What Happens if the AI Copyright Class Actions Settle?

This is the second 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. Read the first part here.

A Look at the Stakes for Authors

As the high‑profile copyright lawsuits against AI companies proceed, the courtroom drama captures headlines. But I’ve long thought that settlement may be the real outcome to watch.

We may already be entering “settlement watch” territory in one of the fastest-moving AI cases, Bartz v. Anthropic. I wrote earlier this week about how Judge William Alsup is currently weighing a motion for class certification under Rule 23, which would allow the class representatives to bring suit on behalf of the rights of millions of book authors. In a notably rare move, he granted “permission to negotiate class settlement” before class certification is resolved (something he ordinarily prohibits, though other judges allow), even telling the parties he had “invited the parties to settle the case, and sooner rather than later.” 

The encouragement came on the heels of hearings Judge Alsup held on the subject of whether Anthropics’ actions in reproducing and training on copyrighted works to train its AI models is fair use. On the training issue, he indicated that he likely would find such use to be fair use, but signaled skepticism over Anthropic’s data sources: “I have a hard time seeing that you can commit what is ordinarily a crime, but get exonerated because you end up using it for a transformative use.”

Why care about settlement?  

To be clear, there is no firm indication of a settlement just yet. But, across these AI cases, the incentives to settle are powerful.  Plaintiffs’ firms in many of these suits work on contingency—they only get paid if there is a settlement or victory. Meanwhile, litigation is expensive, especially in a case like this with complicated discovery and defendants who are motivated to defend their business model. Just take a quick peek at the class action firm representing the plaintiffs in Bartz  – you’ll see a listing of billions of dollars in wins, almost entirely in settlements.

Defendants may also have strong incentives to settle. This is especially true in copyright cases where they face potentially astronomical liability: up to $150,000 per infringed work, with claims possibly reaching into the billions. Certainly, defendants have countervailing interests–a strong fair use victory in these suits would be a boon to their business model. But for most of the big AI companies in these lawsuits, they have demonstrated ample willingness to just pay licensing fees instead. 

More cynically, it is not a stretch to observe that for biggest AI companies, they may actually have an incentive to leave the larger fair use question ambiguous so long as those firms can be sheltered from additional litigation for their activities: doing so would be pulling up the ladder after themselves, but would give them a competitive advantage while making new market entry for startups difficult.  The class action mechanism makes settlement a potent tool to achieve those ends because it would allow defendant AI firms to secure legal certainty for use of a vast corpus of materials that may be impossible to achieve with any other legal mechanism except legislation. 

Would a court approve a settlement? 

Under Rule 23, a court can approve a class settlement only if it is “fair, reasonable, and adequate.” A critical moment is the fairness hearing, where class members can raise objections or opt out.

Our closest point of comparison may be the Google Books case, where the parties formulated an incredibly complex agreement, known as the “Amended Settlement Agreement” that spanned over 150 pages (not including exhibits!). The subject in that case was several million books digitized by Google for its Google Books project. Though that settlement agreement had many forward-looking provisions for ongoing use and licensing that may or may not be an issue here, there would likely be some commonalities.

One is that even if a settlement is just for payment related to past use, someone would need to figure out how to notify potential class members of their rights and then pay out funds from the settlement to them. In the Bartz case, which is focused just on digitized books used in Anthropic’s AI training, the parties and court are already thinking about this in the context of Bartz’s motion for class certification. So far, the solutions are incredibly incomplete. For example, the court asked the parties to critique a proposal where the classes would be “registered owners of the right to reproduce or prepare derivative works” of the “books” at issue and that notice provided to putative class members “shall refer readers to a searchable website listing all scanned books by author and title and ISBN. The notice shall state that anyone who wishes to correct the registered copyrights must do so by MM-DD-YYYY.” 

To begin, the books at issue are not so easily identifiable–  from the filings, it is clear that they were published by publishers from across the world, across a long period of time spanning all of the last century, and across a wide variety of fields. Identifying exactly which of these books in the data sets at issue even have an ISBN could be a major task. But more significantly,  there simply is no such registry of who holds the rights to control reproduction and distribution in a given book. Unlike some other areas such as with some areas of music, there is no database that comes anywhere close to providing the scale and complexity of rights data needed to figure this out.  This is why in the Google Books Settlement, Google had to agree to create and fund–for $34.5 million dollars–a Books Rights Registry to track down owners, maintain a database of information about rights, and administer funds.  Something similar would likely be necessary here. 

Second, there is the issue of who is actually included in this class and who represents them. I covered some of this in my prior post about class certification. The short version is that these cases are brought by a small group of creators–in cases like Bartz, a group of three fiction and popular non-fiction authors—who have seemingly very little in common with your typical historian, or literary scholar, or law professor, or so on. Whether the court will allow a settlement to proceed with such limited class representatives—including ones who may have views on the underlying legal issues that are in direct opposition to many class members—is an important question about fairness.

Third, it seems likely that the AI defendants would push hard in such a settlement to cover not just past usage but future conduct—e.g.,for the development of new models based on the same content. This was one of the major problems with the Google Books settlement. In that case, hundreds of authors and academics objected (one notable example, from Pamela Samuelson, highlighting opposition to the class on issues of particular relevance to Authors Alliance members). The court ultimately rejected the deal, calling it “overreaching.” That proposed settlement would have given Google the right to license access to millions of books for future uses—including orphan works and out-of-print titles—creating dominant market advantages. 

Whether and to what extent a settlement would attempt cover future licensing is really unknown, and presumably the litigants have learned some of the same lessons we have from the Google Books settlement. But, its hard to imagine Anthropic or any other AI company passing up the opportunity to at least try to clear so many of its ongoing legal headaches in one swoop.

Could a Settlement Lock in a Training Data Monopoly?

Yes—and that may be the most consequential issue.

If only the largest AI companies settle—Anthropic, OpenAI, Meta—they could secure legal access to a massive corpus of copyrighted works. Even if a settlement didn’t cover future-looking uses, one that effectively granted a legal blessing over past use in developing existing models would mean that they had created a de facto compulsory license for themselves, unavailable to smaller competitors.

Entrants without the capital or clout to join similar settlements would be left behind, facing unclear legal risks or prohibitively expensive negotiations. The result? A skewed marketplace in which incumbents enjoy both technical and legal advantages, cemented not by Congress or regulation, but by court-approved private deals. 

Courts evaluating class settlements under Rule 23 aren’t required to perform an antitrust analysis, though there is an opportunity for the issue to be raised. In the Google Books Settlement context, several parties including the Department of Justice, which has substantial expertise and authority on this issue, did just that on antitrust grounds. Whether we will see today’s DOJ take the same stance is an open question. 

What It Means for Authors Who Disagree

Even authors who oppose the plaintiffs’ claims—or support AI developers’ fair use arguments—could be swept into a binding settlement unless they affirmatively opt out. Missing a notice deadline, misunderstanding legalese, or simply not realizing you’re part of the class might mean giving up your rights without ever choosing to.

This is especially relevant for authors who favor open licensing, or who already benefit from using generative AI in their own creative or scholarly work. A settlement could limit not only their ability to claim infringement in the future, but also the fair use arguments they support today. As Professors Deborah Gerhardt and Madelyn Wessel remind us, fair use is like a muscle, and particularly in the research and scholarly context, it can atrophy quickly if we give up on using it. 

Final Thoughts

The class action AI suits like Bartz v. Anthropic case is more than a copyright dispute—it’s a test of who—authors, publishers, big tech, Congress, or the courts—gets to define and control the intellectual building blocks of future technology that more than 200 million people now actively use every day. 

A settlement might resolve a legal standoff. But it could also determine the structure of the AI market, who controls access to training data, and whose voices—not to mention which words—constitute the future of machine learning.

For authors, this is a moment to watch closely. Not just because of what ends up in the settlement, but because of who ends up paying, monetarily and otherwise—and who gets to say no.