# What did the 2025 US court rulings say about training AI on books?

Date: 2025-07-02 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/2025-us-court-rulings-say-about-training/

**Answer:** In late June 2025, two federal judges in San Francisco ruled that training language models on authors' books was fair use in the cases before them, Bartz v. Anthropic and Kadrey v. Meta. But the reasoning differs sharply, Anthropic still faces a trial over pirated copies, and the Meta judge stressed that his ruling does not make such training lawful in general.

Within three days, two judges of the US District Court for the Northern District of California ruled on the same question in AI cases: can a company copy books to train a language model without asking the authors? On June 23, 2025, Judge William Alsup ruled in *Bartz v. Anthropic*. On June 25, Judge Vince Chhabria ruled in *Kadrey v. Meta*. Both decisions went mostly the companies' way on training. Read closely, though, they point in different directions, and neither one is the final word.

This is an explainer of what the two orders say, not legal advice. Both are trial-court decisions, so they bind only the parties in those cases and can be appealed.

## A few terms first

**Fair use** is a defense in US copyright law (Section 107 of the Copyright Act). It allows some copying without permission. Courts weigh four factors: what the copying is for, the kind of work, how much was copied, and the effect on the market for the original. The fourth factor, market harm, is often treated as the most important.

A use is **transformative** when it serves a new purpose rather than standing in for the original. **Summary judgment** is a ruling made without a trial, when the judge finds the key facts are not genuinely in dispute. A **shadow library** is an online collection of pirated books, such as LibGen.

## Bartz v. Anthropic: training is fair use, pirated library copies are not

Three authors, Kirk Wallace Johnson, Andrea Bartz and Charles Graeber, sued Anthropic, the company behind Claude. According to the order, Anthropic gathered books in two ways: it downloaded millions of pirated books from Books3, LibGen and another pirate collection, and it bought print books, removed the bindings and scanned the pages. It kept all of this as a central library and drew subsets from it to train its models.

Judge Alsup split the question into three uses and reached three answers:

| Use | Judge Alsup's ruling |
|---|---|
| Copies used to train the language models | Fair use; he called the purpose highly transformative |
| Scans of print books Anthropic had bought | Fair use, as a format change of copies it owned |
| Pirated books kept in the permanent library | Not excused; this goes to trial, including damages |

Two details matter. First, the authors did not claim that Claude ever showed users infringing text from their books, and the order notes that the public service filters what goes in and out. The judge wrote that a case built on infringing outputs would be a different case. Second, he compared training to a person reading many books in order to learn to write, and he treated the competing writing that might result as outside what copyright protects against. He also said that buying a book after pirating it does not undo the earlier copy, though it could affect the damages.

## Kadrey v. Meta: Meta wins, but on the record, not the principle

Thirteen authors, mostly well-known novelists, sued Meta for downloading their books from shadow libraries to train its Llama models. Judge Chhabria granted Meta summary judgment on fair use for that training. His reasons, however, were narrow.

The authors made two main arguments about market harm. They said Llama can reproduce bits of their books, and they said Meta took away a market for licensing books as training data. The judge rejected both: the model could not output enough of their text to matter, and authors are not entitled to that licensing market. The argument he thought could have won, which he called **market dilution**, was that AI output can flood a market with similar books and hurt sales of the originals. The plaintiffs offered almost no evidence on it, so the fourth factor favored Meta.

He then went out of his way to limit the decision. He wrote that his ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful," only that these plaintiffs argued the wrong points. He also said that in many circumstances training on protected works without permission will likely be illegal, and he openly disagreed with Judge Alsup's schoolchildren comparison. A separate claim, that Meta shared the books with others while downloading them by torrent, is still open.

> **Try it:** Read the conclusion section of each order (the last pages) and write one sentence for each in your own words: what the company won, and what is still unresolved. Compare your two sentences with the table above.

## Where the two judges agree and disagree

They agree that training a model is a different purpose from reading or selling the book, so the first factor leans toward the companies. They also agree that the outcome turns on evidence in each case.

They disagree on how much the market factor should count. Judge Alsup gave weight to how transformative training is and did not see AI-driven competition as a copyright harm. Judge Chhabria said harm to the market outweighs the purpose of the copying, and that a product able to generate endless competing works could be exactly that harm. Appeals courts may eventually have to choose between these views.

## What do they mean for authors?

For now, no court has said that AI companies may freely train on any book they find. Anthropic still faces a trial over its pirated library, and how a company got the books has become a central issue. Judge Alsup drew a line between books that were bought and scanned and books that were downloaded from pirate sites.

The Meta ruling also shows what may decide future cases: evidence about real markets. Judge Chhabria pointed to market dilution as the strongest argument and said these plaintiffs did not prove it. Authors and publishers considering their options will likely watch for cases that bring that kind of evidence, along with any appeals of these two orders.

If your own books may be in a training set, the practical steps are modest: learn which datasets were named in these cases, follow updates from authors' organizations, and speak with a copyright lawyer before making decisions about your rights. Both orders are public, and reading their introductions and conclusions is the most reliable way to see what was and was not decided.

## Sources

- Bartz v. Anthropic PBC, No. C 24-05417 WHA: Order on Fair Use: https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_2.pdf
- Kadrey v. Meta Platforms, No. 23-cv-03417-VC: Order Denying the Plaintiffs' Motion for Partial Summary Judgment and Granting Meta's Cross-Motion: https://storage.courtlistener.com/recap/gov.uscourts.cand.415175/gov.uscourts.cand.415175.598.0.pdf