Anthropic AI: Seismic Settlement

The amount of media attention paid to generative Artificial Intelligence has increased tenfold over the last five years.  Much of this is no doubt fueled by the 40+ lawsuits that have targeted AI service providers since 2022.

But all of that has suddenly been overshadowed by the proposed settlement earlier this month of the class action case Bartz, et al. v. Anthropic PBC pending in the Northern District of California (Case No. No. 3:24-cv-05417). The terms of settlement include, astoundingly, creation of a fund of at least $1.5 billion to be distributed to members of the Class (this would be the largest award in a copyright infringement case in the history of the US), destruction of the datasets populated by works retrieved by Anthropic from two pirate sites, and a past release only for Anthropic conduct up to August 25, 2025.

This settlement comes on the heels of the Court’s Order on Anthropic’s Motion for Summary Judgment on the issue of fair use. To the benefit of Anthropic and other AI service providers, the Court held that scanning copyrighted material into a database for the purpose of training an AI engine was clearly defensible as a fair use. To the detriment of Anthropic and others who would emulate its penny-wise strategy, its actions in resorting to pirated copies of copyrighted works to populate its dataset was likely infringing, and so that question would proceed to trial.

It was probably in large part the burden, expense, and risk of a trial on the issue of Anthropic’s use of pirated copies that prompted it to agree to the extraordinary terms in the proposed settlement presently before the court.

The Class

As certified by the Court, the Class is limited to the actual or beneficial owners of timely registered copyrights in ISBN/ASIN-bearing books downloaded by Anthropic from the two pirate libraries - Pirate Mirror Library and Library Genesis. (The Class is limited to timely registered works so that all Class members have access to statutory damages. The restriction to ISBN/ASN-bearing books is necessary to enable association of each book with its corresponding copyright registration.) And, the Class is further limited to owners of the right to make reproductions, one of six separately denominated rights under the umbrella of copyright. The Class is not limited to authors, however, but instead includes everyone who owns the reproduction right. This might include both the publisher if relevant rights have been assigned to it, but in those cases would also include authors as beneficial owners because of their right to receive royalties on proceeds realized by the Publishers from commercial exploitation of the right assigned.

The Settlement Terms

As noted above, the terms of settlement require Anthropic to fund $1.5 billion for distribution to members of the Class. This represents $3,000 per work for each of an estimated 500,000 qualifying works. (Although Anthropic downloaded as many as 7 million copies from the two pirate sites, after eliminating duplications and cutting the total down to timely registered works bearing ISBNs/ASINs, we are left with approximately half a million works, with some work still to be done to settle on a final list.)

How Do You Know If You Are a Class Member

This was an “opt-out” class action, so you do not need to do anything to be a member of the class. But if you are a publisher or author of a book or books that might have been available from one of the two pirate sites noted above, you might be. And to facilitate your timely receipt of notices relating to your participation in the suit (including the opportunity to opt out of it), you should provide your current contact information and book titles to the court-appointed class counsel via the Class website at: www.anthropiccopyrightsettlement.com/register

You will need to provide your name, the title(s) of your book(s) and their ISBN(s), the name(s) of the publisher(s), your postal address, your personal email address, and information on your loan-out company if you use one.

Caveats

Two caveats, as of this writing.

First, in a Sunday order on September 7, the judge expressed his disappointment that counsel had left important questions to be answered in the future, including questions about the final composition of the Works List, the Class List, the Claim Form, and, particularly for works with multiple claimants, the processes for notification (for opt-out, so-called re-inclusion, and claims, whether a given choice is exercised by one, some, or all coclaimants), allocation, and dispute resolution. It is expected that proposals on these important points will be made with input from representatives of the authors and publishers, the Association of American Publishers, the Authors Guild, and selected industry experts, then challenged by Anthropic, then approved by the Court, all by October 10.

Second, as noted at the top of this article, this is but one case among many - sixteen brought by writers, sixteen by actors and YouTubers, five by music companies, two by artists, and one by programmers. At least one of those cases has produced a contrary result on the fair use issue addressed by the Court in Anthropic. In Kadrey v. Meta (Case No. 3:2023-cv-03417), another judge in the Northern District of California reluctantly held for the defendant, Meta – that its use of copyrighted works to train its AI engine was a fair use, but noted that it reached this result only because the 13 plaintiffs in that case had failed to develop a record in support of the contention that Meta’s unauthorized use substantially diminished the market for their works (the most important of the four factors considered in any fair use analysis). This was not a class action, and so the result binds only the thirteen plaintiffs in the case and points the way to others to achieve a different result.

So, stay tuned. Odds are that this will get interesting before it gets predictable.