The market-dilution theory in the Google Gemini copyright case

The allegation worth circling first in the Google Gemini copyright lawsuit filed by major publishers is almost absurdly small: $0.39. Hachette, HarperCollins, Macmillan, Penguin Random House, Simon & Schuster, and the named author plaintiffs allege that Gemini can generate a 100-page murder mystery in about 20 minutes for $0.39, a fact the complaint uses to make its broader theory legible without asking the court to accept every premise about generative AI at once.[1]

That figure is not just color. It is pleading architecture. A copyright complaint that says an AI model was trained on books still has to do more work: identify the protected works, explain the allegedly infringing acts, anticipate fair-use arguments, and, on market harm, show something more concrete than unease about technological scale. The $0.39 murder mystery allegation gives the plaintiffs a proposed bridge from model capability to market injury. If a reader can obtain a genre-length substitute quickly and cheaply, the complaint says, the harm is not limited to copying during training. It extends to outputs that compete with books themselves.[1]

The complaint calls Gemini, in substance, an engine of substitution. That is the part of the filing that matters most. The plaintiffs are not content to say Google copied books to build a model. They plead a market-dilution framework in which Gemini allegedly competes with copyrighted works through three related channels: reproducing protected expression, generating chapter-length substitutes for instructional and educational materials, and flooding markets with synthetic works that depress the value of books even when no particular title is reproduced verbatim.[1]

The complaint turns market harm into a sequence of alleged substitutions

Market harm is often pleaded at a level of abstraction that is easy to nod past and hard to prove. This complaint tries a different route. It does not rely on one kind of injury. It layers injuries, moving from the most familiar copyright harm to the most difficult economic theory.

The first channel is the cleanest legally. If Gemini generates verbatim or near-verbatim passages from protected books, the plaintiffs can frame the output as copying that reaches beyond hidden model training and into the marketplace. That does not automatically win the case. The outputs still have to be examined, compared, attributed, and tested against defenses. But it gives the court something concrete to look at: words on one side, words on the other, and a question of protectable expression.

The second channel is more commercially pointed. The complaint alleges that Gemini can generate textbook and instructional material from generic prompts, including chapter-length outputs that could substitute for purchased educational content.[1] This matters because instructional works occupy markets where readers often buy for function as well as expression. A student, tutor, teacher, or self-directed learner may not need the same authorial voice if the immediate task is to understand a topic, draft a lesson, or obtain a worked explanation. The legal question, however, remains narrower than the business anxiety: it is not whether AI can generate useful study material, but whether the allegedly infringing system harms cognizable markets for the plaintiffs’ protected works.

The third channel is where the complaint becomes most ambitious. Aggregate market flooding is not about a single reader choosing a Gemini output instead of a single named book. It is about volume. The theory is that a system capable of producing unlimited low-cost genre works, guides, summaries, and instructional materials can depress the value of entire categories of human-authored works. The $0.39 murder mystery does its real work here. It makes the market-flooding allegation visible as a price-and-scale story, not just a complaint about technological unfairness.[1]

The difficulty is that these channels do not all require the same proof. Near-verbatim reproduction can be litigated output by output. Textbook substitution requires evidence about the role the output plays for users. Market flooding requires a market definition, an economic mechanism, and proof that any observed loss is caused by Gemini rather than by ordinary competition, changing reader behavior, retail pricing, subscription models, piracy, library access, social media attention, or other generative AI systems.

Why Kadrey matters to this pleading

The most important comparison is not the total number of AI copyright suits, though the docket has become crowded. It is Kadrey v. Meta, because Judge Vince Chhabria’s treatment of market harm clarified the missing piece in a way that any careful plaintiff’s lawyer would notice. As summarized in 2026 litigation updates, Kadrey treated the training-use question as capable of falling within fair use on the record before the court, while faulting the plaintiffs for not producing market-dilution evidence strong enough to carry their theory.[2]

That distinction matters. A court can be skeptical of a record and still leave a doctrinal door open. Kadrey did not prove that AI book-training cases can never show market harm. It showed that saying a model was trained on copyrighted books is not enough. The Hachette complaint reads like a filing drafted into that gap: identify substitutive outputs, allege concrete price and speed, plead market flooding, and tie the theory to licensing markets and internal knowledge.

There is a useful discipline in that approach. It resists two lazy moves. The first is the publisher-side shortcut that treats ingestion of books as the end of fair-use analysis. The second is the technology-side shortcut that treats every AI training dispute as if it were Authors Guild v. Google in a new wrapper. Neither shortcut answers the output problem. A model that returns non-expressive search snippets is not the same legal object as a model alleged to generate expressive, reader-facing substitutes for books.

Kadrey therefore sharpens the issue rather than deciding it. The Hachette plaintiffs still need evidence. But they have identified the evidentiary category they need with more precision than earlier pleadings: not merely that Google benefited from books, and not merely that publishers dislike uncompensated training, but that Gemini allegedly supplies outputs that displace demand or dilute value in markets copyright law recognizes.

Internal documents help only if they connect knowledge to market injury

The complaint also leans on alleged internal Google materials. The plaintiffs say Google documents recognized that copyrighted books improved Gemini’s performance compared with public-domain-only training, and that Google internally flagged use of publisher-provided Google Play Books for AI training as a heightened risk that could expose the company to fines in the tens or hundreds of billions of dollars.[1][3]

At the pleading stage, those allegations are useful for knowledge, value, and willfulness. They help the plaintiffs argue that books were not incidental training debris, but high-value inputs. They may also undermine any suggestion that Google reasonably viewed all available book text as legally interchangeable with public-domain material. But a quoted internal document is not yet an authenticated trial exhibit. The complaint was filed on July 10, 2026 in the Southern District of New York, Google has not responded, and the excerpts will have to survive the ordinary tests of context, admissibility, and interpretation.[1]

The same caution applies to the licensing allegations. The complaint points to Google’s willingness to license some publisher content in other settings, including content arrangements connected to products such as Google News Showcase and academic or scholarly uses, while allegedly declining to license the plaintiffs’ books for Gemini training.[1][4] That is relevant because fair-use market harm includes not only lost sales of existing books but also harm to traditional, reasonable, or likely-to-develop licensing markets. Still, a licensing market cannot be invoked as a magic word. The plaintiffs will need to show that the market is real enough, defined enough, and close enough to the use at issue.

There is a difference between evidence that Google sometimes pays publishers and proof that Google was legally required to pay these publishers for this use. The former helps the complaint get traction. The latter is what litigation is for.

The hard proof problem: capability is not substitution

The most vulnerable move in the market-dilution theory is the leap from capability to conduct. A system that can generate a cheap murder mystery does not, by itself, prove that readers use those outputs instead of buying mysteries. A system that can generate a chapter-length explanation does not, by itself, prove that schools, students, or libraries buy fewer textbooks because of it. A flood of AI-generated works does not, by itself, identify which copyrighted markets have been diluted or by how much.

That is where discovery becomes decisive. The plaintiffs will want usage data, prompt data, output samples, product metrics, internal market analyses, licensing negotiations, retention studies, and documents showing how Google understood Gemini’s competitive relationship to books. They may also need survey evidence or behavioral data showing that users treat Gemini outputs as substitutes rather than supplements. An expert can model market dilution, but the model will need inputs that are not simply assumed from technological possibility.

Causation will be especially unforgiving. Publishing markets are not clean laboratories. A decline in sales for a category of books, if one exists, may have many explanations. If the plaintiffs focus on genre fiction, Google may point to other AI tools, self-publishing trends, subscription reading programs, discounting, piracy, or shifts in reader attention. If the plaintiffs focus on educational materials, Google may point to open educational resources, institutional procurement changes, tutoring platforms, or ordinary digital transformation. The complaint’s theory is elegant; the evidentiary record will be messy.

Market definition will matter too. The plaintiffs cannot simply say “books” and be done. A literary novel, a commercial mystery, a test-prep guide, a trade nonfiction book, and a textbook do not necessarily face the same substitution dynamics. Some readers buy for the author. Some buy for information. Some buy because a course requires a title. Some borrow, browse, or sample. The more precisely the plaintiffs define the affected markets, the more plausible the economics may become; the more broadly they define them, the easier it will be for Google to attack causation and measurement.

The parallel cases sharpen the stakes without deciding this one

The Meta case filed by publishers and authors in May 2026 in the Southern District of New York matters because it uses a closely related market-dilution theory against another major AI system.[5] If the cases proceed, plaintiffs may seek overlapping discovery about book ingestion, output behavior, licensing markets, and economic harm. That does not make the theory stronger on the merits, but it may make the evidentiary project more organized.

The Anthropic settlement also sits in the background, though it should not be overread. In September 2025, Anthropic agreed to a $1.5 billion settlement in litigation involving roughly 500,000 books, a figure widely described as about $3,000 per work.[6] That settlement creates settlement pressure and valuation vocabulary. It does not establish that Google infringed, that Gemini outputs substitute for books, or that the same valuation method applies here.

Bartz v. Anthropic, Thomson Reuters v. Ross, and Authors Guild v. Google are useful mainly as boundary markers. Bartz is relevant for the distinction between training and allegedly unlawful acquisition; Thomson Reuters is relevant because a competitive AI use can fail fair use on a different record; Authors Guild remains important because the Second Circuit approved Google’s book-search use where the product was transformative, non-expressive, and did not provide a market substitute for books.[2][7] None of those cases answers the Gemini output question without a record.

Even the size of the AI copyright docket should be handled cautiously. Public counts vary by methodology and date: one tracker listed 87 U.S. copyright suits against AI companies as of March 2026, while publishing-industry coverage has described a larger universe by mid-2026.[8][5] The important point is not the exact number. It is that courts are being asked to sort a set of disputes that look similar at the headline level but turn on different records: what was copied, how it was obtained, what the model does, what users receive, and what market evidence exists.

What the complaint has built, and what it has not

As a pleading, Hachette v. Google is sophisticated because it understands the weakness it has to cure. It does not merely allege unauthorized training. It tries to show how training, output, user substitution, licensing displacement, and category-wide dilution fit together. The theory is built to survive the response that AI training is just another information-processing use. It says Gemini is not only processing books; it is allegedly producing works that can stand in the marketplace where books used to stand.

But sophistication is not proof. The $0.39 murder mystery allegation is a powerful way to plead substitutive capacity. It is not yet evidence that readers are abandoning books for Gemini-generated mysteries. The alleged internal documents may help show knowledge and value. They do not yet prove market injury. The licensing allegations may help identify a market Google chose not to enter. They do not yet establish the market’s legal scope or the amount of harm.

That is the hinge. The complaint has built the most refined market-harm theory yet in the AI book cases, precisely because it treats market dilution as something to be pleaded through mechanisms rather than recited as a conclusion. Whether it becomes more than that will depend on discovery: whether the plaintiffs can convert the “infinite substitution machine” allegation into evidence of actual substitution, measurable dilution, and causation strong enough to survive the next procedural stage.

References

1. Class Action Complaint, Hachette Book Group, Inc. et al. v. Google LLC, Association of American Publishers, July 10, 2026.
2. AI in litigation series: An update on AI copyright cases in 2026, Norton Rose Fulbright.
3. Publishers sue Google over Gemini AI training, The Guardian, July 14, 2026.
4. Google faces another AI training lawsuit from major publishers, TechCrunch, July 14, 2026.
5. Publishers and Authors Sue Meta, Alleging Massive Copyright Infringement Behind Its Llama AI Service, Publishing Perspectives, May 2026.
6. Anthropic agrees to pay $1.5bn to settle AI copyright lawsuit with authors, The Guardian, September 5, 2025.
7. Authors Guild v. Google, Inc., Justia, October 16, 2015.
8. Latest U.S. Map of Copyright Suits v. AI Companies: Total 87 (Mar. 5, 2026), ChatGPT Is Eating the World, March 5, 2026.