What Was Filed
On July 15, 2026, Hachette Book Group, Cengage Learning, Elsevier, and novelist Scott Turow filed a putative class action against Google in the U.S. District Court for the Southern District of New York, alleging that Google trained its Gemini models on millions of copyrighted books and journal articles without authorization. The nearly 60-page complaint alleges Google drew training text from three sources: material scraped from behind paywalls, works obtained from pirate sites, and — most pointedly — books and journal articles that publishers had uploaded to Google Books and Google Play under license terms that limited Google to snippet-view search, not AI training. The suit also alleges Google stripped or altered copyright management information to obscure the origin of the material, and cites an internal Google document warning that using copyrighted books this way could be "highly problematic for Google," with exposure of "$10Bs–$100Bs in potential fines." It is the same plaintiff group — Elsevier, Cengage, Turow, and Hachette — that sued Meta earlier this year over Llama's training data, suggesting a coordinated publisher-side litigation strategy against every major model developer at once.
Provenance Is the Whole Case
The complaint is carefully built around how Google obtained the text, not merely that it trained on copyrighted work. That distinction now has real legal teeth. In Bartz v. Anthropic, Judge William Alsup ruled that training an AI model on lawfully acquired books can be fair use, but that acquiring the books through piracy — in that case, downloading them from shadow libraries like LibGen — is not protected regardless of what happens downstream. Anthropic settled the resulting class action for $1.5 billion in August 2025, one of the largest copyright settlements in U.S. history, covering roughly 500,000 pirated works. The Hachette-Elsevier complaint against Google is transparently modeled on that theory: it does not ask a court to rule that AI training itself infringes copyright, a question still unsettled nationally. It asks a court to rule that Google's specific acquisition method — alleged piracy, alleged paywall circumvention, and alleged scope creep beyond a licensed search product — was unlawful, and that the AI training use built on top of it inherits that taint.
Steelmanning the Publishers
The publishers' case deserves to be taken on its strongest terms. If Google truly ingested paywalled journal content and pirated e-books to train a commercial product used by billions, that is not a novel legal question dressed up in AI language — it is garden-variety infringement, and the fact that a large language model sits downstream doesn't launder the acquisition. The Google Books allegation is sharper still: publishers negotiated, and courts blessed, a specific bargain in Authors Guild v. Google (2015) — snippet-view search in exchange for indexing, not full-text ingestion for a generative product that can reproduce or closely paraphrase copyrighted expression. If Google used that scope-limited corpus for a fundamentally different purpose without renegotiating consent, publishers have a legitimate grievance about a broken promise, not just a copyright technicality. The U.S. Copyright Office's own May 2025 Part 3 report on AI training reached a similar line: training that is genuinely transformative and general-purpose leans toward fair use, but commercial use of vast troves of copyrighted work to produce expressive content that competes with the originals — especially when accessed illegally — falls outside established fair use boundaries.
Why the Broader Backlash Still Misreads the Risk
Where this reasoning should not be stretched is into a general presumption against AI training on copyrighted material. The same Copyright Office report and the Alsup ruling it echoes both preserve a workable, and important, distinction: training on lawfully licensed or purchased text remains fair use because it is transformative — the model doesn't reproduce the work, it learns statistical patterns from it, much as a search index or a plagiarism-detection tool does. Treating every training run as presumptively infringing, rather than scrutinizing how the underlying corpus was obtained, would punish AI companies that did license or purchase their data lawfully exactly as harshly as those that scraped shadow libraries — collapsing an incentive structure that should reward clean acquisition. It would also disproportionately burden smaller AI labs that can't absorb billion-dollar litigation risk, entrenching the handful of incumbents, Google included, who can. Congress's own response so far reflects this more careful instinct: the Generative AI Copyright Disclosure Act, still pending, would require AI developers to file a public notice of copyrighted works used in training datasets — a transparency fix that makes provenance auditable, rather than a ban on AI training itself.
The Editorial Takeaway
The Hachette-Elsevier suit is likely to matter more for what it forces into discovery than for any injunction it might eventually win: internal Google communications about training data sourcing, and whatever the company actually did with the Google Books corpus after 2015. If the piracy and paywall-scraping allegations hold up, Google should pay for it under the same acquisition-based theory that cost Anthropic $1.5 billion — proportionate liability tied to how the data was obtained, not a referendum on generative AI as a technology. Regulators and courts should keep drawing that line precisely, because blurring it in either direction — letting provenance-laundering models off the hook, or treating all AI training as inherently unlawful — would make U.S. copyright law worse at the one thing it's supposed to do: reward the humans who created the underlying work while still allowing genuinely transformative technology to develop.