Bartz v. Anthropic: Class Certification at Issue for Book Authors in AI Copyright Litigation

This is the first of a two-part post exploring the class action features of the Bartz v. Anthropic case and what to watch for if the parties in it, or the other copyright AI suits, pursue a settlement. 

Back in October, I asked a simple but unresolved question: “Who represents you in the AI copyright suits?” Now, eight months later, we’re getting closer to an answer—at least for some authors. In Bartz v. Anthropic, one of the fastest-moving lawsuits over the use of copyrighted works to train generative AI systems, the plaintiffs have asked the court to certify a class of authors whose books were allegedly copied without permission.

That procedural question—whether a few named plaintiffs can speak for thousands or millions of authors of books—is now before the court alongside a second, and even more consequential, one: Does fair use allow AI companies to copy books for training purposes at all? With motions for both class certification and summary judgment now pending, Bartz is a critical test case for the future of copyright and AI.

Class certification is what allows a few named plaintiffs to represent thousands—or even millions—of rightsholders. That procedural move changes everything: It multiplies the potential damages at stake, dramatically increases the plaintiffs’ leverage in settlement talks, and, if successful, means that either by judicial resolution or settlement, the result will be binding for everyone in the certified class who does not opt out. So it’s worth looking closely at how the Bartz plaintiffs define their class, who they’ve chosen to represent it, and what implications that has for authors more broadly.

What Is the Class in Bartz, and Why Does It Matter?

Getting a court to certify a class is a big deal. It allows a few named plaintiffs to represent a much larger group of people—here, potentially millions of authors—whose claims are similar. This can multiply the stakes of a case significantly. For example, statutory damages under copyright law can be up to $150,000 per work infringed. Multiply that by thousands or millions of books, and the potential liability becomes enormous. That leverage often encourages defendants to settle.

Compared to other AI lawsuits that sweep in every kind of online creator, the proposed classes in Bartz are more targeted: they focus on book authors. But even this narrower scope involves a potentially massive number of rightsholders. The plaintiffs propose two classes:

• Pirated Books Class: All natural persons, estates, literary trusts, and loan-out companies that are legal or beneficial owners of copyrighted works that: (a) were registered with the United States Copyright Office within five years of the work’s publication; (b) were registered with the United States Copyright Office before being downloaded by Anthropic, or within three months of publication; (c) are assigned an International Standard Book Number (ISBN) or Amazon Standard Identification Number (ASIN); and (d) were downloaded by Anthropic as part of [REDACTED].

• [REDACTED] Books Class: All natural persons, estates, literary trusts, and loan-out companies that are legal or beneficial owners of copyrighted works that: (a) were registered with the United States Copyright Office within five years of the work’s publication; (b) were registered with the United States Copyright Office within three months of publication or before being [REDACTED]; (c) are assigned an International Standard Book Number (ISBN) or Amazon Standard Identification Number (ASIN); (d) were [REDACTED]; and (e) were used by Anthropic in LLM training.

We’re really guessing at the scale and scope of the books included in these classes, but the briefs give us a few hints. For one, it seems clear that they include a wide diversity of authors. Anthropic highlights in its brief, for example, that some books are academic books that may be produced as “work for hire,”  that some are likely owned by foreign rightsholders, that some are older, published before 1964. And on overall size, Anthropic states that at least one data source contains 5 million books. 

Finally, we learn a few more details from Judge Alsup’s request for supplemental briefing . He asks the parties to respond to an alternative proposal for classes that names Books3, PiLiMi, and LibGen as sources for the “pirated” class. Judge Alsup also identifies the second class as the “scanned book” class. Though we do not know where Anthropic ultimately obtained the books that were scanned, we do see in the Plaintiff’s brief in support of its motion for class certification that Anthropic “acqui[red]” these books. 

Anthropic, in opposing the motion for class certification, suggests that the proposed classes would cover “millions” of rightsholders. To represent this enormous group, the plaintiffs have identified three authors:

• Andrea Bartz, author of thrillers like The Lost Night, The Herd, We Were Never Here, and The Spare Room, published by Penguin Random House;
  
• Charles Graeber, nonfiction author of The Good Nurse and The Breakthrough, published by Hachette; and
  
• Kirk Wallace Johnson, author of nonfiction works including To Be A Friend Is Fatal, The Feather Thief, and The Fishermen and the Dragon.
  

Whether these authors “fairly and adequately” represent the proposed classes is a legal question—but it’s also a governance one. If the class is certified and the case settles, only those within the class will be bound by the deal. If the class is too broad, it could include authors with very different legal interests or views about AI training—raising serious fairness concerns about these authors directing legal arguments on others behalf. 

How will Class Certification Be Determined? 

The court must evaluate a motion for class certification under Rule 23 of the Federal Rules of Civil Procedure which states that a class action may be certified only if: 

(1) the class is so numerous that joinder of all members is impracticable;

(2) there are questions of law or fact common to the class;

(3) the claims or defenses of the representative parties are typical of the claims or defenses of the class; and

(4) the representative parties will fairly and adequately protect the interests of the class.

Additionally, for the type of class action brought here, the court must conclude that “questions of law or fact common to class members predominate over any questions affecting only individual members, and that a class action is superior to other available methods for fairly and efficiently adjudicating the controversy.”

In its April 2025 brief opposing class certification, Anthropic raises a host of objections—ones that we mostly agree with. It rightly points out the complexity of trying to establish just what books are included, given the lack of reliable book data, a daunting task, especially with one so large as is proposed here.  The company also warns that the proposed class would necessarily exclude many affected authors—especially those whose books are not registered in the U.S., or whose rights status is ambiguous. It argues:

“If Plaintiffs prevail on their fair use theory, authors whose works were used to train Claude but who are not part of the certified class—perhaps because their works are not registered or because their publishers reside abroad—will no doubt file follow-on cases making similar arguments.”

That’s a recipe for ongoing litigation, not resolution. It also reflects a deeper problem: identifying affected class members is often impossible. Many books were published by foreign publishers not clearly listed in U.S. copyright records. Others are orphan works, whose rights holders are missing or untraceable. Some authors published under pseudonyms or with defunct presses. And crucially, the right to authorize AI training often depends on the author’s publishing contract—a private agreement that’s usually not public, and often unclear even to the author. 

In our experience, many authors aren’t sure which rights they hold, or whether they’ve retained the right to license their books for machine learning purposes. We know that across the books at issue there are a likely hundreds of thousands to millions of publishing contracts allocating rights for the book, along with option agreements, sublicensing deals,  and so on. That makes responding to a notice of inclusion in a class—let alone affirmatively asserting rights—deeply confusing. 

Anthropic questions whether the three named plaintiffs fairly represent the broader author population of the subject books. Their brief explains: 

Claude assists people, including authors, with writing tasks. Many members of the classes benefit from the use of AI tools and do not wish to see them crippled by licensing. Academic authors in particular use AI in their professional work, and many of them use books in the same way Anthropic does: to develop LLMs. Similarly, Oxford University Press surveyed academic researchers and found that 76% of researchers use AI tools in their research. Plaintiffs cannot adequately represent a class of people when a significant number of people in that same class benefit from the very conduct they claim is unlawful. (citations omitted). 

The implication is clear: the proposed class sweeps together authors with divergent views, experiences, and economic stakes. Some authors might oppose licensing their works for AI. Others might welcome it. Some use generative tools in their own writing and support broad fair use interpretations. Still others may be unaware that this is happening at all. These perspectives cannot easily be reconciled into a single litigation position—and Anthropic argues that this prohibits binding them under a unified class.

These objections may carry real weight with Judge Alsup, who is known for his analytical rigor and skepticism of overbroad class actions. But they also echo longstanding debates about whether class actions are a good vehicle for resolving complex copyright disputes.

What’s Next?

Judge Alsup’s decision on class certification—and on the fair use issue—could shape the legal landscape for years to come. If the class is certified, it could significantly increase the pressure on Anthropic to settle. If it’s denied, it may signal that individualized lawsuits—or new legislative solutions—are the only way forward.

Either way, the question of “who speaks for authors” remains urgent but possibly unanswerable. In the next post, we’ll turn to what happens if these lawsuits settle—and whether those settlements will bind, include, or leave out the very creators whose rights are most at stake.