After two and a half years since the first lawsuit against AI companies was filed in January 2023, we had in 1 week the first 2 announced settlements (pending court approval).
First, the parties in the copyright suit filed by voice actors Vacker against ElevenLabs announced a settlement after mediation (the role of ADR we will analyze in a subsequent post).
Then, the parties in the copyright suit filed by 3 book authors Bartz in a class action certified by Judge Alsup against Anthropic announced a settlement in principle after mediation.
We don’t have the settlement details yet. And, the lawsuit against Anthropic will be subject to the requirements of Rule 23(e) for Judge Alsup’s application and approval. Based on how the case has transpired, approval seems likely, provided the terms of the settlement don’t look like a bad deal for the plaintiffs’ class.
The Domino Effect
There’s still a lot left for the class settlement to be approved–there’s no guarantee it will.
But let’s assume it is approved. Will it have a domino effect on the other copyright lawsuits?
Looking at the map of the 42 other lawsuits still pending, one might think a domino effect won’t happen. After all, these cases involve different companies, different types of works, different judges, and potentially enough differences in the training of the AI models to matter. Plus, the 2 district court decisions in Bartz v. Anthropic and Kadrey v. Meta both found the AI training part of the respective case was a fair use–albeit with other parts of the decisions that were quite unfavorable to the defendants. With those fair use precedents, the defendants in the other cases might feel they have a decent leg to stand on in their litigation.
But, on the other hand, there are plenty of reasons why we could see at least the cases against OpenAI and Microsoft play out in a similar fashion to the Anthropic case. Meta has thus far escaped a class action (removing one big element of risk from their lawsuit), and Google may stand in a different position in terms of the datasets it used to train its model. I focus on the cases against OpenAI and Microsoft below.
1. If class actions are certified, they put into play “doomsday” amounts of statutory damages.
The biggest pressure on the AI companies to settle is the amount of statutory damages in play. Because the companies all trained their AI models on millions of copyrighted works, that makes the number of works potentially infringed (assuming no fair use) at astronomical levels. In Anthropic’s suit, the number was some number under 7 million (from 2 different pirated books datasets).
The number might have been much lower, let’s say, 10,000 works. But we might not even find out a precise number now given the announced settlement. Using a range of 10,000 – 6 million works, we asked ChatGPT to run a Monte Carlo simulation to project the probabilities of what Anthropic faced. Most scenarios looked bad, and many looked potentially business-ending or bankruptcy-forcing to try to discharge the liability obligation and start fresh.
2. For the LLMs of openai, microsoft, and meta, the companies allegedly used some of the same shadow libraries.
Another big reason why we may see a domino effect in the other cases is that OpenAI, Microsoft, and Meta are all alleged to have used some of the same pirated books datasets from shadow libraries, and, if my memory is correct, are all alleged to have used torrenting.
In fact, it was revealed in a filing that Benjamin Mann took part of the downloading from shadow libraries first at OpenAI in 2019 when he worked there and then later in 2021 when he co-founded Anthropic. Potentially a doubly whammy, in effect.
3. if other courts follow Judge alsup’s approach to shadow libraries instead of Judge chhabria’s, fair use won’t be a defense.
Of course, one big variable is whether courts adopt the approach of Judge Alsup in treating the downloading from shadow libraries.as a separate use, for the purposes of fair use analysis. Judge Alsup held that Anthropic’s downloading was for the purpose of building a library–and that’s not fair use. By contrast, Judge Chhabria held that Meta’s downloading was for the further fair use purpose of training its AI model Llama.
These different approaches might not be inconsistent because the facts related to what each respective company did could have been different.
And one difference between Anthropic and OpenAI is that Anthropic kept the copies it downloaded–a fact found by Judge Alsup that he emphasized in his order, whereas OpenAI reportedly destroyed the Books1 and Books2 datasets after using them to train its earlier model(s). Thus, even under Judge Alsup’s approach, it’s at least possible that OpenAI is in a stronger position for fair use (arguing no library building but instead all training).
4. the new york times lawsuit against openai and microsoft is not a class action but it, alone, involves over 13 million works.
Even if no class is certified in the cases against OpenAI, it also faces the suit filed by New York Times that involves 13 million articles (the exact amount of works would be reduced for each daily compilation consisting of various articles). On top of this lawsuit, other news media have sued OpenAI. So, OpenAI already faces “doomsday” statutory damages even without class certification.
5. The lawyers for the book authors v. anthropic also represent book authors in the other lawsuits
Last but not least, the book authors in the other suits are represented by some of the same attorneys, especially the ones from Susman Godfrey.
In fact, the lead attorney in the Bartz lawsuit v. Anthropic is Justin A. Nelson. Nelson is the interim lead class counsel for book authors in the MDL litigation against OpenAI and Microsoft. Plus, Susman Godfrey lawyers represent the New York Times.
So, we should fully expect that Nelson and the other plaintiffs’ attorney will use the same AI playbook against OpenAI and Microsoft they used against Anthropic. Because some of the facts related to the pirated books datasets are common, there is at least a decent chance the same dynamics will play out in the MDL litigation for the class plaintiffs.
Here’s Nelson after securing a settlement of $787.5 million from Fox News in the suit filed by Dominion Voting systems:
But will plaintiffs demand too much money from openAI and microsoft?
One of the biggest countervailing factors to any settlement in the MDL litigation against OpenAI and Microsoft is simply that, given the 13 or so lawsuits, collectively the plaintiffs demand an astronomical amount of money to settle the case.
In other words, the plaintiffs’ demands in the MDL Litigation could be “too big to settle.”
ChatGPT agrees with our analysis
For what it’s worth, ChatGPT largely agrees with our analysis. It found an increased probability of settlement in light of Anthropic’s settlement, with the highest increase in the Authors Guild and book authors’ cases (15 – 20 percent boost), which are the most similar in facts to Bartz v. Anthropic. ChatGPT provided descriptive reasons to support its analysis not included here.
3 responses to “Will Anthropic’s announced settlement of book authors’ class action copyright suit have domino effect on other lawsuits? Reasons yes and no.”
To be eligible for statutory damages, the works must be registered with the copyright office.
Are the 13 million NYT articles registered with the copyright office? If not, they must argue loss of actual revenue as damages.
“The Times owns over 10 million registered, copyrighted works, including those set forth in the incorporated Exhibits (“Times Works”). See
ECF Nos. 170–202.” SAC para. 14.
Group Registration for Newspapers (GRNP):
The Copyright Office allows publishers like The New York Times to register an entire issue at once, rather than filing separate registrations for each individual article, photo, or graphic.
Each daily edition of The Times is typically registered as a “collective work.”
The registration covers both the compilation (the issue as a whole) and the individual contributions created by Times employees (articles, editorials, photos, etc.).
Freelance contributions may need separate registrations depending on contractual arrangements.