Publishers Sue Google Over Gemini AI Training Data Copyright Infringement

A coalition of major book publishers and authors has filed a federal lawsuit against Google, accusing the company of copyright infringement in training its Gemini AI models. The case, brought by Hachette Book Group, Elsevier, Cengage Learning, and author Scott Turow in a New York federal court, alleges that Google copied millions of copyrighted books through Google Books and scraped large portions of the web to train Gemini without permission. For AI builders, this lawsuit underscores the unresolved legal risks around training data provenance and the growing pressure for licensed data sourcing.

What happened

The nearly 60-page complaint alleges that Google first copied books as source material through Google Books, using works it obtained for strictly limited purposes. It also claims Google downloaded web scrapes of virtually the entire internet, including from known pirate sources and behind legitimate paywalls, to train its Gemini models. The suit argues these uses exceeded the scope of existing agreements and that Google never informed authors or publishers about the training use. Internal documents allegedly warned that using books to train AI models was "highly problematic for Google," and the plaintiffs cite potential fines up to $100 billion.

The lawsuit follows a previous attempt by Hachette and Cengage in February to join a pre-existing class action. It adds to a growing wave of AI copyright litigation involving companies like OpenAI, Meta, and Anthropic, with varying outcomes so far.

Why AI builders should care

This case highlights fundamental questions about training data provenance and the limits of fair use in AI model development. Legal experts note that proving what was in a training corpus is extremely difficult, and that any fair use argument could be undermined if the data was acquired unlawfully. The outcome could reshape how AI companies source training data, especially for large language models that rely on web-scale corpora.

For builders shipping AI products, the uncertainty around copyright liability for training data creates practical risks. If courts rule against Google, it could set a precedent that forces stricter data provenance audits and licensing requirements across the industry. Conversely, a ruling in Google's favor (similar to Meta's recent fair use win) could embolden broader scraping practices.

Practical implications

AI developers may need to adjust data sourcing strategies now rather than wait for court rulings. Key actions include:

• Conducting data provenance audits to identify copyrighted material in training corpora.
• Negotiating licensing agreements with publishers and content owners, though current offers are reportedly unattractive (around $10 per title for perpetual AI training licenses).
• Building model governance frameworks that track training data lineage and document permissions.

Publishers are likely to push for licensing revenue models and demand transparency in training data sources. The case also raises questions about who bears responsibility for infringing outputs: the model developer or the end user who prompts for copyrighted content.

Caveats

This lawsuit is in its early stages, and no court has ruled on the merits. The fair use defense remains an open legal question, with different outcomes in related cases. A federal judge ruled in Meta's favor in 2025, finding that AI training met fair use requirements, while OpenAI's case is still pending. The outcome of this Google case could take years to resolve, and appeals are likely regardless of the initial ruling.

Additionally, the specific allegations about Google's internal warnings and the scope of web scraping are based on the complaint and have not been proven in court. Google has not yet responded to the lawsuit.