the market for training data was somehow implicated because that falls outside the protection of the Copyright Act.

Separately, Judge Alsup found that digitizing lawfully purchased books was also fair use, but for a different reason.  That use, the court found, was also “fair” because it did not create new copies, new works, or redistribute copies.  Instead, the defendant replaced its own, lawfully-procured physical copies “with more convenient space-saving and searchable digital copies.” 

However, Judge Alsup drew a stark distinction between Anthropic’s use of books that it purchased lawfully and those that it used after downloading them for free from pirate sites on the Internet.  “Anthropic had no entitlement to use pirated copies,” the court held.  “Creating a permanent, general-purpose library” of stolen books was clear copyright infringement that could not be excused as fair, the court held.

Judge Chhabria issued his decision in Meta days after Judge Alsup’s decision in Anthropic.  Judge Chhabria also found the defendants’ use of the copyrighted work to be “fair,” but for different reasons and with some criticisms of Judge Alsup’s reasoning.

As in Anthropic, authors sued Meta for using illicit copies of their works to train a large language model called Llama.  And, as in Anthropic, the court found that generative AI was necessarily a transformative use, satisfying factor one, because Llama does not and was not intended to re-create the works on which it was trained.  By contrast, Llama can assist with “a wide range of functions,” including finding recipes, translation, “creative ideation,” and generating diverse texts.

Unlike Judge Alsup, Judge Chhabria in Meta found that the fact that some of the copyrighted works were obtained illicitly did not weigh against fair use.  Even if Meta used works obtained illicitly and in “bad faith,” “[t]he purpose of fair use is to allow new expression that won’t substitute for the original work, and whether a given use was made in good or bad faith wouldn’t seem to affect the likelihood of that use substituting for the original.” The court held that this was particularly true where defendants did not show that Meta’s use of illicitly obtained copies promoted further improper infringement by others.

As in Anthropic, Judge Chhabria found that factor two (the nature of the copyrighted work) -- which included “highly expressive works,” including “mostly novels, memoirs, and plays” -- favored plaintiffs.  In line with the court in Anthropic, Judge Chhabria held that the third factor (the amount and substantiality of the portion of copyrighted material taken) favored Meta.  He held that although Meta copied the entirety of the plaintiffs’ books, the amount it copied was reasonable given its relationship to Meta’s transformative purpose.

Finally, in evaluating factor four (market harm), Judge Chhabria concluded that, because the market for training data is not one “that the plaintiffs are legally entitled to monopolize,” this factor weighed in favor of fair use.

Having reached the same ultimate conclusion as Judge Alsup, Judge Chhabria in Meta discussed at length a possible argument that the plaintiffs could have made but failed to present sufficiently: that Meta had “copied their works to create a product that will likely flood the market with similar works, causing market dilution.” Judge Alsup in Anthropic had not considered this argument.

According to Judge Chhabria, a harm that could be recognized by factor four is not the harm of “direct substitution,” where an author’s book is supplanted by an exact or virtual replica of the book.  Rather, Judge Chhabria imagined a world in which there is “indirect substitution,” where “AI-generated books could successfully crowd out lesser-known works or works by up-and-coming authors.”  Notably, Judge Chhabria observed that, had the plaintiffs presented “any” evidence that a jury could use to find that the plaintiffs faced such “market dilution,” the claim would have needed to go to a jury.  But on the record before him relating to the 13 plaintiffs in Meta, the court had “no choice” but to grant summary judgment.  The court’s decision “stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”

Given the variable interpretations of the law by the three courts that have thus far ruled on these issues, the many cases still before courts, and the fact-bound nature of each court’s analysis, the contours of the application of the fair use doctrine to AI remain uncertain. Right now, AI companies sued for infringement have won more than they have lost.   However, as Judge Chhabria noted in Meta, whether a defendant can viably claim fair use will turn on the quality of the claims alleged, the circumstances of each case, and the evidentiary support presented.  These decisions are likely to be appealed.  Future court decisions at the district and appellate level as well as potential legislative action will likely clarify these issues.

For additional information on the intersection of artificial intelligence and copyrights, read our related client alerts:

• U.S. Copyright Office Releases Part 2 of Artificial Intelligence Report, Clarifying Copyrightability of Generative AI Outputs 
• U.S. Copyright Office Releases Third Report on AI and Copyright Addressing Training AI Models with Copyrighted Materials