Anthropic agrees to pay $1.5B US to settle author class action over AI training

Anthropic told a San Francisco federal judge on Friday that it has agreed to pay $1.5 billion US to settle a class-action lawsuit from a group of authors who accused the artificial intelligence company of using pirated copies of their books to train its AI chatbot, Claude, without permission.

Anthropic and the plaintiffs in a court filing asked U.S. District Judge William Alsup to approve the settlement, after announcing the agreement in August without disclosing the terms or amount.

“If approved, this landmark settlement will be the largest publicly reported copyright recovery in history, larger than any other copyright class action settlement or any individual copyright case litigated to final judgment,” the plaintiffs said in the filing.

The proposed deal marks the first settlement in a string of lawsuits against tech companies including OpenAI, Microsoft and Meta Platforms over their use of copyrighted material to train generative AI systems.

As part of the settlement, Anthropic said it will destroy downloaded copies of books acquired through pirating sites LibGen and PiLiMi (Pirate Library Mirror). Under the deal it could still face infringement claims related to material produced by the company’s AI models.

In a statement, Anthropic said the company is “committed to developing safe AI systems that help people and organizations extend their capabilities, advance scientific discovery, and solve complex problems.” The agreement does not include an admission of liability.

“This historic settlement is a vital step in acknowledging that AI companies cannot simply steal authors’ creative work to build their AI just because they need books to develop quality LLMs,” Authors Guild CEO Mary Rasenberger said in a statement.

“These vastly rich companies, worth billions, stole from those earning a median income of barely $20,000 [US] a year. This settlement sends a clear message that AI companies must pay for the books they use just as they pay for the other essential components of their LLMs.”

Although an estimated seven million books were downloaded by Anthropic from piracy sites, according to the Authors Guild, only around 500,000 works are covered in the class action, meaning the settlement amounts to roughly $3,000 US per author. 

Writers Andrea Bartz, Charles Graeber and Kirk Wallace Johnson filed the class action against Anthropic last year. They argued that the company, which is backed by Amazon and Alphabet, unlawfully used millions of pirated books to teach its AI assistant Claude to respond to human prompts.

Creative work stolen

The writers’ allegations echoed dozens of other lawsuits brought by authors, news outlets, visual artists and others who say that tech companies stole their work to use in AI training.

The companies have argued their systems make fair use of copyrighted material to create new, transformative content.

Alsup ruled in June that Anthropic made fair use of the books to train Claude, but found that the company violated their rights by saving more than seven million pirated books to a “central library” that would not necessarily be used for that purpose.

A trial was scheduled to begin in December to determine how much Anthropic owed for the alleged piracy, with potential damages ranging into the hundreds of billions of dollars.

The pivotal fair-use question is still being debated in other AI copyright cases. 

Vancouver author J.B. MacKinnon recently launched class-action lawsuits against NVIDIA, Meta, Anthropic and Databricks Inc. in B.C. Supreme Court, alleging that his and other Canadian authors’ works have been used illegally for AI training. 

Another San Francisco judge hearing a similar ongoing lawsuit against Meta ruled shortly after Alsup’s decision that using copyrighted work without permission to train AI would be unlawful in “many circumstances.”