News · 2026-07-20

Judge Grants Final Approval to Anthropic's $1.5 Billion Book-Piracy Settlement

A federal judge gave final approval to Anthropic's $1.5 billion settlement with a class of book authors, entered judgment, and dismissed the case with prejudice. Judge Araceli Martinez-Olguin also ordered Anthropic to destroy the pirated book files at the center of the case within 30 days. This closes one of the largest copyright disputes in AI history, but it settles a dispute rather than deciding the underlying legal question of whether training a model on copyrighted text is lawful.

Key facts

• The number: a non-reversionary $1.5 billion fund, plus interest, approved on July 20, 2026.
• Who and where: Judge Araceli Martinez-Olguin in the Northern District of California, in Bartz v. Anthropic.
• Fees: counsel asked for $187.5 million; the court awarded about $101.6 million (~6.8% of the fund).
• Primary source: the court's own final approval order.

The case grew out of how Anthropic assembled its training data. To understand why this matters, you need one distinction the court drew a year ago. In June 2025, Judge William Alsup issued a fair-use ruling that split the behavior into parts: using copies to train a specific model was fair use, and digitizing books the company had lawfully bought into an internal searchable library was fair use too. But downloading pirated books from shadow libraries like LibGen to build a permanent, general-purpose collection was not fair use, and the judge held that all four fair-use factors favored the authors. The settlement resolves that piracy-and-retention piece.

Think of it like the difference between borrowing a book to study from and keeping a stolen copy in your basement forever: the court treated acquiring and hoarding the library as its own act, separate from whatever the model later learned. That is why the money attaches to the pirated corpus, not to the act of training.

What the approval actually does is mechanical but consequential. The settlement agreement sets a payment schedule: a $300 million installment due within five business days of final approval, and two $450 million installments due within 12 and 24 months of preliminary approval. The court awarded counsel $101,561,111 plus $2.64 million in expenses, held back an $18.22 million cost reserve, granted $15,000 service awards to each of three named representatives, and withheld 10% of the fee pending a final accounting. Within 30 days of judgment, Anthropic must destroy the original LibGen and PiLiMi files and everything copied from them, then certify it did so. The court declined to order model deletion, output attribution, or new licensing schemes, calling them outside the settlement's scope.

Why it matters: the settlement puts a concrete, enormous price on building a training corpus from pirated sources, which is a different and narrower question than whether AI training infringes at all. The release covers past claims tied to the Works List through August 25, 2025, and explicitly excludes claims about the model's outputs, future conduct, and works not on the list. The agreement states plainly that it is not an admission of liability and not a license to torrent, scan, or train on copyrighted works.

Participation was high. As of April 16, claims covered 440,490 of 482,460 Works List titles, about 91.3%. There were 350 timely opt-outs covering 1,802 works and 54 objections or comments; the court overruled the objections and allowed two late opt-outs for excusable neglect. Those figures show broad buy-in and genuine dissent at the edges, not unanimity.

The honest caveat is about what this does not decide. This is a Rule 23 approval, fee, and judgment order, not appellate precedent and not a ruling that training is fair use. One point of confusion worth flagging: a widely shared Authors Guild statement welcoming the deal is dated September 25, 2025, and addresses preliminary approval, not today's final order. The Guild says it disagrees with the fair-use holding and is looking to other cases to test AI outputs that compete with authors' own work. For the wider fight over whether models can be trained on copyrighted material, this is the end of one chapter, not the book.

Primary source, verified: read the paper →

Ground Truth. (2026, July 20). Judge Grants Final Approval to Anthropic's $1.5 Billion Book-Piracy Settlement. Ground Truth. https://groundtruth.day/news/anthropic-1-5-billion-book-settlement-final-approval.html

@misc{groundtruth:anthropic-1-5-billion-book-settlement-final-approval,
  title  = {Judge Grants Final Approval to Anthropic's $1.5 Billion Book-Piracy Settlement},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/anthropic-1-5-billion-book-settlement-final-approval.html}
}

Topics: ai-law · copyright · anthropic · settlement · policy