Anthropic’s $1.5 billion settlement could set precedent for other AI copyright suits

Anthropic Co-Founder & CEO Dario Amodei speaks onstage during TechCrunch Disrupt 2023 at Moscone Center on Sept. 20, 2023 in San Francisco. Anthropic is the first major AI company to reach a settlement in multiple lawsuits by authors alleging tech companies used copyrighted material to train their systems. (Photo by Kimberly White/Getty Images for TechCrunch)

A $1.5 billion proposed settlement due to be paid by AI company Anthropic could set a precedent for how other AI companies navigate the training of their models on copyrighted material.

The proposed settlement is a result of a class action lawsuit, Bartz v. Anthropic, from last year, in which authors filed a complaint against Anthropic for using copyrighted materials to train its large language models (LLMs), including Claude. 

It’s a common practice for many large AI companies, said Yelena Ambartsumian, AI governance attorney and founder of tech consultancy firm Ambart Law.

“[American companies] made a bet that, instead of paying for this content, licensing it, which would cost billions of dollars, the bet was okay, ‘we’re going to develop these LLMs, and then we’ll deal with the fallout, the lawsuits later,” Ambartsumain told States Newsroom about the issue last month. 

In June, San Francisco-based federal Judge William Alsup delivered a nuanced ruling on the suit, saying that Anthropic’s use of the books for training was legal under the fair use doctrine, which allows limited use of copyrighted material for educational or creative purposes, but Anthropic’s copying and storage of more than 7 million pirated works infringed on the author’s rights. Anthropic was ordered to destroy the original files of the downloaded materials and any copies, the ruling said. 

Anthropic announced the settlement this week, aiming to avoid a trial set to take place in December. The $1.5 billion settlement comes from projected payments of $3,000 per work for a list of about 500,000 works that have been identified.

But Alsup shared concerns about the settlement in a hearing on Tuesday, saying it was lacking details about the claim process for affected authors and a thorough list of works covered by the deal. The missing details could leave Anthropic open to future litigation by more authors, and leave affected authors out of the loop for receiving what they are owed, he said. 

He’ll be revisiting the settlement in another hearing on Sept. 25, and asked for a “drop-dead list” of total books by Sept. 15.

“His concern is, okay, we have this huge settlement fund, but we haven’t even identified who the claimants are,” Ambartsumian told States Newsroom after the settlement was announced. 

In his ruling, Alsup did not find fault with the training methods of the models, so Anthropic will not have to delete or alter their large language models themselves, just the copies of the original works they were trained on, Ambartsumian said. Users of Claude or its other models will likely not feel the effects of the ruling, she said, and the company probably won’t take a huge business hit because of it, given the company’s multi-billion dollar annual revenue.

But Anthropic will be the first to feel the costs of the bet it — and other large AI companies — made in their strategies for training. It’s the first case to reach a settlement among a collection of copyright lawsuits facing AI companies like OpenAI, Microsoft and Meta. 

The costs associated with training and keeping up with existing LLMs have deterred many companies from developing their own models, Ambartsumian said. In addition to bracing for legal fees like Anthropic’s, many companies feel that if they released a new tool, an existing AI company would likely release their own, more popular version shortly after. 

Though Anthropic’s case would likely set precedent for how other large tech companies’ copyright cases can be handled, the details of the settlement still need work, Ambartsumian said. 

“It’s the $1.5 billion question,” she said. “The class was certified, which was a big deal. It’s a really large hurdle in these cases, but now the question is, we certified the class but we don’t know who they are.”