• by Wendy Davis , Yesterday

Anthropic has agreed to settle a class-action complaint by authors who claim the company wrongly trained its large language model on copyrighted books.

Attorneys for the company and authors on Tuesday informed U.S. District Court Judge William Alsup in San Francisco of the proposed settlement. Terms have not yet been disclosed.

If accepted by Alsup, the deal will resolve a lawsuit dating to last August, when the authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson alleged in a class-action complaint that Anthropic built its business on "largescale copyright theft."

Anthropic allegedly digitized books it had purchased, and also downloaded digital copies of books from piracy sites, in order to train its large language model.

The artificial intelligence company argued to Alsup that using books for training purposes is fair use -- regardless of whether the books were purchased legally or downloaded from piracy sites.

Alsup issued a mixed ruling in June, when he held that said Anthropic's use of books it had purchased was transformative, and therefore a "fair use" that didn't infringe copyright.

But he also ruled that the company was not entitled to claim fair use regarding millions of books it downloaded from piracy sites. Alsup said in a subsequent ruling that the authors' complaint could proceed as a class-action on behalf of copyright owners of books in two libraries -- LibGen and PiLiMi -- containing pirated books.

Those rulings, taken together, exposed Anthropic to the prospect of billions in damages.

Anthropic last month asked Alsup to allow an immediate appeal of the ruling depriving the company of a fair use defense for pirated books. The company also petitioned the 9th Circuit Court of Appeals to hear a fast appeal of Alsup's decision to certify a class of copyright owners.

Anthropic and the authors reached a settlement before Alsup or the 9th Circuit ruled on those motions.