Bartz v. Anthropic: The Line Between Learning and Piracy

Three authors — Andrea Bartz, Charles Graeber and Kirk Wallace Johnson — sued Anthropic in 2024, alleging their books had been copied to build and train its Claude models.⁴ The case matters to anyone who assembles a training corpus because it produced the first substantive US ruling on whether that copying is fair use, and then the largest copyright settlement in the country's history — and the two point in opposite directions. Training on books was held lawful; the way Anthropic obtained many of them was not.

What you should be able to judge after reading this: why the same conduct can be both "exceedingly transformative" and a nine-figure liability, why the settlement leaves the most important question for builders unanswered, and what actually changes about how a lawful corpus is sourced.

What Judge Alsup actually held

Fair use is a statutory defence. Section 107 of the Copyright Act tells a court to weigh four factors: the purpose and character of the use, the nature of the work, the amount used, and — the factor that decides most modern cases — "the effect of the use upon the potential market for or value of the copyrighted work."¹ No factor is dispositive; the court weighs them together.

On 23 June 2025, Judge William Alsup split the conduct into three parts and decided each separately.² First, using the books to train the models was fair use: the court called the purpose "exceedingly transformative," reasoning that a model learning statistical structure from text to generate new text is closer to how a writer learns from what they read than to republishing the book. Second, digitising books Anthropic had lawfully purchased in print — destroying each physical copy as it was scanned — was also fair use, a format shift that added no new copies to the world. Third, and decisively, downloading pirated copies from shadow libraries to build a permanent central library was not protected, and that question was headed for trial.²

That distinction is the holding's centre of gravity. The transformative nature of training did the work of clearing the first hurdle; the source of the copies did the work of creating the liability.

Why the number is $1.5 billion

The settlement's size is not a measure of the authors' lost sales. It is a function of statutory damages. US copyright law lets a plaintiff elect a fixed penalty per infringed work — $750 to $30,000 for ordinary infringement, and up to $150,000 where the infringement is wilful — regardless of proven harm.⁷ Court filings indicated Anthropic had downloaded more than seven million books from pirate sites; after certification, roughly 482,460 works qualified for the class.⁷ Multiply that class by the wilful maximum and the theoretical exposure runs past $70 billion — a bet-the-company number even after a fair-use win on training.

After Judge Alsup certified a class of rightsholders whose works came from shadow libraries, the parties settled the piracy claim rather than try it.⁴ The $1.5 billion fund, net of fees and costs, works out to roughly $3,100 per claimed work — a small fraction of the statutory ceiling, but spread across hundreds of thousands of titles.⁷ By the May 2026 fairness hearing the claims rate had reached about 92.8 per cent of eligible works, with only a few hundred opt-outs.⁵

1. 2024

   Complaint filed⁴

   Three authors sue over books used to build and train Claude.

2. Jun 2025

   Summary judgment²

   Training and lawful digitisation are fair use; pirated copies are not.

3. Jul 2025

   Class certified⁴

   A class of rightsholders whose works came from shadow libraries.

4. Sep 2025

   $1.5B settlement⁷

   The piracy claim is settled rather than tried, and preliminarily approved.

5. May 2026

   Fairness hearing⁵

   Final approval taken under submission, with fee questions outstanding.

The matter is not closed. The case passed to Judge Araceli Martínez-Olguín for final approval, and at the 14 May 2026 hearing she took it under submission — pressing mainly on attorneys' fees (an objection put the multiplier near 6.9 times lodestar) and on the cost reserve — rather than approving from the bench.⁵ As of this writing the court had ordered supplemental briefing on a handful of late opt-outs and had not yet entered a final approval order.⁶

What the settlement does not settle

A settlement resolves liability; it does not make law. The piracy question that was headed for trial will now go unadjudicated, so the only binding doctrine to come out of Bartz is the summary-judgment holding: training can be fair use, acquisition can still infringe.⁷ Two live uncertainties sit on either side of it.

The gap between two judges in the same courthouse is the warning: the fair-use win rests on a market-harm theory a better-argued plaintiff may yet dislodge. On the other side, Anthropic's own certification narrows the win further. The company states the settlement releases only legacy training claims, not any claim based on model outputs, and certifies that it did not use the LibGen or PiLiMi material in commercial models.⁸ Output liability and web-scraped data remain untested.

What an engineer should take from it

The practical lesson inverts the headline. "AI training is fair use" is the quotable line, but the enforceable rule that emerged is narrower and more operational: provenance is the liability. The same model, trained on the same text, is defensible or ruinous depending on whether the copies were lawfully acquired.² For anyone assembling a corpus, three things follow.

1. Treat acquisition as the control point

   Keep records of where each part of a corpus came from and on what licence. A shadow-library shortcut converts a fair-use defence into a statutory-damages exposure measured per work.⁷

2. Do not read the win as covering outputs

   The settlement expressly leaves output-based claims alive; a model that reproduces protected text is a separate risk the training holding does not touch.⁸

3. Assume the market-harm theory is unsettled

   A future plaintiff who documents market dilution, as Judge Chhabria invited, could reach a different result on the same training conduct.³

Go deeper

Read the statute first: 17 U.S.C. § 107 is one paragraph, and the four factors are the frame every one of these cases argues within.¹ Then follow the primary record on the docket — the summary-judgment order, the class-certification order, and the settlement papers are all filed under No. 3:24-cv-05417 in the Northern District of California.⁹ To see how contested the reasoning is, read the Bartz and Kadrey orders side by side: same court, same month, opposite emphasis on market harm.³ Watch next for the final-approval order in this case, and for the first plaintiff to bring the market-dilution record Chhabria said was missing.⁶