Last week, the first two US court rulings on AI were handed down in cases where the plaintiffs were authors claiming their copyrighted works were infringed upon. While the same issue lies at the heart of both cases—is it fair use for AI companies to train on copyrighted works?—and both judges found in favor of AI companies, the rulings are quite different from one another.
The two cases in question were both heard in the US District Court for the Northern District of California. They are:
- Bartz v. Anthropic | Judge Alsup | Authors Guild explainer | Authors Alliance explainer
- Kadrey v. Meta | Judge Chhabria | Authors Guild explainer | Authors Alliance explainer
The Authors Guild’s dissection of both rulings is readable and helpful, although keep in mind the Authors Guild has their own AI case pending on behalf of authors. For a balancing perspective, I’ve included links to the Authors Alliance, an organization that prioritizes broad access to information and education. (Its values are in line with that of the Internet Archive.)
What distinguishes the Anthropic case? The judge had to determine whether Anthropic’s training of its AI chatbot, Claude, on 7 million books constituted fair use. Anthropic built its library on pirated books and by purchasing print books that it digitized. The judge ruled that Anthropic’s acts of training were fair use (legal), but that it was not legal for Anthropic to knowingly download and copy pirated works for its own training library. For that reason, the case will move to trial in December to determine damages to be paid to authors for infringement. Each act of infringement carries a minimum penalty of $750 and maximum penalty of $150,000—mind-blowingly large figures, considering the number of works involved. However this case is not yet class certified; a decision on that is still pending. If it is not class certified, only the named authors will receive damages.
What distinguishes the Meta case? Again, the judge had to determine whether Meta’s training of its AI models constituted fair use; Meta also used pirated materials. Meta won, but as the Authors Guild points out, “only on technical grounds, a matter of procedure, not on the merits of the law.” Due to the lack of evidence presented by the author-plaintiffs, the judge could not rule in their favor, even though he found Meta’s fair-use arguments unpersuasive. In his ruling, the judge basically invites someone else to bring a better case that could be won, as he doesn’t see how Meta’s use is fair.
Given the many other cases still pending, these rulings are by no means the final word—only the opening salvo. In May, the Authors Guild summarized the ongoing legal cases relevant to authors, although at least one more case has been announced since.
On Monday, I interviewed Authors Guild CEO Mary Rasenberger and chief legal officer Kevin R. Amer about the two cases and what might happen next. Note that either side could settle before a trial comes to pass, or one side could appeal before or after the trial.
The following interview has been edited for length and clarity.
Jane Friedman: Let’s talk about the Anthropic case first. Given the quantity of titles in play—7 million—those are a lot of damages to manage as far as getting the money to the authors. Can you speak to how that would unfold?
Rasenberger: So that’s assuming class certification is granted. If it isn’t granted, then they’re only going to be looking at the named authors’ books. Assuming class certification is granted, then it’s 7 million. … I think there are duplicates, and there are probably some in the public domain. But it’s still going to be millions of books and billions in statutory damages. So in a class action, there are these firms and this is what they do. They’re brought in and work with the court and the parties, and they send out notices. Then they handle distribution of the damages also, generally speaking. … It’s a lot of work, but it’s doable.
In the Anthropic case, I thought it was interesting the judge said it was okay [fair use] for print copies to be digitized for AI training, and that seems …
Rasenberger: Crazy, right?
Well, yeah. [laughs]
Rasenberger: We think there are some serious errors in this case. We think it’s winnable on appeal. There is, you know, a lot of case law saying that’s just not true. Now most of that case law falls in the Second Circuit, but the same rationale would apply. There’s cases like the Internet Archive case, which is about books. There was a case that got settled that publishers brought against Audible. ReDigi was a service that was saying that you can resell your digital files … but it involved copying. And the court said, “No, you can’t. That’s copying.”
The Meta judge, I will say, makes the same mistake [as the Anthropic judge]. They basically say, “Well, it’s just format shifting, so there’s nothing wrong with that.” Well, that [indicates] a total misunderstanding of how the industry works. You sell your rights separately. When you sell your print rights to someone, you don’t think you’re selling your digital rights to them too. And if they want digital, they better pay for it. [Otherwise] it really undermines the market.
So let’s bring in the Meta case. It does seem like these two rulings are somewhat in conflict with each other despite their finding of fair use. What do you think authors need to know about how these judges differ and how that might resolve?
Rasenberger: For one thing, the Anthropic judge says that if you’re getting your books by pirating them, that’s piracy. That’s infringement. That’s not fair use. And the Meta judge just completely discounts the fact that they were getting the books from pirate websites, which is very strange because it’s criminal-level piracy. It just makes no sense.
[Editor’s note: Chhabria’s ruling says, “The plaintiffs are wrong that the fact that Meta downloaded the books from shadow libraries and did not start with an ‘authorized copy’ of each book gives them an automatic win. … But Meta is also wrong to suggest that its use of shadow libraries is irrelevant to whether its copying was fair use. … The law is in flux about whether bad faith is relevant to fair use.”]
Amer: On the market harm analysis, they really are in pretty direct conflict. Because the Anthropic judge made this analogy to human learning and said that what an AI system is doing is really no different than, you know, if you had some impossibly efficient person who could read millions of books and then turn around and create their own works. We think it just completely disregards the practical realities of how these things work and how they can harm the market as a whole.
The Meta court went completely in the other direction, and we think correctly said this is nothing like human learning. There’s no human, obviously, who can ingest millions and millions of works and within seconds just spit out works in the same style, and [the judge] accepted the argument that this is potentially really devastating to the writing profession. What [the Meta court] said was, “The plaintiffs in this case didn’t provide sufficient evidence about how that affected their works.” And so it ultimately ruled for Meta, but we were really pleased that the judge understood the real market harm that’s happening here.
Yes, the Meta judge [Judge Chhabria] seemed really critical of the case that was brought, like he was inviting someone else to bring a better case.
Rasenberger: The decision is odd in that sense because he basically says this should not be fair use and here’s why. And he really focuses on, as Kevin explained, the dilution to the marketplace, which devalues every single work and therefore decreases the incentives to create. … That’s exactly what we’re concerned about with AI, is the saturation of the marketplace, as well as the kinds of books that directly compete with specific books or with specific authors, like the kind of experience you had. We’re still seeing those even though Amazon is trying to keep them off their site. And then there’s of course lost licensing income.
Judge Chhabria says, “You guys focused all your arguments on this lost licensing income, going back and forth about whether there’s a licensing marketplace.” And he sort of says that there isn’t [a marketplace], even though there is. … In both the Meta case and Anthropic, the judges dismiss that part of the market harm analysis completely, preemptively. They just say authors don’t have a right to that licensing income. So Judge Chhabria said, “Well, you guys spent all your time worrying about this licensing income. That wasn’t the harm here. The harm here is dilution to the marketplace.”
He also dismisses the harm from actual substituted books. Maybe he didn’t have enough examples of it. I don’t know, but he doesn’t talk about it. He said, “Plaintiff, you did not give me the evidence I wanted.” And he talked about this at the hearing. … “This does not sound like fair use to me. I don’t get this,” he said to the defendants. And then to the plaintiffs’ attorneys, “Where’s that evidence I need? Why didn’t you submit a report?”
So I thought, based on that, he well could have said, “I can’t decide summary judgment. I can’t decide the issue of fair use of training on summary judgment because I don’t have the evidence that I need.” And then put that off to trial. But he didn’t do that. Instead, he scolded the attorneys and basically says anybody else can bring a case who provides this evidence that I just totally laid out for you that you need to provide.
The lawyers will say, “Well, we tried to provide that evidence,” but it’s very hard to prove, to provide evidence of future market harm, right? We haven’t seen the flooding of the marketplace yet. So what you need then is an expert who can just lay out the argument that Chhabria makes in the introduction [to his ruling], which is, given the way the publishing industry works, and given how easy it is to create certain kinds of books that will necessarily compete with commercial, human-written books, this is the inevitable outcome.
We [the Authors Guild] will be building those kinds of reports. The problems with both the Anthropic and Meta cases is they hurried discovery along. And in Anthropic, the discovery wasn’t even done yet, and it’s really unusual to hear summary judgment when discovery is still open. So that is a reversible error also. Judge Alsup didn’t allow parties to complete discovery. … But we think he wanted to hurry the case up, and the same with Judge Chhabria in the Meta case. You know, they’re in the exact same courthouse, so it feels a little bit like a competition. [laughs]
Does any of this change the Authors Guild case to the extent you’re able to talk about it?
Rasenberger: No, we feel like we’ve got a really, really good case there. We thought long and hard about the case, our lawyers did, before bringing it. … It’ll be easier, particularly [since] the longer [we have] to get into the use of AI to generate books, the easier it is to establish the evidence that courts are looking for. And again, I think it’s an error on the part of the courts, because they’re supposed to be looking at potential harm to the market, not actual harm. But judges often have a really hard time with potential harm. Even if they’re saying potential, they want evidence of actual.
Will the Second Circuit [where the Authors Guild case will be heard] be looking at these early rulings when they’re making a determination?
Rasenberger: Well, they’ll be interested, I’m sure. It could impact them. I don’t know in what way it would, but they most certainly do not have to follow these cases because they’re in separate circuits. If they agree with the reasoning, they can adopt the reasoning, but they do not have to.
And if their rulings are in conflict, does that then set up some sort of showdown at the Supreme Court?
Rasenberger: Yes. So the Meta and Anthropic cases I think inevitably will go up to the Ninth Circuit [Court of Appeals]. The losing party’s going to appeal. And if the decisions are in conflict at the circuit level, then it is very ripe for a Supreme Court case, and it is a case that I think the Supreme Court should and would take. But I’m not convinced the Ninth Circuit is going to side with either of these decisions.
Bottom line: These early rulings indicate to me there will ultimately be a legal path to training AI models using copyrighted works; it’s just a matter of how much money rights holders will receive. If the Anthropic ruling were the last word on the matter, then it would be considered fair use for AI companies to purchase authorized copies and use them for training without payment or licensing. That would mean authors or publishers could do nothing to prevent AI models from training on their books. But such questions are far from settled.
For a long time, I’ve been emphasizing that the two cases to watch, out of more than two dozen, are the Authors Guild case as well as the New York Times case against OpenAI and Microsoft. Those cases have been consolidated with some others in the Second District in New York, and discovery continues in those cases. (It has been contentious.) But there’s now a third case to watch: the one Disney and Universal have brought against Midjourney. Publisher Keith Riegert recently commented on the importance of that case.
For further analysis of these rulings
- Read professor James Grimmelmann’s analysis of both cases on Bluesky
- Read professor Matthew Sag’s analysis of the Anthropic case ruling from Judge Alsup
- Bloomberg Law looks at the mixed Anthropic ruling, quoting both Grimmelmann and Sag
- A strong analysis of the Meta case ruling from Judge Chhabria by Ashley Belanger at Ars Technica
Jane Friedman has spent her entire career working in the publishing industry, with a focus on business reporting and author education. Established in 2015, her newsletter The Bottom Line provides nuanced market intelligence to thousands of authors and industry professionals; in 2023, she was named Publishing Commentator of the Year by Digital Book World.
Jane’s expertise regularly features in major media outlets such as The New York Times, The Atlantic, NPR, The Today Show, Wired, The Guardian, Fox News, and BBC. Her book, The Business of Being a Writer, Second Edition (The University of Chicago Press), is used as a classroom text by many writing and publishing degree programs. She reaches thousands through speaking engagements and workshops at diverse venues worldwide, including NYU’s Advanced Publishing Institute, Frankfurt Book Fair, and numerous MFA programs.
Good interview. It’s interesting that the judges have a hard time assessing potential market harm while acknowledging that no human comparison to AI exists, in terms of capacity to produce competitive product—their dilemma makes me wonder if a private concern is that the horse is out of the barn. Time will tell. Thanks Jane.