Anthropic settlement a big step for AI law

There was huge news last week in the colliding worlds of copyright protection and development of artificial intelligence. The artificial intelligence developer, Anthropic, recently settled a very high profile copyright suit brought as a class action by book authors. This lawsuit has been described by Wired as the “most significant artificial intelligence copyright lawsuit in history.” The results of an adverse decision against Anthropic would have wiped it out.

The issues the court grappled with were partially decided on summary judgment; Anthropic both won and lost. Anthropic, the developer of Claude, an artificial intelligence service, had copied sets and subsets of millions of books in order to train large language models to power Claude’s artificial intelligence service. The books used were both purchased and taken from pirate sites on the Internet; the purchased books were digitized by Anthropic if acquired in hard copy.

There are few, if any, cases that squarely address these issues. The U.S. District Court for the Northern District of California had to deal with issues which appeared to be novel when applied to artificial intelligence, but nevertheless rely on basic copyright law and the doctrine of fair use as codified in the Copyright Act.

All of the books used to train Claude were available for purchase. Anthropic did not have the plaintiffs’ consent to use any of the books plaintiffs authored for the purpose of training Claude. Anthropic had undertaken talks with various publishers in order to obtain licenses to use the books to train Claude, but those discussions were terminated and Anthropic embarked on a plan to buy some books and to find other books on pirate websites. It used all of these books, bought and stolen, for training purposes. From the books acquired Anthropic’s engineers selected sets and subsets they felt were best for training Claude. Each book used was copied multiple times, some text was removed and Claude literally memorized those sections of books that were offered to it for this purpose. Many of the books were copied in their entirety.

The books that were acquired were then placed in a digitized library, some for later use as some of these books were not yet used to train Claude. When confronted with the possibility that Anthropic had committed copyright infringement by copying these books, purchased and pirated, it raised the defense of fair use, an affirmative defense meaning it acknowledged the copying but denied infringement occurred, claiming its use was fair so that copying these books would not be copyright infringement.

In the Ninth Circuit fair use is a question of law for the judge with factual questions determined by the trier of fact, although this case never got beyond summary judgment, which was granted and denied in part by the trial judge. The judge ruled that Anthropic’s actions in using books that it had purchased to train Claude was “exceedingly transformative “ and a fair use. Using books that it had pirated from websites without purchase, the judge held was ‘’irredeemingly infringing “. As to those books not yet used to train Claude, but to create a library, the digitization of those books that were purchased was also considered a fair use, and the digitization of the pirated books was not considered a fair use.

The judge carefully examined section 107 of the Copyright Act which provides four non-exclusive factors to apply to determine if a use is fair. They are: purpose and character of the use; the nature of the work; the amount and substantiality of the portion used compared to the entire work; and the effect of the use on the potential market for the work.

As to purpose and character of the use, the court held that no portion of the books were seen by Claude’s end users, and that no output provided to Claude’s users infringed the books’ copyrights. It held “the purpose and character of the use was transformative-spectacularly so.” Transformative means that the new work does not take the place of, or substitute for, the old, but becomes itself altogether another work. The court ruled that the plaintiffs could not exclude anyone from using their books for training or learning purposes if the books were purchased, and that the training was generative and transformative because Claude‘s product resulted in new content. Using the purchased books for training purposes was a permissible fair use under the Act.

Adding books to Claude’s library, even if not used yet for training, was not infringing if the books were purchased. Anthropic had the right to dispose of each copy of a purchased book as it wanted, and that included a mere change in format from hardcopy to digitization. However, it was a different story regarding the more than 7 million books that had been pirated in order to train Claude or to add to Claude’s library.
Merely having in mind the purpose of future research was not talismanic, turning infringement into fair use, because taking these books without payment would, the court held, destroy the academic publishing market.

The partial grant and denial of summary judgment seems to be pretty well reasoned, but of course it was not the final word on whether using copyrighted works, properly acquired to create AI services is fair use as this was only the decision of a trial court and is not binding on any other court, although it might provide some helpful guidance to another court grappling with the same issues. There was going to be a trial on the issue of whether using pilfered copies of books would be infringing or a fair use because the book authors had not moved for summary judgment on this issue and it remained open for determination, although the judge had made up his mind that these uses were infringing.

The terms of the settlement have not been released and they will not be finalized until early September. This was a major test case for the AI industry. This resolution or the court’s rulings are not binding on any other court or any other parties engaged in publishing or development of artificial intelligence but as to the question of stolen works, it feels right and creates a huge risk for those AI developers who rely on non purchased books to train their AI services.

It would be prudent for all publishers to join together and do what ASCAP and BMI did for music publishers-sell licenses for the use of content to train AI services and divide the royalties among the copyright holders, thereby permitting the use of their books to train artificial intelligence without the threat and uncertainty of copyright litigation. At some point the industry has to deal with these issues and understand that payment must be made for the use of protected content to train AI, whether in the form of an outright purchase, or a license for this use.

Jim Astrachan is a counsel to Corey Tepe LLC and has taught intellectual property law in the two Maryland law schools since 1999.