Anthropic Copyright Settlement Pays Authors, Keeps Fair Use Win

A record payout that does not change the rule

The largest copyright recovery in U.S. history is now final, and it looks different depending on who you ask. On July 20, 2026, U.S. District Judge Araceli Martínez-Olguín gave final approval to Anthropic's $1.5 billion class settlement with authors and publishers, saying it gives them "meaningful relief." The money covers pirated books that Anthropic used to build its Claude chatbot, at roughly $3,000 per title.22 Plaintiffs' attorney Justin Nelson called it the largest known copyright recovery in history.22

The central legal question came out differently. Anthropic still holds the ruling that training AI on books it legally acquired is fair use, and the company keeps saying so. Deputy general counsel Aparna Sridhar repeated the point after approval: "Training AI on books is fair use under copyright law."1 NPR's follow-up coverage put it bluntly: the case could be seen as more of a win for AI companies than for publishing. Within the narrow terms of this lawsuit, the court never found it illegal for Anthropic to train on copyrighted books as long as it paid for them.1

My view is that NPR has it right. The settlement priced a past mistake in how Anthropic got its data. It did not set a toll for using books to train AI. For the companies being sued, and for the lawyers advising them, that difference matters more than the headline number.

How the case split

Andrea Bartz, Charles Graeber and Kirk Wallace Johnson filed the class action in 2024. They accused Anthropic of feeding millions of digitized copyrighted books into the models behind Claude.2 In June 2025, Judge William Alsup split the case in two. He held that training was fair use, calling it "exceedingly transformative." But he sent the claims over pirated books from Library Genesis and Pirate Library Mirror toward trial.2 Court records show Anthropic downloaded at least 5 million books from LibGen in 2021 and at least 2 million from PiLiMi in 2022.10

The pirated-copy claim is what forced a settlement. Willful infringement can bring statutory damages of up to $150,000 per work, and the judge found more than 7 million pirated copies. The potential trial exposure was enormous.2 Reuters reported that the liability could have reached a trillion dollars.15 Coverage disagrees on when the deal was struck. NPR reported in early September 2025 that the parties had reached a settlement2, while a later Reuters overview says Anthropic settled in December.15 Either way, the order of events is clear: Anthropic won on training and paid to end the piracy problem.

Alsup granted preliminary approval and then retired, so final approval fell to Martínez-Olguín.22 Under the deal, Anthropic also has to destroy its pirated dataset. The settlement covers only past training data, not what Claude produces.12

Why $3,100 a book feels smaller than $1.5 billion

The big number shrinks once it is divided up. After costs, the court awarded class counsel about $101.6 million, down from the $187.5 million they requested, which was 12.5% of the fund.2928 One analysis put the fee at roughly $2,953 an hour across the attorney time the court credited. Meanwhile the named author plaintiffs had their service awards cut from the requested $50,000 to $15,000 each.26 Writer Beware calculated about $1.38 billion left for rightsholders across 440,490 claimed works, or about $3,127 per work.

That per-work amount is often split again. When a publisher still holds an exclusive license, the default is a 50/50 split between author and publisher. Only sole rightsholders, such as self-published authors or authors whose rights have reverted, get the full amount.21 Writer Beware argues that publishers are the real winners. It notes that one house, Bloomsbury, has 14,087 titles in the class, which is worth roughly $21 million at half the estimated payout. When claim notices went out in early September, many authors found publishers claiming books whose rights had already reverted, publishers claiming 100% instead of 50%, and literary agencies filing claims even though agents generally don't own rights in a book.28

Graeber, one of the lead plaintiffs, says he is owed about $3,100 for each of his two books, including The Good Nurse. He also told NPR that after more than two years of travel and turned-down work, he is "much poorer for this settlement, ironically."1 Payment timing is still unsettled. Writer Beware reports that several appeals of the final approval were filed, while South Korean press reported that Anthropic plans to start paying rightsholders on November 15.5

Coverage generally agrees that claim participation was high. The court and Anthropic cite about 91% of covered works claimed22, and one law-firm analysis puts the rate near 93%.18 Reach has also turned out to be global. Settlement lists include works by Nobel laureate Han Kang, with at least 22 Korean-authored titles identified among the 482,460 covered works.10

The doctrine is still split

For lawyers on both sides, the settlement's most important effect may be what it prevented. Authors Alliance noted objectors' complaint that the settlement ends any chance of an appeals court reviewing Alsup's fair-use ruling, whichever way it might have gone.23 That leaves trial-court decisions in the same district pointing in different directions.

In Kadrey v. Meta, Judge Vince Chhabria also ruled for the AI company but reasoned differently. He treated Meta's shadow-library downloads as part of the training purpose rather than a separate infringing use, and decided the case on market harm, which the plaintiffs had not proved.17 He also warned that AI training "in many circumstances" would not be fair use and that generative AI could flood the market with competing content.15 Alsup, by contrast, dismissed that market-dilution theory, comparing it to complaining that teaching schoolchildren to write well creates competition.11 Bartz has made the dilution argument directly: "The algorithm is being used to essentially try to put us out of a job."1

In practice, legal analysts agree that where the data came from now matters as much as what the model does with it. Firms advising developers say that companies training on lawfully obtained copies are on much firmer ground than those that built libraries from pirated sources.18 That explains Anthropic's change in approach. After the dispute grew, it started buying large numbers of physical books in 2024 under an internal effort reportedly called the "Panama Project," with plans to destructively scan them.5

The legal-tech exception

The AI case most relevant to the legal industry went against the AI company. On September 30, 2026, the Third Circuit affirmed that Ross Intelligence infringed Thomson Reuters' copyrights by training its legal research tool on Westlaw headnotes. The court held that the material was copyrightable and that Ross's use was not fair.14 The panel stressed that Ross wanted to be a direct competitor and used the material for a very similar purpose.14 Reuters called it a first-of-its-kind appellate ruling.13

How far the ruling reaches is disputed. Ross's tool was not generative AI; it was a case-search product built to compete with its source.12 Even an AI-industry group read the decision as implicitly confirming that highly transformative, general-purpose generative training is likely fair use.13 That reading has merit, but the warning for legal-tech builders is plain. Training a narrow tool on a competitor's editorial content in order to replace that competitor is the weakest fair-use position there is. Vendors that build research, drafting or summarization products on licensed legal databases should treat Ross as binding risk, not as an outlier.

What comes next

The next tests are already scheduled. Music publishers in a second lyrics case against Anthropic are seeking about $3 billion, and their claims draw on discovery from Bartz that turned up more than 20,000 song-related files.16 Cross-motions for summary judgment in the consolidated OpenAI litigation in New York were filed on September 4, 2026.17 Abroad, Europe's rules are stricter. The EU AI Act blocks training on works whose rightsholders have opted out, and a Munich court sided with the music rights society GEMA against Suno.5

The Authors Guild sees licensing as the long-term answer, but its policy director acknowledges that such deals remain rare while the major AI companies are tied up in litigation.1 That is the gap the Anthropic settlement leaves open. It made piracy expensive without making training expensive. Until an appeals court rules on generative training, or Chhabria's market-dilution theory wins at trial, AI companies have a fairly clear path: buy the books, scan them, and argue fair use.