Anthropic and the book scanning machines: Then we’ll just eat them – culture

GA big relief for Anthropic in the summer of 2022. The AI ​​company was looking for new texts with which to train its models. The competition is ahead, things have to happen quickly. On July 1, 2022, a new, freely accessible digital internet archive containing millions of books will be released. The site’s admins write that they are “deliberately violating copyright law in most countries.” Ben Mann, one of the Anthropic founders, sent a link to the site to colleagues and commented: “Just in time!!!”

That’s what it quotes Washington Post from court documents from a trial that ended with a settlement on Monday. What was negotiated was Anthropic’s illegal use of such archives for training its AI. The company, now valued by investors at $965 billion, is paying $1.5 billion to some of the authors whose books gave it intelligence to its AI in the first place.

At times there were three-digit billion sums in the room

The now legally binding deal was announced in September, in time before a trial would have taken place in December. Anthropic absolutely wanted to prevent that. The statutory damage amounts in the US Copyright Act range between $750 and $300 per work, and up to $150,000 can be due for intentional infringement. Anthropic had downloaded more than seven million books. At times, three-digit billion sums were therefore available as compensation.

According to the authors’ lawyer, the $1.5 billion now agreed is still the largest amount ever paid in a copyright case. Around 500,000 rights holders joined the class action lawsuit. The amount corresponds to a payout of around $3,000 per book, a sum that is at the lower end of the legally defined range.

Some of the rights holders complained and demanded more money. Judge Araceli Martínez-Olguín still signed the agreement. Calls for more money for authors, she said, were “not based on a realistic overall assessment of the risks and rewards of a trial.”

During the course of the trial, Anthropic argued that the use for internal research purposes fell under the principle of “fair use” and was therefore permitted under US copyright law. The initially responsible judge, William Alsup, followed this argument. In a decision last June, he compared feeding the books to AI to giving them to schoolchildren to read. Copyright is not there to “keep authors away from competition,” not even from intelligent machines. The problem from his point of view, however, was the way in which the books got to the AI ​​models. Downloading the stolen books itself, regardless of how they are used, violates copyright law.

Millions of books are said to have already been shredded

Anthropic has therefore apparently switched to other methods of getting fresh reading material. Judge William Alsup already referred to this in his decision. In February 2024, Anthrophic commissioned Tom Turvey to “get all the books in the world”. The tech manager had built Google’s online library Google Books for Google, which was very close to the copyright limit. Now he has started buying books for Anthropic from all kinds of dealers, preferably one from each work, so that they can be processed in machines: the spines are removed, the pages are cut, scanned and then disposed of. Millions of copies would have already been turned into waste paper on the way to the artificial super brain.

Alsup’s statement goes on to say that Turvey tried at times to acquire usage licenses from publishers – that would have been the amicable way. But Turvey then let the conversations fall asleep. And started the book-eating machines instead.

A few days ago, several large publishers sued Google because the company used books from Google Books to train its AI Gemini. It is another of various processes that are taking place around the world around the topic of artificial intelligence.