Publishers Sue Google Over Alleged Gemini Training on Copyrighted Books

Google is facing a new class action lawsuit from major publishers and authors who say the company trained its Gemini AI system on copyrighted books without permission. The complaint, filed Tuesday in federal court in New York, argues that Google not only used protected texts to build its models but also altered or removed copyright information to hide what it had done, making the case one of the latest major tests of AI training and copyright law.

The lawsuit brings together well-known plaintiffs including Hachette, Cengage, Elsevier, novelist Scott Turow, and S.C.R.I.B.E., and adds to a growing wave of litigation from rights holders who say generative AI companies have built products on top of creative work they did not license. The case arrives at a moment when courts are still sorting out how long-standing copyright rules apply to machine learning systems that ingest vast libraries of text, images and audio to generate new content.

For Google, the dispute is especially sensitive because it involves a company with a long history of working with publishers through Google Books, a product that was built to index and preview books rather than reproduce them in full. The plaintiffs say that historical relationship does not give Google permission to use those same titles for AI training, especially when the company allegedly never sought consent for Gemini.

What the publishers say Google did

At the center of the complaint is an allegation that Google copied books from limited-access programs and used them to train Gemini models without authorization. The plaintiffs argue that this was not a stray or accidental use, but an intentional decision to feed copyrighted material into the company’s AI pipeline.

According to the lawsuit, Google relied on books that had been made available through scope-limited publishing arrangements, along with titles uploaded to the Google Play Store. The key claim is that those books were used for model development even though the company had not obtained a license covering AI training.

The complaint says Google “illegally copied” works from programs designed for narrow uses and did so despite knowing it did not have permission to train Gemini on them.

The filing also alleges that Google intentionally modified or removed copyright metadata on some works, which the plaintiffs say was done to obscure the source of the training data and make it harder to trace where Gemini’s underlying material came from.

Why the Google Books history matters

The lawsuit’s narrative depends in part on Google’s long-standing relationship with publishers. Google Books was created to make books searchable, not to turn them into training data for generative AI. That distinction could become important as courts assess whether a company that had permission for one use can later repurpose the same material for something fundamentally different.

Under Google Books, readers can typically see short snippets of text and bibliographic details, but not full books. The plaintiffs argue that this limited access model does not translate into an open-ended right to ingest the same books into an AI training system that can produce new text on demand.

How does this lawsuit fit into the broader AI copyright fight?

This lawsuit is part of a wider legal battle over whether AI companies can train models on copyrighted works under U.S. law. Publishers, authors, artists and other rights holders have filed a growing number of cases against companies including Google, Meta, OpenAI and Anthropic, claiming those firms built valuable AI products by copying creative works without a proper license.

At the same time, AI companies have argued that training on copyrighted material can qualify as fair use, particularly when the original works are used to teach systems to generate new outputs rather than to redistribute the originals. That defense has gained some traction in early court rulings, though the legal landscape remains unsettled and highly dependent on the facts of each case.

Two early decisions in California have leaned in favor of AI companies, suggesting that training on copyrighted material may fall within fair use under existing U.S. law. But those rulings do not end the debate. Copyright law was written long before modern generative AI existed, and judges have been left to stretch decades-old doctrine to fit new technical realities.

That uncertainty means the New York case could be important even if it does not immediately change the law. A different federal court will be weighing the dispute, and a new judge may view the facts, the licensing history and the alleged use of the books differently from the California courts.

Why this case could be different from other AI lawsuits

This case stands out because it involves a company that already had a relationship with the plaintiffs through book search and digital distribution products. Instead of a simple allegation that a tech platform scraped the open web, the complaint says Google had access to books through arrangements that were never meant to cover model training.

That nuance matters. If a court concludes that a company cannot stretch a narrow license for search or preview purposes into an AI training right, the decision could affect not just Google but other AI firms with similarly layered content agreements.

The plaintiffs also point to an internal Google document they say shows the company understood the risks. In that document, Google allegedly warned that using copyrighted books for AI training could be “highly problematic” and might expose the company to enormous penalties.

The lawsuit cites an internal Google assessment that reportedly flagged potential exposure in the tens of billions to hundreds of billions of dollars if copyrighted books were used without authorization.

If proven, that detail could strengthen the publishers’ argument that the company knew the legal risks and proceeded anyway. Google has not publicly responded to the complaint.

What the plaintiffs are asking the court to decide

The lawsuit seeks to turn a broad policy argument into a concrete legal question: did Google have the right to use these books for Gemini training, or did it cross the line into infringement?

The answer will likely turn on several overlapping issues:

• whether the books were lawfully obtained for a limited purpose;
• whether AI training is protected as fair use;
• whether any alleged removal of copyright data suggests bad faith;
• and whether damages should reflect the scale of the alleged copying.

Because the complaint is a class action, it could also become a vehicle for a larger group of authors and publishers to seek remedies if the court finds Google’s conduct unlawful. That possibility makes the case more than a one-off dispute over a handful of titles; it could become a template for how rights holders pursue AI companies in the future.

What do the recent court decisions mean?

Recent rulings in California have given AI firms some reason for optimism. In those cases, judges concluded that training on copyrighted works could be treated as fair use, at least under the specific circumstances presented to the court.

But those victories are not sweeping endorsements of the AI industry’s legal position. Fair use is highly fact-specific, and different judges can weigh the same broad issue differently depending on how a company acquired the material, how it used it and whether the use harmed the market for the original works.

Another reason the California decisions do not settle the matter is that copyright law remains rooted in an era before the internet, cloud computing and large language models. Courts are being asked to apply a doctrine built around copying books, recordings and films to systems that absorb immense datasets and produce new text or images at industrial scale.

That mismatch is why these cases matter so much. Each lawsuit adds another data point as judges, publishers and AI companies try to define the boundaries of lawful model training.

How significant was the Anthropic settlement?

The Anthropic case showed that even when AI firms win some legal arguments, they can still face enormous financial exposure if courts conclude that the underlying data collection was unlawful. The company agreed to pay $1.5 billion after being accused of pirating the material used to train its systems, the largest copyright payout in U.S. history.

That settlement changed the conversation around AI training because it made the scale of potential liability impossible to ignore. It also highlighted the difference between a company’s legal theory and the practical costs of defending it in court.

Roughly half a million writers were eligible for payments of at least $3,000 under that settlement, though many opted out in order to keep pursuing separate claims. That detail suggests some rights holders believe individual settlements may not fully resolve the broader issue of how AI companies acquired training data in the first place.

What the Anthropic deal signals for Google

For Google, the Anthropic outcome is a warning as well as a precedent to watch. It suggests that courts and plaintiffs alike may be willing to attach serious monetary value to unauthorized data use, especially when the alleged infringement involves books and other highly valuable creative works.

Still, the legal posture of every case is different. A settlement does not establish a universal rule, and Google could continue to argue that its use of books for training falls within lawful fair use or within the scope of existing content relationships.

Timeline: how the dispute over AI training has escalated

The current case is the product of several years of tension between the publishing industry and AI developers. As generative AI systems became more capable, the question of what they were trained on moved from a technical issue to a central legal and business dispute.

Who are the plaintiffs, and why do they matter?

The plaintiffs in the Google case represent a mix of major publishing houses, academic content providers and an individual author, which gives the complaint both commercial and creative weight. Their involvement suggests that concerns about AI training are not limited to one corner of the literary world.

Hachette and Elsevier are among the best-known names in trade and academic publishing, while Cengage is a major education company whose materials are widely used in classrooms. Scott Turow brings the perspective of a prominent novelist and lawyer, and S.C.R.I.B.E. adds another rights-holder voice to the group.

That broad coalition matters because AI training disputes often split into two different conversations: one about the economics of publishing, and another about authorial control over individual works. In this suit, both concerns are present.

What happens next?

For now, the case enters the normal early stages of federal litigation, where Google is likely to challenge the claims and the plaintiffs will try to preserve the broadest possible reading of their complaint. The first major fights may involve whether the case can move forward as a class action and how much discovery the court will allow into Google’s training data and internal communications.

Those details may prove decisive. AI training cases often turn on what the companies knew, what data they used and whether they can document permissions or licenses. If the court allows extensive discovery, the litigation could surface more information about how Google assembled Gemini’s training set and how it evaluated legal risk.

The outcome may also shape negotiating leverage across the AI industry. If publishers can win traction against Google, other companies may face stronger pressure to secure licenses, limit training datasets or redesign their data pipelines to reduce legal exposure.

Why the industry is watching closely

The AI sector is watching because this is not just a copyright dispute; it is a question about the economics of building generative systems at scale. If courts increasingly require licenses for high-quality training material, the cost of developing frontier models could rise sharply.

At the same time, publishers are watching because the case could determine whether their archives become a new source of licensing revenue or remain a contested resource that AI companies can use under fair use. The broader market implications are substantial either way.

The bigger legal and business stakes

The fight over Gemini training sits at the intersection of copyright law, product strategy and the future of digital content. If the plaintiffs prevail, AI companies may have to rethink how they source data, how they document permissions and how they explain model development to investors and regulators.

If Google wins, the ruling could strengthen the industry’s position that model training is transformative and protected by fair use. But even a win would not erase the pressure to negotiate licenses, because many rights holders are already treating AI training as a commercial market rather than a legal gray area.

That shift is already visible in the lawsuits themselves. Creators are no longer asking only whether AI companies can use their work; they are increasingly asking what the licensing value of that work should be if it is used to build a product that competes with books, articles and reference materials.

For Google, the stakes are especially high because Gemini is central to the company’s effort to keep pace with rivals in generative AI. Any ruling that constrains training data or increases costs could affect not only legal exposure but also product development speed and model quality.

Bottom line

The Google lawsuit is another major chapter in the battle over whether AI companies can train models on copyrighted books without explicit permission. It also tests whether a company’s old content relationships can be repurposed for a new generation of AI products, or whether that leap requires fresh consent, licensing and compensation.

With courts still divided and the law still catching up to the technology, the answer is far from settled. But this case, filed in New York by a coalition of major publishers and authors, could help define the next phase of the AI copyright fight.