Publishers sue Google over Gemini training

Three major publishers and a bestselling author have sued Google in federal court in New York, accusing the company of copying millions of copyrighted books to train its Gemini artificial-intelligence models. According to The Guardian, the plaintiffs include Hachette Book Group, Cengage Learning and Elsevier, alongside author Scott Turow. They argue that books provided to Google under narrow, older arrangements—such as Google Books snippet search and ebook sales through Google Play—were repurposed into training data for commercial AI systems without permission or payment.

The filing lands in a legal environment where the basic inputs of generative AI—large, messy corpora assembled at industrial scale—collide with a rights system built for discrete acts of copying and distribution. Publishers have spent years supplying Google with content under programs designed to make books searchable or sellable, but those programs were negotiated around limited uses: bibliographic display, searchable excerpts, and storefront distribution. The lawsuit’s core claim is that the same supply chain became a back door into full-text copying for model training, converting a licensing relationship into a data pipeline.

TechCrunch reports that the case is framed as a class action and alleges Google removed or altered copyright information to conceal that Gemini models were trained on “stolen materials.” The complaint also points to internal Google discussions acknowledging legal exposure, including potential fines in the tens of billions of dollars, suggesting the company treated copyright risk as a balance-sheet variable rather than a bright-line constraint. The plaintiffs are seeking statutory damages, a permanent injunction, and an order requiring destruction of allegedly unauthorized copies used in training—remedies that, if granted, would target not only future behavior but the accumulated assets inside a model’s training set.

The economics behind the dispute are straightforward: training on books is valuable because books are dense, edited, and structured—exactly the kind of text that makes models more coherent. Yet the commercial upside from that coherence largely accrues to the AI provider, not to the authors, editors, and publishers whose work supplies the signal. The lawsuit highlights substitution risk in blunt terms, arguing that an AI system can generate long-form genre fiction quickly and cheaply, turning a catalog of paid creative labor into a template for near-zero-marginal-cost imitators. Whether courts treat that as infringement, fair use, or something in between will determine whether future training data is acquired through negotiated licenses or through litigation after the fact.

Google did not respond to comment requests cited by The Guardian and TechCrunch. The case now sits in the Southern District of New York, where a judge will be asked to decide whether a book-scanning era deal about searchable snippets also implicitly authorised the mass copying needed to build Gemini.