Google Faces Class Action Over Books Used To Train Gemini

A coalition of prominent publishers and authors has initiated a proposed class-action lawsuit against Google, alleging the tech giant unlawfully copied millions of copyrighted books and journal articles to train its advanced artificial intelligence model, Gemini. The complaint, filed in the U.S. District Court for the Southern District of New York on July 10, centers on claims that literary works supplied through Google Books, Google Play Books, and Google Scholar, alongside content scraped from the web, were exploited without permission or compensation, thereby infringing on intellectual property rights. This legal challenge underscores the escalating tensions between content creators and AI developers over the foundational data fueling generative AI technologies.

The Heart of the Allegation: Unauthorized Data Sourcing

The lawsuit was brought forth by Hachette Book Group, Cengage Learning, Elsevier, acclaimed novelist Scott Turow, and his company, S.C.R.I.B.E. These plaintiffs represent a significant cross-section of the publishing industry, from major trade publishers to academic powerhouses and celebrated individual authors. Their core contention is that the works provided to Google’s various platforms were intended for specific, limited uses—such as search indexing, previewing, or digital sales—and that training a commercial AI model like Gemini falls far outside the scope of those initial agreements.

The complaint specifically targets three distinct methods by which Google allegedly acquired the copyrighted material:

1. Direct Supply via Google Services: Millions of books and journal articles were reportedly ingested from Google Books, Google Play Books, and Google Scholar. Publishers contend that their contracts or terms of service for these platforms did not grant Google the right to use this content for AI training.
2. Web Scraping: The lawsuit claims Google also copied works obtained through web scrapes, including content found on pirate sites and from paywalled subscription libraries. This method bypasses any direct contractual relationship and relies on the assumption that publicly accessible (or illicitly accessible) web content is fair game for data collection.
3. Copying During Training: Regardless of the initial acquisition method, the act of reproducing and processing these works within Google’s systems for the purpose of training Gemini is cited as a separate act of unauthorized reproduction.

The Association of American Publishers (AAP) publicly announced the lawsuit on the same day it was filed, highlighting the collective concern among its members regarding what they perceive as widespread copyright infringement in the AI development space. This move signifies a coordinated effort by the publishing industry to assert its rights in the face of rapidly evolving AI capabilities. Google has yet to issue a public statement specifically addressing the complaint, and it is crucial to note that no court has yet ruled on the validity of these claims. The central legal question revolves around whether permission granted for one specific use automatically extends to the entirely new application of training a sophisticated AI model.

Chronology and Context: A History of Content Disputes

Google’s current legal battle is not its first skirmish with copyright holders over digitized books. The present class action evokes echoes of the landmark Authors Guild v. Google lawsuit, which began in 2005. That case challenged Google’s ambitious project to scan millions of books from library collections to create Google Books, a massive digital library that allowed users to search and view snippets of copyrighted works. After years of litigation, including a proposed settlement that was ultimately rejected, the Second Circuit Court of Appeals ruled in 2015 that Google’s digitization and display of book snippets constituted fair use, primarily because it was transformative and provided public benefit by facilitating search without supplanting the original works. This earlier precedent, however, focused on the act of displaying snippets for search, not on using entire works to train a generative AI model. This distinction is critical to the current dispute.

The rise of generative AI in recent years has ignited a new wave of copyright litigation across multiple creative sectors. Companies like OpenAI, Stability AI, and Microsoft, alongside Google, have faced lawsuits from artists, writers, programmers, and news organizations. These cases typically allege that AI models are trained on vast datasets of copyrighted material without permission or compensation, leading to outputs that can mimic, reproduce, or compete with the original works. The plaintiffs in these cases argue that such training constitutes direct infringement, while AI developers often invoke the "fair use" doctrine, contending that training is a "transformative" use that does not require licensing.

The timeline leading up to the current lawsuit highlights the growing friction:

• Early 2023: Generative AI models like ChatGPT gain widespread public attention, demonstrating unprecedented capabilities in text and image generation.
• Late 2023 – Early 2024: A surge of copyright lawsuits is filed against AI companies by various content creators, including visual artists, authors, and news publishers.
• June 25, 2024: Google publishes a policy paper, explicitly defending AI training on public web data as a "transformative, non-expressive use" under fair use protections. This paper also mentions mechanisms like Google-Extended, a robots.txt token, allowing websites to opt out of AI training.
• June 2024: Digital Content Next (DCN), a trade association representing premium publishers, sends a cease and desist letter to the Common Crawl Foundation, a non-profit that maintains a massive open dataset of web crawls, demanding they stop scraping their content for AI training. DCN asserts that copyright law operates on an "opt-in" rather than an "opt-out" basis.
• July 10, 2024: Hachette, Cengage, Elsevier, Scott Turow, and S.C.R.I.B.E. file the class-action lawsuit against Google. The Association of American Publishers announces the filing, emphasizing the industry’s unified stance.

Legal Pillars of the Complaint: Copyright and DMCA Violations

The complaint against Google brings forth four distinct counts, primarily rooted in the Copyright Act and the Digital Millennium Copyright Act (DMCA):

1. Unauthorized Reproduction (Google Books & Other Services): This count alleges that Google illegally reproduced copyrighted works initially obtained through its own services (Google Books, Play Books, Scholar) for the purpose of AI training. The plaintiffs argue that the original agreements for these services did not encompass such broad, transformative use.
2. Unauthorized Reproduction (Web Scraping): This count focuses on works allegedly copied by Google through web scraping, including from pirate sites and paywalled subscription libraries. This method, the plaintiffs contend, involves direct infringement, as no prior permission or license existed for these copies.
3. Unauthorized Reproduction (During Training): This count specifically targets the act of copying and processing the works within Google’s computational infrastructure to train Gemini. Even if initial acquisition methods were somehow deemed permissible (which the plaintiffs dispute), the internal reproduction for training is argued to be a separate act of infringement.
4. Removal of Copyright Management Information (DMCA Violation): This count alleges that Google removed or altered copyright management information (CMI) associated with the works, in violation of the DMCA. CMI includes information identifying the author, copyright owner, and terms and conditions for use. Its removal can make it harder to trace the origin and ownership of content, hindering copyright enforcement.

The plaintiffs are seeking significant remedies, including monetary damages for past infringement, a permanent injunction to prevent future unauthorized use, a detailed accounting of all copyrighted works used to train Gemini, and court orders mandating the deletion of any unauthorized copies.

Damning Internal Documents and the "Fair Use" Debate

Perhaps one of the most compelling aspects of the complaint comes from its alleged citations of internal Google documents. The filing quotes what it describes as a Google internal document that labels the use of books from Google Play Books for AI training as "highly problematic for Google," with potential fines ranging from "$10Bs-$100Bs." Another line, attributed to Gemini’s lead engineer, reportedly states, "we don’t do deals for data we already have or already possess." While these quotes are presented by the plaintiffs and have not been publicly verified by Google, if proven accurate, they could suggest a corporate awareness of the legal risks involved and a deliberate strategy to avoid licensing negotiations for content already within their grasp, regardless of original use agreements.

Google’s broader defense, outlined in its June 2024 policy paper, hinges on the doctrine of "fair use." Fair use is a legal defense to copyright infringement, allowing limited use of copyrighted material without permission for purposes such as criticism, comment, news reporting, teaching, scholarship, or research. Courts typically consider four factors:

1. The purpose and character of the use: Is it commercial or non-profit educational? Is it "transformative" (i.e., does it add new meaning or expression, or alter the original with new purpose or character)? Google argues AI training is highly transformative.
2. The nature of the copyrighted work: Factual works tend to have broader fair use allowances than highly creative works.
3. The amount and substantiality of the portion used: How much of the original work was used?
4. The effect of the use upon the potential market for or value of the copyrighted work: Does the use harm the market for the original?

Google’s "transformative, non-expressive use" argument directly addresses the first factor, positing that AI training fundamentally changes the purpose of the copyrighted material from expressive content to raw data for pattern recognition, without directly reproducing the original expression in its outputs. However, critics argue that AI outputs do compete with original works and that the training itself involves unauthorized reproduction, irrespective of output. The publishers in this lawsuit are likely to argue that training a commercial AI model that can generate text in the style of authors, or summarize scholarly articles, directly impacts the market for their copyrighted works.

The Limitations of Crawler Controls and the Opt-Out Debate

The lawsuit also highlights the inadequacy of current technical controls, such as robots.txt files, in addressing the alleged infringements. Google’s Google-Extended robots.txt token is designed to allow website owners to opt out of having their content used for future Gemini training and some grounding uses. However, as the complaint points out, this mechanism is irrelevant to the primary sourcing methods at issue:

• Direct Agreements: Books supplied directly to Google via existing agreements for Google Books or Play Books are not governed by robots.txt files, as they were provided through direct contractual or licensing channels.
• Web Scrapes from Third-Party Domains: The claims regarding web scrapes from pirate sites or subscription libraries involve content hosted on domains other than the publishers’ own websites. A publisher’s robots.txt file on their domain cannot regulate how content is used once it appears on a third-party site or in a dataset like Common Crawl.

This distinction reinforces the argument made by organizations like Digital Content Next, which insists that copyright law is fundamentally an "opt-in" system, not an "opt-out" one. Content creators should grant permission for specific uses, rather than being forced to actively block unwanted uses of their work. The fact that BuzzStream data in January indicated that 79% of top news sites already block at least one AI training bot suggests widespread concern among publishers, but these technical blocks are often insufficient for the types of data acquisition methods alleged in the Google lawsuit.

Precedent and the Path Forward

The current lawsuit also builds upon, and seeks to differentiate itself from, earlier rulings concerning AI training data. In 2024, two Northern California district judges issued rulings on fair use in the context of AI training. One ruling, involving Anthropic, denied summary judgment on claims related to pirated central-library copies, suggesting that using illicitly obtained content for AI training might not easily qualify as fair use. Another ruling, involving Meta, found certain training uses to be fair on the specific record before it but stressed that the decision was narrow and specific to those plaintiffs and their particular circumstances.

The plaintiffs in the Google case stated that they chose to file in New York after initially considering intervening in the ongoing "In re Google Generative AI Copyright Litigation" in California. Their decision to pursue a separate suit in New York suggests they believe their claims, particularly those concerning content directly supplied to Google’s own services and the alleged internal corporate knowledge, fall outside the scope of that proposed class and warrant a distinct legal challenge.

The next procedural step will involve Google’s formal response to the complaint, which could be an answer denying the allegations or a motion to dismiss the case on legal grounds.

Broader Implications for the AI and Publishing Industries

This class-action lawsuit carries profound implications for the future of artificial intelligence development, copyright law, and the economic landscape of content creation.

• The Future of AI Training Data Acquisition: A ruling in favor of the publishers could fundamentally reshape how AI models are trained. It might necessitate a shift from mass, unlicensed data scraping to a licensing-based model, where AI developers would need to negotiate "data deals" with copyright holders. This could lead to a new revenue stream for creators but also significantly increase the cost and complexity of AI development.
• Economic Impact and Valuation of Content: The outcome could establish a precedent for how the immense value generated by AI is distributed. If AI models are found to be built on infringed content, it could lead to massive damages awards, potentially in the "10Bs-$100Bs" range, as Google’s internal documents allegedly suggest. This would underscore the economic value of copyrighted works as essential inputs for AI.
• Clarification of Fair Use in the AI Era: This case, alongside others, will be pivotal in defining the boundaries of the fair use doctrine in the context of generative AI. The legal system is grappling with how to apply existing copyright frameworks to novel technological uses.
• Creator Rights and Compensation: At its heart, the lawsuit is about compensating creators for their intellectual labor. Many authors and artists feel their work is being exploited to enrich tech companies without their consent or fair remuneration. A successful outcome for the publishers could empower creators across industries to demand better protections and compensation.
• The "Black Box" Problem and Transparency: The plaintiffs’ request for a detailed accounting of the works used to train Gemini highlights a critical challenge in AI litigation: the "black box" nature of AI models. Proving exactly what data an AI was trained on can be difficult, and court-ordered transparency could set an important precedent for future AI development.
• Impact on AI Model Diversity and Development: If access to diverse training data becomes severely restricted or prohibitively expensive, it could impact the capabilities, diversity, and quality of future AI models. This raises questions about whether robust AI development can proceed ethically and legally without broad access to existing knowledge and creative works.

The lawsuit by Hachette, Cengage, Elsevier, and Scott Turow against Google is more than just another copyright dispute; it is a critical test case that will help shape the regulatory and ethical framework for artificial intelligence, potentially influencing the balance of power between technology giants and the creators whose works underpin the digital world. The outcome will be closely watched by industries worldwide as they navigate the complexities of the AI revolution.