The complaint against Gemini is being covered as another entry in a losing streak. Read the ruling everyone is citing as a defeat for authors, and it turns out to be a set of instructions. The new filing follows them line by line.

Hachette Book Group, Cengage Learning, Elsevier, and the novelist Scott Turow filed a class action against Google on July 10 in the Southern District of New York, alleging the company copied millions of books, textbooks, journal articles, and other written works to train its Gemini models without permission. Turow's publishing entity S.C.R.I.B.E. is also named. The plaintiffs want damages and an injunction.

The complaint accuses Google of "one of the most prolific infringements of copyrighted materials in history," and takes a swipe at the company's retired motto on the way.

Most of the coverage has filed this under a familiar heading: publishers keep suing, courts keep siding with AI companies, here goes another one. Two California decisions in June 2025 found that training on copyrighted books was fair use. The implication is that the publishers are pushing a rock uphill.

That reading gets one of those two decisions almost exactly backwards.

The ruling that was not a win

On June 25, 2025, Judge Vince Chhabria of the Northern District of California granted summary judgment to Meta in Kadrey v. Meta. Headlines recorded a win for AI.

Read the opinion and it is closer to a threat.

Chhabria found Meta's training highly transformative, then spent the bulk of his analysis on the fourth fair use factor, market harm, where he laid out a theory he called market dilution: an LLM trained on books can flood the market with competing works, damaging authors even when every individual output is lawful and original. No other use, he wrote, has "anything near the potential to flood the market with competing works." He called Judge Alsup's contrary analogy, which compared LLM training to teaching schoolchildren to write, "inapt," and accused his colleague of "blowing off the most important factor in the fair use analysis."

Then he ruled for Meta anyway, because the authors in front of him had not made that argument.

He said so in terms that read like a memo to future litigants. His decision, he wrote, did not establish that Meta's training was lawful. It established only that "these plaintiffs made the wrong arguments and failed to develop a record" in support of the right one. He added that in cases like this one, plaintiffs will often win where the record is better developed.

A federal judge told the plaintiffs' bar what evidence to bring and predicted they would prevail with it. Twelve months later, a nearly identical roster of publishers filed against Google.

The complaint reads like the roadmap is open on the desk

The Hachette filing is not arguing that Gemini regurgitates books. That was the losing theory.

It argues substitution. The complaint alleges Gemini produces near-verbatim copies of portions of works, replacement chapters for academic textbooks, alternative versions of famous novels, and knockoffs that copy the creative elements of originals. It alleges the model tailors output to mimic the choices of specific authors. It alleges a chatbot can generate a 100-page murder mystery in roughly 20 minutes "for a mere $0.39."

That figure is doing precise work. Chhabria's complaint about Alsup's schoolchildren analogy was that it ignored speed and cost, the ability of one person to mass-produce commercial text. The publishers put speed and cost in the pleading, with numbers attached.

They also brought the plaintiff coalition Chhabria's theory needs. Hachette covers trade fiction. Cengage covers educational publishing. Elsevier publishes The Lancet and Cell. That spread lets the class argue dilution across the whole book market rather than in one genre, which is the record Chhabria said was missing.

The same group sued Meta in the same Manhattan courthouse in May, joined there by Macmillan and McGraw Hill. They have run this play before, in this courthouse, with these lawyers.

Google's own precedent is the trap

Here is the part that makes this case different from every other AI training suit, and it has gone almost entirely unmentioned.

Google already won the landmark book-scanning fair use case. In Authors Guild v. Google, decided in 2015, the Second Circuit held that Google's digitizing of tens of millions of books and display of search snippets was fair use. The panel was unanimous. The opinion was written by Judge Pierre Leval.

Leval is the reason the phrase "transformative use" exists in American copyright law. His 1990 Harvard Law Review article introduced the standard the Supreme Court adopted four years later, and it is the doctrine Alsup leaned on when he called Anthropic's training "exceedingly transformative". The framework that gives AI companies their best defense came from the same judge who wrote Google's book-scanning win.

Now look at how that win was scoped. The Second Circuit held the copying was fair because the purpose was search and because the snippet display was limited enough that the revelations "do not provide a significant market substitute for the protected aspects" of the originals. Non-substitution was the condition of the victory. It was the load-bearing wall.

Authors Guild v. Google is binding precedent in the Second Circuit. The Southern District of New York sits in the Second Circuit. The publishers filed there.

So Google walks into a courtroom bound by a ruling it won, which says book copying is fair use when it does not substitute for the original, to face a claim that it took books from Google Books and Google Play and built something the plaintiffs say substitutes for the original. The precedent that protects Google contains the exact limiting principle the publishers intend to hang it with. Alsup and Chhabria sit in the Ninth Circuit and bind nobody in Manhattan.

Venue was not an accident.

What Google has going for it

The publishers have real problems, and the coverage skips these too.

Both California judges rejected the argument that a licensing market existed and that AI companies wrongfully bypassed it. Neither found it cognizable. That theory is 0 for 2, and it is central to the publishers' damages story here.

Worse, Chhabria's own opinion contains a detail that cuts against them badly. Meta's head of generative AI had discussed spending up to $100 million on licensing. The deals did not happen, and the reason was not corporate greed. Publishers largely do not hold AI training rights for books. Individual authors do, there is no collective licensing body for those rights, and where publishers hold them at all they tend to hold them regionally. A company willing to write a nine-figure check could not find anyone able to sell it what it wanted to buy.

The publishers say Google could have licensed, pointing to HarperCollins striking a deal with Microsoft in 2024. Google will point to Chhabria's paragraph and ask which rights, from whom.

The claim that Google stripped copyright management information to hide its training data has a problem too. That is a DMCA theory, and Chhabria granted Meta summary judgment on exactly those claims in Kadrey. It has already failed once against nearly the same lawyers.

And there is the correction worth making, because the trade press keeps repeating it: Anthropic was not fined $1.5 billion. It settled for that figure, and Judge Alsup postponed preliminary approval in September 2025 over concerns the agreement was incomplete and that class counsel were binding authors to a deal without adequate consideration. A settlement is not a judgment. It creates no precedent and decides no question of law. Every outlet describing it as a fine is telling readers a court ruled something that no court ruled.

The document problem

What settled the Anthropic case was not a fair use holding. It was discovery.

The Hachette complaint suggests Google has the same exposure. The plaintiffs cite internal Google material allegedly warning that training on copyrighted books would be highly problematic for the company and could produce "$10Bs-$100Bs in potential fines." Other internal communications quoted in the filing allegedly note that book publishers were likely to view LLM training on their books as infringement, and flag heightened risk around fair use defenses.

Fair use is a defense a company raises. Willfulness is a state of mind a plaintiff proves. Statutory damages for willful infringement run to $150,000 per work, and documents in which a company's own staff price the exposure at tens of billions are the kind of evidence that converts an argument about doctrine into an argument about numbers.

Google did not immediately respond to a request for comment.

What to watch

The case may be swept into In re Google Generative AI Copyright Litigation, the consolidated proceeding the company is already defending. That would slow it and dilute the specificity the publishers built the complaint around.

The live question is whether Judge Chhabria's dilution theory travels. It is dicta from a district judge in another circuit, binding on no one, and it may not survive contact with the Supreme Court's Warhol decision, which suggests substantial market harm undercuts a finding of transformative use in the first place rather than sitting downstream of it. The Copyright Office endorsed something close to Chhabria's view in its 2025 report, and its director was dismissed days later.

What the publishers have done is take a loss and read it as a syllabus. The only question left is whether a judge in Manhattan agrees the assignment was completed.