Google Faces Class Action Over Books Used To Train Gemini

The plaintiffs in the case include some of the most influential names in the publishing industry: Hachette Book Group, Cengage Learning, and Elsevier. They are joined by acclaimed novelist Scott Turow and his company, S.C.R.I.B.E. The core of their argument rests on the claim that Google leveraged its vast repositories—specifically Google Books, Google Play Books, and Google Scholar—to ingest massive quantities of text to refine Gemini’s capabilities. According to the complaint, this use far exceeded the original purposes for which the works were provided to Google, transforming a collaborative digital library effort into a proprietary training ground for a commercial AI product.

The Core Allegations and Internal Disclosures

The lawsuit brings four specific counts against Google. Three of these counts allege unauthorized reproduction under the Copyright Act. These charges cover the systematic copying of works through Google’s internal services, the downloading of web-scraped content, and the subsequent copying that occurs during the iterative training process of the AI model. The fourth count alleges a violation of the Digital Millennium Copyright Act (DMCA), specifically accusing Google of removing copyright management information (CMI) from the works to facilitate their use in training datasets.

One of the most striking elements of the filing is the inclusion of what are described as internal Google communications. The plaintiffs cite documents where Google employees allegedly expressed internal anxiety regarding the legal risks of their training methods. One quoted document suggests that using books from Google Play Books for AI training was viewed internally as "highly problematic for Google," with the potential for statutory damages ranging from "tens of billions to hundreds of billions of dollars."

Furthermore, the complaint attributes a statement to a lead engineer for Gemini, who reportedly told colleagues that the company should not "do deals for data we already have or already possess." If proven true, these statements could suggest a "willful" infringement, a legal distinction that significantly increases the potential for higher statutory damages in copyright litigation.

The Evolution of the Conflict: A Chronology

The tension between Google and the publishing world is not a new phenomenon, but the advent of generative AI has reframed the debate. To understand the current lawsuit, one must look at the timeline of Google’s relationship with digital text:

1. The Google Books Project (2004): Google began scanning millions of books from university libraries. This led to a decade-long legal battle with the Authors Guild. In 2016, the courts ultimately ruled in Google’s favor, stating that displaying "snippets" of books constituted "fair use" because it was transformative and served as a search tool rather than a replacement for the books themselves.
2. The Shift to Generative AI (2022–2023): With the rise of OpenAI’s ChatGPT, Google accelerated the development of its own LLMs, eventually rebranding its efforts under the "Gemini" umbrella. Unlike the original Google Books search index, Gemini generates full-length text, summaries, and creative works, which publishers argue directly competes with the original authors.
3. The Rise of Crawler Controls (Late 2023): Google introduced "Google-Extended," a robots.txt token allowing webmasters to opt out of having their site content used for AI training. However, publishers argue this is an "opt-out" rather than "opt-in" system, which they claim is a reversal of traditional copyright protections.
4. The July 2024 Filing: The current class action in New York represents a strategic move by publishers to address specific grievances that they feel were not adequately covered in previous or concurrent litigation in California.

The Technical Gap: Why Opt-Outs Failed to Protect Content

A critical point of contention in the lawsuit is the failure of standard technical safeguards like robots.txt or the Google-Extended token to protect the works in question. Google has frequently pointed to these tools as evidence of its commitment to creator control. However, the plaintiffs argue that these tools are irrelevant in this specific context for two reasons.

First, a significant portion of the data used to train Gemini allegedly came from works supplied directly to Google through various partner agreements (such as those for Google Play Books). Because these works were hosted on Google’s own servers under specific distribution agreements, robots.txt—a tool designed for public web crawling—had no bearing on how Google used that internal data.

Second, the complaint alleges that Google utilized "Common Crawl" datasets that included pirated copies of books and articles hosted on third-party "shadow libraries" and pirate sites. Since these copies were hosted on domains not owned by the publishers, the publishers had no way to use robots.txt to prevent Google from scraping them. The publishers argue that Google’s reliance on these "laundering" sites was a deliberate attempt to bypass paywalls and licensing fees.

The "Fair Use" Defense and the Transformation Argument

Google has historically relied on the "Fair Use" doctrine to defend its data ingestion practices. In a policy paper published in June 2024, Google argued that training AI on public web data is a "transformative, non-expressive use." Under U.S. copyright law, fair use is determined by four factors: the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the potential market for the work.

Google contends that AI models do not "copy" the text in the traditional sense; rather, they learn the statistical relationships between words and concepts to create something entirely new. This, they argue, is highly transformative.

However, the publishers and authors in the Gemini lawsuit strongly disagree. They argue that Gemini is a commercial product that directly threatens the market for their books. By providing summaries, answering questions based on the text, and mimicking the styles of specific authors, Gemini acts as a market substitute. The plaintiffs suggest that while the 2016 Google Books ruling protected the creation of a search index, it does not extend to the creation of a generative engine that can replicate the expressive value of the original works.

Broader Industry Implications and Potential Impact

This lawsuit is not an isolated event but part of a broader "AI vs. Intellectual Property" movement. The outcome of Hachette v. Google will likely set a precedent for how "training data" is handled globally.

• Financial Consequences: If the court finds Google liable for willful infringement, the statutory damages could be astronomical. With millions of works involved, even a fraction of the $150,000 maximum per-work fine for willful infringement could reach the "hundreds of billions" mentioned in the internal documents.
• The Licensing Model: A victory for the publishers would likely force Google and other AI developers to adopt a licensing model. This would mirror the music industry’s transition to streaming, where platforms pay royalties for the content they use. We have already seen the beginnings of this with OpenAI striking deals with News Corp, Reddit, and Axel Springer.
• Training Data Scarcity: If tech companies are barred from using copyrighted books without permission, the "well" of high-quality training data may dry up. This could lead to a divide between companies with the capital to buy licenses and smaller startups that cannot afford the entry price, potentially stifling innovation in the AI sector.
• Legislative Pressure: This litigation is putting pressure on the U.S. Congress to clarify the Copyright Act. Lawmakers are currently debating whether new categories of "AI rights" need to be established or if existing laws are sufficient to cover the nuances of machine learning.

Looking Ahead: The Next Legal Steps

As the case moves forward in the Southern District of New York, the legal community is watching closely for Google’s formal response. The company is expected to either file an answer to the complaint or a motion to dismiss, likely leaning heavily on the "fair use" precedents established in the Northern District of California.

In 2024 and early 2025, other judges in California ruled in favor of AI companies in similar cases involving Anthropic and Meta, finding that the mere act of training on copyrighted data was generally protected. However, the New York court may take a different view, especially regarding the "pirated library" claims and the alleged internal admissions of legal risk.

The plaintiffs have requested a detailed accounting of every work used to train Gemini and a court order to delete any unauthorized copies. Should the court grant an injunction, it could theoretically force Google to "unlearn" or retrain its models—a process known as "machine unlearning" that is technically complex and incredibly expensive.

For now, the publishing industry remains firm in its stance. As the Association of American Publishers stated upon the filing of the suit, the goal is to ensure that "the laws of the land apply to the most powerful companies in the world just as they do to everyone else." The battle over Gemini is more than just a dispute over digital files; it is a fundamental debate over the value of human creativity in the age of artificial intelligence.