$1.5B Anthropic Settlement: Books, Fair Use & Payouts

The settlement

On September 5, 2025, Anthropic agreed to pay $1.5 billion to end Bartz v. Anthropic, a class action brought by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson. The plaintiffs' law firm, Susman Godfrey, called it the largest copyright recovery in United States history. [1][3][8]

Atlas interpretation: The size of the figure obscures what it was actually paying for. A federal judge had already ruled, three months earlier, that training a language model on books was lawful. What Anthropic paid for was how roughly half a million of those books ended up on its servers in the first place. [1][4]

Two separate findings, one settlement

On June 23, 2025, Judge William Alsup of the Northern District of California granted Anthropic partial summary judgment. He ruled that training Claude on lawfully purchased books was fair use, calling the use "exceedingly transformative" and comparing it to how a reader becomes a writer. He found all four statutory fair-use factors favored Anthropic on that specific question. [4]

The same order denied Anthropic summary judgment on a separate practice: downloading millions of books from pirate sites, including Library Genesis (LibGen) and the Pirate Library Mirror (PiLiMi), to build a general-purpose library it could draw on for training. Alsup wrote that "pirating copies to build a research library without paying for it...was its own use, and not a transformative one," and sent that question to trial rather than ruling for either side. [4][3]

Atlas interpretation: The training ruling drew most of the initial coverage because it was the first judicial statement that feeding books to a language model can be fair use. It was the piracy finding that actually cost Anthropic money: the company settled two months before the piracy trial, once its exposure was millions of documented unauthorized downloads rather than a contested legal theory about transformation. [4][1]

The payout and how authors collect it

The settlement covers an estimated 500,000 works, which puts the payout at roughly $3,000 per work before legal and administrative costs. Money is split between an author and publisher under a default 50/50 formula for trade and university-press books unless a contract specifies otherwise; self-published authors and those holding full rights keep the entire amount. The deadline to file a claim was March 30, 2026. Named plaintiff Charles Graeber, for example, was reported to be owed about $3,100 for each of his two qualifying books, including his bestseller The Good Nurse. [2][8]

The class is defined narrowly: a work must carry an ISBN or ASIN and be registered with the U.S. Copyright Office either within three months of publication or before August 10, 2022, the date Anthropic's disputed downloading was alleged to have occurred. Anthropic also agreed to destroy the original pirated files and any copies derived from them. [2][3]

From preliminary approval to final sign-off

Judge Alsup granted preliminary approval of the settlement on September 25, 2025, but did not see it through. He moved to inactive status in his final months on the case, and after filing a December 2025 memo objecting to what he called "huge multipliers" in the fee request from several plaintiffs' firms, the case was reassigned to Judge Araceli Martinez-Olguin. [6][7]

At a fairness hearing on May 14, 2026, the court heard from seven objectors out of 54 objections and comments filed, on issues including publisher favoritism, adequacy of compensation, and class counsel conflicts of interest; Martinez-Olguin did not rule from the bench that day. She issued a 23-page order granting final approval on July 20, 2026, cutting the attorneys' fee award to about $101.5 million (down from an initial request as high as $300 million) and reducing the named plaintiffs' service awards from $50,000 to $15,000 each. [5][6]

What the settlement leaves open

Atlas interpretation: Because Anthropic settled rather than take the piracy claim to trial and a possible appeal, the June 2025 fair-use ruling never reached an appellate court and sets no binding precedent outside this case. Copyright suits over AI training against Google, Meta, OpenAI and Midjourney were still pending as of mid-2026, and each of those cases is free to produce its own answer on whether training itself is lawful use. [5]

Atlas interpretation: The settlement also does not resolve the tension it was meant to close for authors. Andrea Bartz said after final approval that the fair-use rationale for training still troubled her because the resulting models can compete directly with the writers whose books trained them, telling reporters, "the algorithm is being used to essentially try to put us out of a job." And a settlement built on U.S. copyright law has no reach over AI developers training models outside U.S. jurisdiction. [8]