Earlier this year, author Maureen Johnson was feuding with Anthropic.
Specifically, I was struggling with Anthropic’s copyright settlement website.
Johnson is the author of 18 books, most of them for young people and many of them best-sellers. The artificial intelligence company Anthropic owes him about $3,000 per book (which will be split 50/50 with his publisher) for several of them. The payments are part of a first-of-its-kind settlement last fall in which Anthropic admitted it downloaded millions of pirated and copyrighted books to train its AI models without the authors’ permission. (According to the New York Times, “As part of the agreement, Anthropic said it did not use any pirated works to create artificial intelligence technologies that were made public.”) A judge determined that using those books without the author’s permission constituted fair use, but piracy did not. Similar lawsuits are pending against Meta and OpenAI. (Disclosure: Vox’s Future Perfect is funded in part by the BEMC Foundation, whose primary funder was also an early investor in Anthropic; they have no editorial input on our content.)
- Anthropic owes a class of half a million authors $1.5 billion in legal settlement for downloading pirated books to train its AI model.
- However, Anthropic’s data set had so many errors that the authors found it difficult to navigate the website created to manage the claim.
- Furthermore, that $1.5 billion represents a very small amount for each individual author in the class, especially after they have split the payment with their publishers.
- The deal will go to court for a fairness hearing on May 14.
The class action lawsuit aimed to level the playing field between individual authors and one of the most valuable companies in the world. To distribute the money to the perpetrators, Anthropic and the plaintiff’s attorneys worked with a claims administrator (a company that specializes in handling compensation claims) to create a website that the perpetrators can use to access a small portion of the record $1.5 billion payout.
But Johnson, like other authors who spoke to Vox, quickly found a problem: The claims site is buggy and unreliable, forcing people to jump through endless hoops to collect the money they’re owed. By March, he had already submitted applications for his 14 eligible titles twice, spending 90 minutes each time painfully filling out the forms.
Now, the claims administrator was telling him that they couldn’t find any of his entries. She was taken through several levels of management, each repeating the same thing.
“It was becoming more and more surreal how little this system worked,” Johnson said.
Eventually, Johnson connected with an employee who she said spent the entire call laughing. He told her he had found her first claim filing from February, but not the new one.
“This system is really complicated,” Johnson said. “It’s just not programmed well.”
In response, Johnson said the employee laughed again. “Coding is hard,” he told her.
Johnson is not alone in her frustrating experience. The authors had six months to register their claims for payment from Anthropic, and many of them had difficulty doing so.
Anthropic regularly touts its ethical and philanthropic bona fides. (The company is here to serve the long-term well-being of humanity! It is the safe and responsible AI company! Claude helped NASA’s Perseverance rover travel to Mars!) But the good it is doing is based on stolen work, and the people who created that work are having trouble getting the small resource they are owed.
“Everyone agrees it’s not the best data.”
All popular big language models were trained on books; that was the only way to get them enough high-quality text to start generating their own. Most of those books were downloaded from pirate libraries, at least in one case on the grounds that it would simply be too expensive to pay for each title. As it became increasingly clear that this was the case, class action lawsuits began to arrive.
Bartz et al. v. Anthropic CBP was the first to be resolved. In September 2025, a judge approved a $1.5 billion settlement between Anthropic and nearly half a million writers it had determined belonged to the class. However, things got complicated when it came time to determine who those half a million writers were.
They had to be authors of books that appeared in one of the three hacked databases Anthropic used in 2021. But trying to create a complete list from those databases proved difficult. Anthropic had not created its own records as it introduced pirated books into its training corpus, so lawyers for both sides had to rely on the pirate sites’ own data. And they had to do it quickly, because the trial had strict deadlines.
“It’s like metadata from crowdsourced pirate libraries,” Dave Hansen, executive director of the advocacy group Authors Alliance, told Vox. (Authors Alliance has filed amicus curiae briefs in the bartz case and published extensive technical explanations for the authors). “I wouldn’t rely on that for much of anything, much less managing legal claims in a large, important lawsuit. But that was the best I had been given to the data sources that were being used.”
“I think everyone agrees that it’s not the best data, but it’s the best they could do in the time frame,” publishing industry journalist Jane Friedman told Vox. “I think it was the reality for the class attorneys. The judge was really expediting matters, so they did the best they could in the time they had.”
Neither Anthropic, its attorneys, the class attorney in this case, nor the claims administrator responded to a request for comment from Vox. But it appears that the plaintiff’s attorneys and the claims administrator worked together to narrow down Anthropic’s initial list of 7 million books to only titles that were under U.S. copyright in 2022.
“Then they used a bunch of other industry sources to enrich that data so they had more information about the current publishers, and then they used that to generate contact information,” Hansen said. “At that scale, it’s really difficult to achieve 100 percent accuracy.” He added: “One of my biggest criticisms of how this deal and process has gone is the data. They just haven’t been very transparent about it.”
From there, the group’s claims administrator and attorneys used that shaky list to create their buggy website, which is how Maureen Johnson eventually found herself on the phone with a man who laughed and told her that coding was hard. Other authors were in a similar situation.
“I have 19 titles in the database,” said Christopher Moore, author of quirky comic novels like Lamb: The Gospel According to Biff, Christ’s Childhood Best Friend. After he had done paperwork for 18 of them, he had to step away from his computer. When he returned the next day to finish the paperwork for book 19, everything had been deleted.
A month passed after he submitted the form the second time, Moore said. “And I got another notice: What about these other titles?” Most of the titles belonged to one of the four other Christopher Moores who worked as authors. One was actually his, Moore said, “but he showed it with some weird Texas copyright.” He filed the claim anyway and is still waiting to hear back.
April Henry, who writes YA mysteries, also found unusual copyright holders on her books. “One of the books on the list appeared to be an audiobook and showed the narrator as one of the copyright holders,” he said.
Meanwhile, he’s struggling to figure out how to handle seven of his 22 books that he wrote with a co-author. “No one ever included in their contract that they would split the rights to reach a legal agreement,” Henry said. “You know what I mean?”
And as authors struggle to navigate the claims process, they do so with mixed emotions.
“That’s not much for their entire catalogue.”
Johnson is still furious about her experience with the claims manager website. “Your AI monster ate all our work,” he said, addressing Anthropic. “Now you’re trying to pay us back with this […] piece of garbage that doesn’t work.”
For many authors, the money did not seem enough, considering that their life’s works had been stolen without their permission. The total deal of $1.5 billion seems like a lot. But divided among so many copyright holders, it doesn’t go that far. There’s also the fact that the $3,000 figure is just an estimate of what authors’ payments will eventually look like. There is actually a set amount of cash available for the class, and the more people participating in the class, the smaller the pool of money available to everyone involved.
“When you think that $3,000 per book times 22 books, you think, ‘I’m getting $66,000,’” Henry said. But then there’s the money that goes to publishers and the money that goes to a co-author. “In some cases, it will end up being like $500 per book,” Henry said. “At first you think, ‘What a windfall!’ But it doesn’t seem like a windfall.”
“To me, it’s quite a career and it will come down to less than $30,000,” Moore said. “That’s not much for their entire catalogue.”
Then there’s the question of what the new world Anthropic helped build with all those stolen books will be like for authors. “We have no idea what the long-term damage this will cause to artists,” Moore said. “I’m on the downhill slope of my career, so there’s not much they can take away from me. But if someone is strong in the middle of their career, this could really hurt them.”
On May 14, the settlement will receive a fairness hearing, where the judge will review a series of complaints from the authors, including what they describe as “inadequate compensation in relation to the harm.”
Meanwhile, Anthropic remains one of the biggest players in technology, currently valued at $900 billion. According to industry headlines: “Anthropic’s Claude makes his way to the top of the AI market.”
Correction, May 6, 2:00 p.m.: This story originally misstated the number of books Maureen Johnson has written; there are 18.