Unbiased

Anthropic's $1.5 Billion Piracy Settlement Does Not Touch Its Training Ruling

Judge William Alsup had already ruled that training Claude on lawfully bought books was fair use and only its pirated library was not, leaving open whether the roughly $3,000 a work the settlement set for that piracy should govern AI training broadly or only the stolen copies that produced it.

9 min read2026-08-14

The Split That Survived a $1.5 Billion Settlement

On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of a $1.5 billion settlement in Bartz v. Anthropic PBC, the largest copyright class-action settlement on record by the court's own repeated description.¹ The non-reversionary fund covers 482,460 books that Anthropic downloaded from the LibGen and Pirate Library Mirror shadow libraries and that meet the class's registration criteria.¹ The order is still inside its own appeal window, so distribution to authors has not begun.¹

The number is bigger than the question it answers. In June 2025, Judge William Alsup split the conduct behind it in two. Training Claude's models on books Anthropic had lawfully bought or already owned, he ruled on summary judgment, was "exceedingly transformative" and a fair use under all but one of copyright's four factors.² Those factors come from the statute both sides argue over, 17 U.S.C. § 107.³ They weigh the purpose and character of a use, the nature of the work, how much of it is used, and the work's effect on its market. Acquiring and permanently retaining more than seven million pirated books from Books3, LibGen, and the Pirate Library Mirror to build that same training set was not a fair use on any factor, Alsup ruled in the same order.² He ordered a trial on damages.

How Judge Alsup's order split the same training program into a win and a loss, factor by factor.²
Section 107 factor	Training on lawfully bought books	Building the pirated library
Purpose and character of the use	Favors fair use: "exceedingly transformative"	Against fair use
Nature of the work	Against fair use, the one factor training lost	Against fair use
Amount and substantiality used	Favors fair use	Against fair use
Effect on the market	Favors fair use	Against fair use

That is the split the settlement bought its way out of, and only half of it. The order's own language draws the line precisely. Class members "give up all claims they could have brought against Anthropic broadly related to past AI inputs." But they "do not give up claims about past AI outputs, nor claims of any kind about future conduct."¹ The settlement also requires Anthropic to destroy the pirated files it downloaded, according to the plaintiffs' own account of its terms.⁴ What $1.5 billion bought was the piracy claims. It did not touch, and could not undo, the ruling that training itself was lawful.

That ruling is not the industry's last word, because the same legal question is being argued in a second courtroom right now. In January 2026, the music publishers Concord Music Group, Universal Music Publishing Group, and ABKCO sued Anthropic for more than $3 billion. They allege it copied roughly 21,000 songs' lyrics by torrenting them, the same acquisition method behind the Bartz piracy claims.⁵ On August 3, 2026, Anthropic filed a motion to dismiss two of the publishers' claims but left the torrenting and training claims themselves in place.⁵

Both sides also lean on an institutional analysis that is not itself settled law. The U.S. Copyright Office's report on generative AI training walks through the same four factors, concluding that training's transformativeness is often real but "a matter of degree," and that copying from pirate sources to compete with the works copied "is unlikely to qualify as fair use" when licensing is reasonably available.⁶ The report was issued as a "pre-publication version" in May 2025.⁶ The Register of Copyrights who signed it, Shira Perlmutter, was removed from her post the next day.⁷ Each side of this fight cites that sequence selectively, and neither treats it as the whole story.

Bartz v. Anthropic PBC N.D. Cal. · No. 4:24-cv-05417-AMO

Stage: Final approval granted July 20, 2026. Still inside its Ninth Circuit appeal window; distribution has not begun.¹
Question: Whether the per-work price the piracy claims produced should set the rate for AI training generally, or only for stolen copies like the ones that produced it.
Stakes: $1.5B resolves the piracy claims only. Claims over AI outputs and future conduct stay open, and the same legal question is live in a second, $3 billion suit.⁵

What is actually contested, then, is not whether training on lawfully acquired work is fair use. One judge has already answered that. It is whether the price a court just calculated for stolen books, roughly $3,000 per work and built on statutory damages meant to punish piracy, is the number that should govern how every other AI training run gets priced and permitted from here.

Compensation and Consent

The roughly $3,000 per work this settlement extracted for piracy should be the model for pricing every unlicensed use of protected work in AI training, because payment and consent are the only approach that has made an AI company answer for taking work it did not buy.

Anthropic did not lack alternatives. The same order it now cites for its training win found the company "'ha[d] many places from which' it could have purchased books." It preferred to steal them to avoid "legal/practice/business slog," as cofounder and chief executive officer Dario Amodei put it.² Between 2021 and 2022, it downloaded more than seven million pirated books from three shadow libraries to build the central library that fed training.² Every fair-use factor went against that conduct, and had the piracy claims gone to trial rather than settled, Anthropic faced statutory damages of up to $150,000 per work for willful infringement.⁸

Mary Rasenberger, CEO of the Authors Guild, reads the result as proof the approach works. She calls it an "excellent result for authors, publishers, and rightsholders generally, sending a strong message to the AI industry that there are serious consequences when they pirate authors' works to train their AI, robbing those least able to afford it."⁹ Maria Pallante, President and CEO of the Association of American Publishers, calls the piracy behind the case "abhorrent conduct that should never be normalized," and the settlement "an important victory in the larger battle to hold big tech accountable."¹⁰

The legal scholar Robert Brauneis argues the case for payment runs deeper than one company's bad faith. Copyright's limits on liability, he argues, presuppose the imperfect, effortful character of human memory and learning, a premise generative AI breaks by making and analyzing perfect copies at superhuman speed and scale.¹¹ The Copyright Office's own report quotes him more bluntly: training "transcends the human limitations that underlie the structure of the exclusive rights."⁶ If the analogy to human learning does not hold, the argument goes, neither does the license to train on someone else's work without paying for it, transformative purpose or not.

The strongest objection to pricing training this way is that no market exists to license it at scale. The plaintiffs' own economist, James Malackowski, argued the opposite in the same case: that a market for licensing books to train large language models was already forming. Alsup did not resolve which economist was right. He wrote only that "[o]ur record could support either account — so this order must assume Authors are correct," without it changing the outcome for lawfully acquired books.² The Copyright Office's own report documents licensing deals already struck in music, stock photography, and news.⁶ If a market can be built in those industries, the price a federal court already calculated is the one a book-licensing market has to clear.

Mary Rasenberger, CEO, The Authors Guild. She led the organization representing the authors named as class representatives in the settlement, and argues its price should extend beyond this one case.⁹

Transformative Use and Access

Training on lawfully acquired work is already settled fair use, and no license covers what a frontier model actually needs to read. Pricing training at the piracy settlement's rate would misprice a transformative use and leave the resulting AI industry to whichever company can absorb the largest bill.

Alsup's own words on training are unqualified: "The use of the books at issue to train Claude and its precursors was exceedingly transformative and was a fair use under Section 107 of the Copyright Act."² Anthropic's Deputy General Counsel, Aparna Sridhar, has repeated that holding since final approval: "Training AI on books is fair use under copyright law."¹² She adds that the settlement itself has already run its course, with "more than 91% of authors and publishers covered by the settlement" having "claimed their share of the payment."¹² On her account, the money closed the piracy claims. It did not put training back in question.

Mark Lemley and Bryan Casey made the access argument years before this case existed. Training data sets, they wrote, "are likely to contain millions of different works with thousands of different owners," so "there is no plausible option simply to license all the underlying [works] for the new use."¹³ Anthropic's own retained economist, Steven Peterson, reached the same conclusion from the company's side of this specific case: "the hypothetical competitive market for licenses covering data to train cutting-edge LLMs would be impracticable."¹⁴ A law professor with no stake in the litigation and an economist paid by one of its defendants arrived at the same conclusion from different directions.

The Copyright Office's own analysis, cited by both sides, places the disputed territory on a spectrum rather than a binary. Uses "for purposes of noncommercial research or analysis that do not enable portions of the works to be reproduced in the outputs are likely to be fair," it found.⁶ But "the copying of expressive works from pirate sources in order to generate unrestricted content that competes in the marketplace, when licensing is reasonably available, is unlikely to qualify as fair use."⁶ Training on lawfully bought books, on that framework, sits nowhere near the piracy pole the settlement priced. Pamela Samuelson, Christopher Jon Sprigman, and Matthew Sag, quoted in the same report, go further: deriving patterns from training data and using them to build something new is "not just transformative, it is highly transformative."⁶

Treating the piracy settlement's price as the going rate would fall hardest on whoever cannot absorb it. The venture-capital firm Andreessen Horowitz, in comments to the Office, warned that "imposing the cost of actual or potential copyright liability on the creators of AI models will either kill or significantly hamper their development." The result, it argued, would be "far less competition, far less innovation."⁶ A price built to punish willful theft, extended to lawful training generally, would not deter piracy again. It would set the entry fee for building a frontier model, payable only by whoever already has $1.5 billion to spend on one.

Aparna Sridhar, Deputy General Counsel, Anthropic. She is the company's own voice in the case it fought and partly won, and has kept stating the training holding as settled even as Anthropic pays to close the piracy claims.¹²

The Split That Survived a $1.5 Billion Settlement

Compensation and Consent

Transformative Use and Access

Sources