Humanity pays authors $1.5 billion to resolve copyright infringement lawsuits

NEW YORK (AP) – Artificial intelligence company Anthropic has agreed to pay $1.5 billion to settle a class action lawsuit by the author of a book that says the company received a pirated version of their work to train chatbots.

The groundbreaking settlement, if approved by a judge on Monday, could mark a turning point in the legal battle between AI companies and writers, visual artists and other creative experts denounce copyright infringement.

The company has agreed to pay the author or publisher about $3,000 for each of the estimated 500,000 books subject to the settlement.

“As far as we can say, it's the biggest copyright restoration ever,” said Justin Nelson, the author's lawyer. “This is the first time in the AI ​​era.”

Author Trio – Thriller Novelist Andrea Burtz and Non-Fiction writers Charles Greber and Kirk Wallace Johnson – I sued last year And now we represent a wider group of writers and publishers whose books were downloaded to train chatbot Claude.

A federal judge dealt with the case June mixed judgmentdiscovered that it is not illegal to train AI chatbots on copyrighted books, and that humanity mistakenly acquired millions of books through pirate websites.

If humanity hadn't been resolved, experts say they lost the lawsuit after it was scheduled for trial in December.

“We've seen strong potential for billions of dollars, which is enough to potentially crippling or even go out of business,” says Thomas Long, a legal analyst at Wolters Kluwer.

San Francisco US District Judge William Alsup is scheduled for a hearing Monday to consider the terms of the settlement.

Humanity said in a statement Friday that if the settlement is approved, it will “settle the remaining claims of the plaintiffs' estate.”

“We are continuing to be committed to developing secure AI systems that help people and organizations expand their capabilities, promote scientific discovery and solve complex problems,” said Aparna Sridhar, the company's assistant advisor.

As part of the settlement, the company also agrees to destroy the original downloaded files.

Books are known To be an important data source (essentially billions of words carefully staked), it is necessary to build a large language model of AI behind chatbots like Anthropic's Claude and its major rival, Openai's ChatGpt.

Alsup's June ruling found that humanity had downloaded more than 7 million digitalised books, which he “knows have been pirated.” It started with nearly 200,000 from an online library called Books3, assembled by AI researchers other than Openai, to match the vast collection that ChatGPT trained.

Bartz's debut thriller novel “The Lost Night,” by Bartz, the main plaintiff in the case, was found in the data set.

Alsup later ingested at least 5 million copies from the Pirate Website Library Genesis or Libgen, and at least 2 million copies from Pirate Library Mirror.

The author guild told thousands of members last month that it expected “loss is at least $750 per work, which could be much higher.” A high-settlement award – about $3,000 per job – could reflect a small pool of books affected after extracting duplicates and non-copyrights.

On Friday, Author Guild CEO Mary Raysenberger said the settlement was “a great outcome for authors, publishers and right-wingers, sending a strong message to the AI ​​industry, robbing the author's work to train AI, and robing those who can't afford it.”

The Danish Alliance of Rights has managed to defeat one of these shadow libraries, saying on Friday that the settlement would be of little use to European writers and publishers whose work is not registered with the US Copyright Office.

“On the other hand, editing AI training datasets by downloading millions of books from known illegal file sharing sites is priced,” said Thomas Heldrup, director of content protection and enforcement for the group.

Meanwhile, Heldrup said it fits into a tech industry playbook for growing the business first and then paying relatively small fines to break the rules compared to the size of the business.

“We understand that these companies view settlements like humanity as the price to run business in a highly competitive space,” Heldrup said.

Founded in 2021 by a former Openy leader, the privately held humanity valued it to $183 billion earlier this week after raising another $13 billion in investment.

Humanity also expects to earn $5 billion in sales this year, but like Openai and many other AI startups, it doesn't report profitability, so instead relies on investors to support the high cost of developing AI technology to expect future payoffs.

A settlement may affect other disputes, including ongoing litigation Authors and newspapers for Openai and its business partners Microsoft, and Case for MetaAnd Mid Journey. And as the terms of human settlement were filed, another group of authors sued Apple in the same San Francisco federal court on Friday.

“This shows that in other cases, creators and AI companies can essentially reach settlements without breaking in court,” says law analyst Long.

Industry including humanity were largely praised Alsup's June ruling. This is because they discovered that by training AI systems with copyrighted works, chatbots are “typically transformative” and that they can create passages of texts that are certified as “fair use” under US copyright law.

Comparing the AI ​​model to “readers aiming to become writers,” Alsup writes that humanity “trained his work to race, recreate, and compensate first, but then turn the hard corner and create something else.”

However, documents disclosed in court show internal concerns among human employees regarding the legality of the use of pirate sites. The company later shifted its approach and hired former Google executive Tom Turvey, who will be in charge of Google Books. Year of copyright battle.

According to court documents, humanity, with his help, bought a large number of books, tore the bindings, and began scanning each page before feeding each page to an AI model. The judge said it was legal but did not reverse previous copyright infringement.

– –

In an earlier version of this story, The Associated Press misreported the name Wolters Kluwer Legal Analyst. His name is Thomas Long, not William.