Context: Last month Claude AI maker Anthropic, which is backed by Amazon and Google, scored a glass-half-full win as Judge William H. Alsup of the United States District Court for the Northern District of California held (June 24, 2025 ai fray article) that training large language models (LLMs) on copyrighted works constitued fair use. But the glass is half-empty for Anthropic because the senior judge also held that Anthropic’s acquisition of the works in question by downloading pirated material entitled the authors to statutory damages and was not excused by the ultimate purpose of LLM training.
What’s new: Late on Monday (July 14, 2025), Anthropic brought a motion asking for an interlocutory appeal, i.e., a very rare exception from the rule that normally only final judgments can be appealed. Anthropic wants the United States Court of Appeals for the Ninth Circuit, whose appellate territory comprises the entire West Coast, to overturn the part that Anthropic lost and hold that the end (LLM training) justifies the means (downloading and use of material from pirate sites). As a hypothetical procedural alternative, Anthropic wants to ask the district court for reconsideration of the part of its summary judgment ruling that was unfavorable to Anthropic. Either way, Anthropic would want the district court to put the case on hold, hoping that it would be cleared of all potential liability.
Direct impact: The first thing to understand is that if a rare interlocutory appeal was allowed here. it would be logically inevitable for the Ninth Circuit to address the entire fair use question, not just the part that Anthropic lost over pirated works. But interlocutory appeals are an extraordinary step, and Judge Alsup is known to have strong opinions and is not going to change his mind on whether Anthropic’s use of pirated material should be condoned. It is also known that he actually believes the appropriate next step would be a settlement (May 23, 2025 ai fray article). That would also be a logical resolution of the matter in our view, and it could be that Anthropic primarily brought the motion to improve its bargaining position. It is, however, also possible that Judge Alsup takes a similar decision as his colleagues Judge Jon S. Tigar (in the same district), who granted class-action lawyers an interlocutory appeal of a Digital Millennium Copyright Act (DMCA) question as we predicted in that Doe v. Github case (September 28, 2024 ai fray article), or Judge Stephanos Bibas (in the District of Delaware) in Ross Intelligence v. ThomsonReuters (April 16, 2025 ai fray article). Also, Anthropic might succeed with a petition to the Ninth Circuit after Judge Alsup denies the motion, but such a petition would face a high hurdle.
Wider ramifications: In any of the scenarios outlined above where an interlocutory appeal might happen, Bartz v. Anthropic would become the first-ever AI fair use appeal to be filed on the U.S. West Coast, and depending on what happens in the aforementioned Delaware case (where the Third Circuit now has to decide), the second or even the first in the entire United States. Also, it is only a question of when a fair use appeal will be brought in the Second Circuit in light of various AI copyright cases (such as the one by the New York Times) pending in the Southern District of New York. Internationally there is the potential of an AI copyright case involving somewhat similar questions as fair use reaching the European Court of Justice (ECJ), though under EU law there is no direct equivalent of fair use to be clear. A German collecting society will get its day in the Munich I Regional Court soon, and that court could refer certain questions directly to the ECJ.
Here’s Anthropic’s motion, followed by further commentary:
If an interlocutory appeal happens, the whole fair use question will be on the agenda
The most important aspect of whether a district court allows an interlocutory appeal or whether the appeals court accepts such a petition is not even the merits. It’s about whether there would be an objective benefit from near-term appellate clarification of a question and also has to do with case management. As far as the latter is concerned, the district court could, in the absence of a settlement, simply hold a copyright infringement trial (as planned), where the book authors would be awarded a certain amount of damages, and Anthropic could either settle then or appeal.
Anthropic essentially argues that its ultimate purpose is wonderful, it’s absolutely fair use, and therefore it doesn’t really matter that it pirated the material in the first step. The Ninth Circuit wouldn’t be able to hold that the end justifies the means without also ruling on whether the end actually is fair use. And on that question, the book authors’ class action will still disagree, even if Anthropic was the only (or at least the first) party to seek an interlocutory appeal. They could bring a contingent motion for a cross-appeal; after a final judgment it would be a non-contingent (cross-)appeal; but in this case it’s actually a situation where the Ninth Circuit can’t really avoid the question anyway when it discusses the case with the parties, given that Anthropic can’t win unless the Ninth Circuit feels strongly about the LLM training part.
The book authors could win if the Ninth Circuit says that no matter how “fair” the ultimate use was, using pirated material is unlawful, period. But the appeals court wouldn’t prejudge the case, so it wouldn’t know in the beginning whether that’s going to be the conclusion.
Furthermore, if the whole litigation was put on hold anyway, the book authors would also want to appeal the part of the summary judgment ruling that was unfavorable to them. So from a case management point of view, it would then also be a given that fair use would be analyzed as a whole and not just with a view to the part Anthropic lost.
Anthropic is taking considerable risk by not settling at this stage
Anthropic knows that it’s taking a risk here: It could be cleared of all liability, but the whole situation could become a lot worse if the holding that LLM training is fair use got overturned. In that case, Anthropic would be jumping out of the frying pan into the fire.
It is in the nature of class-action litigation that the parties and their lawyers, who typically work on a full-contingency basis, want to get paid. That is presumably why Judge Alsup suggested a settlement, and it is why a settlement would still make sense.
It’s possible that Google and Amazon are strongly influencing Anthropic. For Anthropic itself, it would make more sense to agree on a reasonable amount to pay to the book authors and their lawyers. But the involvement of Stanford professor Mark Lemley, who is extremely well-respected but paid by Google and has been consistently Google-aligned on every major competition or IP issue for decades (even on issues where Google was and is on the wrong side of history), is one indication of Google taking a very strong interest in this case. Anthropic’s leadership will have to ask itself whether Amazon and/or Google are basically using Anthropic, and potentially risking Anthropic’s future, because of their interest in obtaining appellate clarification of some wider questions. Anthropic would basically be fighting a proxy war where Google and Amazon stand a lot to win, but Anthropic stands the most to lose if it doesn’t settle now. Anthropic also has interests beyond just the book authors’ case, but given that Claude AI’s key differentiator is the ability to generate long texts, this case is of particular relevance to Anthropic. Amazon and Google have very far-reaching business interests beyond this case.
Questions of law
The basis on which Anthropic hopes its motion will succeed is that it argues there are two novel questions of law:
- whether fair use is analyzed by reference to the defendant’s ultimate use of a copyrighted work or by parsing the defendant’s use into subsidiary steps; and
- whether a defendant’s acquisition of a copyrighted work from an unauthorized source counsels strongly against fair use, even if the defendant’s ultimate use of that work is transformative.
The first question is about whether the acquisition of the relevant works should be judged in its own right (like saying “you downloaded pirated material, so you infringed no matter what you did thereafter”) or whether everything should be amalgamated (“let’s look at what you did as a whole and yeah, on balance this was fine apart from a cosmetic issue at the start”). Therefore, the first question comes down to whether there are actions such as downloading and storing pirated works that can give rise to liability in their own right or whether everything must be lumped together with an “end justifies means” attitude.
The second question is about “whether a defendant’s initial acquisition of a copyrighted work from a third party who distributed the work without permission counsels strongly against fair use—even if the defendant’s ultimate use of that work is transformative.” This is not about whether the initial act could be unlawful in its own right, but whether the piracy part should be given a lot of weight in the fair use analysis.
It’s hard to see how the second question could even be reached unless Anthropic prevailed (in theory) on the first. If the answer to the first question is that separate steps matter, the ultimate use wouldn’t play a role. But Anthropic’s lawyers may hope that the second question would be an alternative road that would lead to the result they desire (Anthropic being cleared of infringement on fair use grounds).
Implications of Judge Chhabria’s decision in Kadrey v. Meta in the same district
Anthropic’s motion references a very recent decision by a different judge in the same district that is actually a double-edged sword: Judge Vince Chhabria‘s partial summary judgment decision against a class action and in favor of Meta (June 25, 2025 order (PDF)).
The part that Anthropic likes about Judge Chhabria’s decision is that he didn’t consider the download and use of pirated works as important and independently unlawful as Judge Alsup did.
But there are dicta (things that a judge says about the law that are not outcome-determinative in the case at hand, but could matter elsewhere) in Judge Chhabria’s decision that take a far less fair-use-friendly position than Judge Alsup.
Meta only won partial summary judgment on fair use against the Kadrey class of copyright holders because the class-action lawyers (who later, and possibly too late, brought in reinforcements from the Boies Schiller firm) had failed to plead certain facts that could have resulted in a finding of no fair use. Here’s what Judge Chhabria wrote in his order:
“Given the state of the record, the Court has no choice but to grant summary judgment … [A]s should now be clear, 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.”
While Judge Alsup’s decision in Anthropic’s favor is not based on an evidentiary failure, Judge Chhabria’s is, and that was clear at Judge Chhabria’s summary judgment hearing (parts of which we followed by video). While Judge Alsup’s take is that LLMs do something completely different than books and considered that fact pretty much dispositive, Judge Chhabria noted that “[t]here is certainly no rule that when your use of a protected work is ‘transformative,’ this automatically inoculates you from a claim of copyright infringement.” Judge Chhabria is not worried about innovation: even if using copyrighted works is not deemed fair use, there’s enough revenue potential in AI systems that authors can be compensated and AI providers will make a ton of money.
The most important part of Judge Chhabria’s take, with a view to all those AI copyright cases where the fair use defense is typically the one that decides everything, is that he essentially advocates an indirect substitution theory. Even if copyright holders can’t show that he output of those LLMs directly competes with their works, Judge Chhabria is way more copyright holder-friendly than Judge Alsup and concerned about the indirect market impact of LLMs “flood[ing] the market” at record speed (in fractions of a second, basically) with works that do compromise the opportunity for human authors to earn their livelihood.
If Anthropic was granted an interlocutory appeal, the book authors have no shortage of arguments for why the whole question of fair use needs to be addressed on appeal anyway, but Judge Chhabria’s ruling also supports a review of the question of whether LLM training constitutes fair use.
Procedurally, it’s not clear what will happen next in the Meta case. There is a possibility of an interlocutory appeal there, but also one of a near-term final (albeit partial) judgment leading to an appeal.
Procedural steps
Anthropic requests a hearing on August 28, 2025. The court will decide, and it might even do so without a hearing.
The book authors’ lawyers will now get to respond, and Anthropic will then reply. We may now see the book authors move for an interlocutory appeal of the part they lost, but presumably on a contingent basis where they’d rather proceed to trial on the current basis, but want to appeal the part they lost in case Anthropic gets to appeal at this stage.
Ultimately, the whole question of whether LLM training constitutes fair use will sooner or later reach the Supreme Court. The question is just in which case(s).
Counsel
Counsel for Anthropic
- Cooley’s Kathleen Hartnett, Ephraim McDowell and Alexander J. (“Alex”) Kasner.
- Arnold & Porter Kaye Scholer, San Francisco: Douglas A. Winthrop, Joseph Farris, Jessica L. Gillotte and Estayvaine Bragg.
- Arnold & Porter Kaye Scholer, Los Angeles: Angel T. Nakamura, Oscar Ramallo and Allyson Myers.
- Stanford professor Mark Lemley of Lex Lumina, who has been working for Google (a key Anthropic partner) for decades.
- Latham & Watkins, San Francisco: Joseph R. Wetzel and Andrew M. Gass.
Counsel for the book authors
- Susman Godfrey’s Justin A. Nelson (lead counsel; he’s involved with virtually every major AI copyright case, on the plaintiff side, apart from certain cases against Meta), Rohit Nath, Alejandra C. Salinas, Collin Fredricks, Rohit Nath, Jordan W. Connors and J. Craig Smyser.
- Lieff Cabraser Heimann & Bernstein’s Rachel Geman (co-lead counsel), Jacob S. Miller, Danna Z. Elmasry, Daniel M. Hutchinson and Reilly T. Stoler.
- Cowan DeBaets Abrahams & Sheppard’s Scott J. Sholder and CeCe M. Cole.