Three publishers—Hachette Book Group, Cengage Learning, and Elsevier—along with author Scott Turow and others, filed suit against Google in the U.S. District Court for the Southern District of New York on July 10, 2026, alleging copyright infringement in the development of Gemini. The complaint does not rest on an abstract, all-encompassing argument about whether AI training itself is permissible. Instead, it separates works accumulated in Google Books and elsewhere for different purposes, documents obtained from the Web, and the copies generated during the training process itself—arguing that reproduction rights were infringed at each of these distinct stages. At stake is whether Google can repurpose the book data it spent years building for search as a competitive asset for generative AI.
The complaint also cites an internal Google document warning that using publisher-provided books for AI could expose the company to "$10 billion to $100 billion in potential penalties." However, this figure is neither the amount the plaintiffs are seeking nor damages determined by any court. It is an internal risk assessment that the plaintiffs cite to argue that "Google was aware of the legal risk."
The Gemini Lawsuit Broken Down Into Four Claims
The case number is 1:26-cv-05870. S.C.R.I.B.E., which holds the copyright to Scott Turow's works, has also joined as a plaintiff, and the suit seeks class certification to represent authors and publishers who suffered similar harm. The works at issue include general books and textbooks, as well as academic papers handled by Elsevier.
The 57-page complaint asserts four claims. The first concerns unauthorized reproduction of works held in Google Books, Google Play Books, Google Scholar, and similar services. The second concerns reproduction of works obtained through Web scraping, and the third concerns copies generated during the training process. The fourth alleges violation of the Digital Millennium Copyright Act, claiming that copyright notices, author names, and publication information—copyright management information—were removed or altered during preprocessing.
This breakdown matters. Even if AI model training is recognized as a transformative use, the question of where the underlying data came from and under what authorization remains separate. Further, the plaintiffs count as distinct acts of reproduction each step in which acquired works are moved to high-speed storage, loaded into memory, tokenized, and reprocessed every time a new model is created.
The relief the plaintiffs seek is also broad. In addition to an injunction against the alleged infringing conduct and damages, they ask the court to require Google to identify, on a work-by-work basis, the books and papers used for Gemini and to disclose how they were collected, processed, and encoded. For data found to constitute infringing copies, the plaintiffs also seek destruction under court supervision. The complaint does not specify a fixed damages amount.
Can Google's Google Books Victory Extend to Gemini?
Google has a major legal victory concerning the full-text digitization of books. In 2015, the Second Circuit Court of Appeals held that Google Books's practice of copying entire books to make them searchable constituted fair use. Search results display only short snippets, which do not meaningfully substitute for purchasing the original work. The court emphasized that the purpose of making information about books easier to find is different from the purpose of reading the original work, and it weighed this combination of factors heavily.
The plaintiffs here turn that winning rationale on its head. They argue that what was upheld in the Google Books case was copying for the purpose of search and snippet display, and that training a commercial model to generate expressive content serves a different purpose. The complaint cites instances in which Gemini reproduced textbook chapter structures and text closely resembling the original, and generated novel summaries and similar works. The plaintiffs emphasize the distinction between Google Books, which drives traffic toward search, and Gemini, which can potentially generate substitutes for works in the market.
Meanwhile, Google took the position, in a policy statement published in April 2025, that training foundation models using content from the open Web constitutes transformative fair use under U.S. law. The U.S. Copyright Office has also acknowledged that building general-purpose models from massive, diverse datasets can be transformative. Training generative AI is not automatically unlawful.
Still, the Copyright Office rejected a one-size-fits-all conclusion. It stated that one must first look at the work, the source from which it was obtained, and the purpose of use, and then evaluate output controls and market impact. In this lawsuit, the key question will be whether Gemini's outputs and guardrails satisfy the "not a market substitute" condition that was decisive in the Google Books ruling.
Before Training Comes the Question of How the Data Was Obtained
The strongest part of the complaint concerns the purpose for which Google originally obtained the data it already held. According to the plaintiffs, e-books provided to Google Play Books were intended for sale, and papers provided to Google Scholar were intended for search and directing users to legitimate access—neither included authorization for AI training. The complaint also alleges that Gemini's lead engineer stated that Google decided it "wouldn't sign contracts for data it already had."
Another acquisition route was the Web. According to the complaint, LaMDA's training dataset, Infiniset, consisted of 2.97 billion documents, of which approximately 371 million documents—12.5%—came from Google's C4 dataset. C4 is a curated selection of documents from Common Crawl. The plaintiffs allege that copyright symbols appeared more than 200 million times within C4, and that it included documents from at least 27 sites identified by the U.S. government as piracy markets, including a former domain of Z-Library. These are allegations that will need to be tested through evidence and Google's response going forward.
Other litigation has offered a framework for distinguishing between acquisition routes. In the 2025 Bartz v. Anthropic case, the process of using books to train a model was found to be fair use. On the other hand, at least on the summary judgment record, obtaining millions of books from pirate sites and storing them in a central library was not found to be fair use. However, this ruling came from a separate case in the Northern District of California and is not binding on the New York court hearing the present case.
A policy report from the U.S. Copyright Office similarly concluded that outcomes vary depending on the source and the use to which the data is put. Nonprofit research or analysis that cannot output the expressive content of a work is more likely to qualify as fair use. By contrast, uses that obtain expressive works from pirate sources—despite licenses being realistically available—and then generate content that competes in the market without restriction are less likely to be recognized as fair use. While this document is not binding on courts, the current complaint is structured to connect readily to this kind of case-by-case evaluation by separating three points of entry: Google Books, Web acquisition, and training.
Demanding a Work-by-Work Ledger Using ISBNs and DOIs
The class definition the plaintiffs propose is specific. It covers rights holders of works with an ISBN, or of academic papers with a DOI or ISSN, that satisfy U.S. Copyright Office registration requirements and were reproduced without authorization through any of Google's services, Web scraping, or the development and training of Gemini. Class certification requires showing that common questions of fact and law apply to numerous rights holders, and this proposed definition has not itself been accepted by the court.
Here, a practical obstacle emerges: work-level data provenance. Google's current Gemini 3 Pro model card discloses categories such as public Web and crawler-obtained data, and commercially licensed data. It also covers user data, data acquired or generated internally, and synthetic data. The card explains that robots.txt is respected during preprocessing. However, Google has not published a list showing which specific books or papers were used, on what basis, for which model.
For Web site operators, there is Google-Extended. This control token lets site owners specify whether their site's content may be used for training next-generation Gemini models or for search grounding, without affecting inclusion or ranking in ordinary Google Search. But it does not resolve the treatment of files previously entrusted to Google Books or Play Books in the past, works reposted on pirate sites, or cases where the rights holder and the site operator differ.
If the case proceeds to discovery, Google will likely be pressed to provide data provenance more granular than its category-level disclosures. The court's first task will be to determine the scope of what can proceed as a class action; beyond that lies the larger question of what authority is needed to move works obtained for search purposes into generative AI, and how much work-level record-keeping should be required. Google's answer to the complaint and the court's ruling on class certification will determine whether Gemini's training data ledger can be brought into the open in court.