Why Tech Giants Can’t Use Pirated Books to Train Models?

(Read time: 3 Minutes)

By Ollie Coull

Generative artificial intelligence has sparked a massive legal war over copyright. Tech companies argue that using public internet data to train Large Language Models (LLMs) is perfectly legal. Authors and publishers argue it is systematic theft. For years, this fight sat in a very confusing legal grey area.

However, a wave of major court rulings has finally drawn a hard line in the sand. The courts have decided that building an AI model is legally fine, but how the tech companies get their training data can lead to total financial ruin.

The landscape completely changed following a massive class-action lawsuit, Bartz v. Anthropic, which ended in a historic $1.5 billion settlement—the largest copyright payout in history. The case explicitly defined when training an AI crosses the line from high-tech innovation into corporate theft.

The Debate Over "Fair Use"

To build an AI that can speak like a human, developers have to feed it trillions of words from books, articles, and websites. When authors sue tech companies for using their writing without permission, the tech companies always use a defense called Fair Use.

Under US copyright law, fair use allows people to use copyrighted material without permission if they are transforming it into something completely new.

At first, tech companies won some major victories using this argument. Judges ruled that using a book to teach an AI how human language works is "transformative." The courts compared it to a human reading books at a library to learn how to write. Because the final AI model doesn't actually copy and paste the book's text when you ask it a question, judges ruled that the act of AI learning was perfectly legal.

The Catch: You Can't Learn From a Stolen Book

However, those rulings contained a massive trap for the tech industry: the fair use defense only works if the company acquired the data legally in the first place.

In the Anthropic lawsuit, lawyers for the authors discovered that the tech company didn't buy the books or use legal digital libraries. Instead, Anthropic had scraped massive datasets from illegal, online "shadow libraries" like Library Genesis (LibGen) and the Pirate Library Mirror—websites used to download pirated books for free.

The court drew an absolute line. The judge ruled that downloading pirated files from illegal websites is a distinct act of copyright infringement. You cannot claim your use of a book is "fair" if your very first step was executing an illegal download.

Facing a trial they were guaranteed to lose and trillions of dollars in potential penalties, Anthropic panicked and agreed to the $13.5 billion settlement. They were forced to pay authors roughly $3,000 per book and ordered by the court to completely delete and destroy the pirated files from their AI training servers.

The New Target: Elsevier v. Meta

This ruling has completely changed how major publishers are attacking tech companies. Instead of arguing about whether AI is good or bad, publishers are focusing entirely on where tech companies get their data.

For example, a massive coalition of the world's largest scientific publishers filed a lawsuit called Elsevier v. Meta in New York. The publishers allege that Meta deliberately hid its identity to download millions of medical journals, textbooks, and academic papers from pirate networks. By focusing on the illegal downloading rather than the abstract concept of AI learning, publishers are easily winning their cases in court.

The Bottom Line

For professionals, engineers, and businesses looking to build or use AI tools, the wild west era of data scraping is officially over. Training a commercial AI tool on data from sketchy or unverified sources is a massive corporate liability. If a company uses an AI model built on a pirated data pipeline, they risk multi-billion-dollar lawsuits and court orders forcing them to delete their entire software infrastructure. In the modern business world, the law is simple: a piece of code cannot clean up a stolen file, and an AI is only as legal as its cleanest source data.

References (APA 7)

Bartz v. Anthropic PBC, No. 3:24-cv-05258 (N.D. Cal. 2025).

Elsevier Inc. v. Meta Platforms, Inc., No. 1:26-cv-03689 (S.D.N.Y. 2026).

Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal. 2025).

U.S. Copyright Act, 17 U.S.C. § 107 (2012) [Limitations on exclusive rights: Fair use].