Inside Anthropic's $1.5B Book-Piracy Settlement

On July 20, 2026, a federal judge signed off on Anthropic paying $1.5 billion to authors whose books it pulled from pirate sites to train Claude. The court called it the largest copyright class action settlement in history.

The case is Bartz v. Anthropic, filed in August 2024 by three authors: thriller novelist Andrea Bartz, plus Charles Graeber and Kirk Wallace Johnson. What started as three names became a class of roughly 482,460 works.

The twist that makes this worth an engineer's attention: Anthropic actually won the central legal question. It paid anyway. Understanding why is a lesson in how the law treats your training pipeline, not just your model.

To build a training corpus, Anthropic downloaded more than 7 million books by torrenting them from two shadow libraries, Library Genesis (LibGen) and the Pirate Library Mirror (PiLiMi). Separately, and legally, it also bought and scanned millions of print books.

That split matters. The whole case turned on how the data was acquired, not whether learning from copyrighted text is inherently illegal.

Two paths to the same tokens, two very different legal outcomes. The pipeline's provenance became the liability.

In June 2025, Judge William Alsup issued a split summary-judgment ruling. Training an LLM on legally acquired books he called "quintessentially transformative" and squarely fair use.

The copies torrented from LibGen and PiLiMi were a different story. Alsup found them "inherently, irredeemably infringing" and set a trial on piracy liability for December 1, 2025.

So Anthropic had the precedent it wanted on training. But the acquisition of the corpus was a separate, losing act, and that is what dragged it toward trial.

In August 2025, Alsup certified a class, one he defined himself rather than one the plaintiffs asked for, covering every eligible copyright owner in the pirated datasets. That converted a manageable dispute into exposure across nearly half a million works.

Statutory damages under 17 U.S.C. §504(c) run $750 to $30,000 per work, and up to $150,000 per work if infringement is willful. Multiply that ceiling across the class and the theoretical exposure lands around $72 billion.

Against a number that could end the company, $1.5 billion is a controlled loss. Winning the fair-use argument did not remove the risk that a jury would call the torrenting willful.

The fund works out to about $3,000 per work, roughly 4x the statutory minimum for ordinary infringement. By default each work's share splits 50/50 between author and publisher, adjusting where contracts say otherwise.

Anthropic also has to destroy the original pirated files from LibGen and PiLiMi, and any copies derived from them, within 30 days of final judgment, with written certification to class counsel.

The deal did not go through cleanly. At a September 8, 2025 hearing Alsup skewered it, said he felt misled, and demanded a full list of the pirated books before he would move. "We'll see if I can hold my nose and approve it," he said. Weeks later, after supplemental filings, he granted preliminary approval, adding, "I don't see how you can get a better deal".

This settles the past, not the future. It covers conduct only through August 2025, and Judge Araceli Martínez-Olguín, who granted final approval, was explicit that it does not release claims based on the output of AI models.

It also grants Anthropic no license to train on these books going forward. The court overruled all 54 objections, including pleas to add source attribution or force model deletion, calling those remedies outside the scope of the suit.

Objectors also flagged a counting quirk: group registrations can bundle many novels under one "work," so the class list may undercount the real number of books affected. A one-time payment, they argued, poorly matches a model that keeps earning.