Zuckerberg turns to YouTube to defend Meta's AI training

When a Big Tech CEO has to sit across from hostile trial lawyers and explain exactly what pirated books his flagship AI ate for breakfast, things get uncomfortable fast.

According to recently unsealed deposition snippets from the Kadrey v. Meta Platforms copyright case, Meta chief Mark Zuckerberg reached for an unexpected defense when grilled over the company's training datasets: YouTube.

The YouTube defense and playing dumb on LibGen

During his deposition late last year, Zuckerberg argued against blanket bans on large datasets by comparing them to video platforms. He noted that while YouTube occasionally ends up hosting pirated content for a period of time, the platform actively tries to take it down, and the vast majority of its catalog is properly licensed.

Zuckerberg then asked whether Meta would want a policy banning employees from using YouTube just because some pirated content slips through. His answer was no, insisting that a blanket ban is not the right move.

That defense was directly aimed at Meta's use of Library Genesis (LibGen), a notorious shadow library that offers millions of unauthorized, copyrighted e-books from major educational publishers. LibGen has already lost multiple lawsuits, faced domain seizures, and racked up tens of millions in infringement penalties.

When pressed directly by prominent plaintiffs' attorney David Boies, Zuckerberg claimed he had never really heard of LibGen and simply had no specific knowledge of the platform.

Unsealed filings paint a very different picture

However, the latest amended complaint filed by plaintiffs—including bestselling authors Sarah Silverman and Ta-Nehisi Coates—tells a far less innocent story.

Court documents allege that Zuckerberg personally signed off on using LibGen to train Llama models, despite explicit pushback from Meta's own AI executives and researchers. Internal communications cited in the filing show employees warned that LibGen was a "data set we know to be pirated" and flagged that touching it could severely undermine Meta's negotiating stance with regulators.

The allegations get even messier. Plaintiffs claim Meta used LibGen to train Llama 3 and is actively using it for next-generation Llama 4 models. Meta allegedly cross-referenced pirated LibGen titles against legitimate commercial catalogs to decide whether paying publishers for licenses was even worth the money.

To make matters worse, the lawsuit alleges Meta researchers attempted to conceal the use of copyrighted works by inserting "supervised samples" during fine-tuning. Meta also allegedly grabbed e-books from Z-Library for training as recently as April 2024, despite the site being targeted in federal criminal indictments and domain seizures.

Why it matters

In our view, Zuckerberg's comparison between LibGen and YouTube falls flat on its face.

YouTube operates as a legitimate user-generated content platform under DMCA safe-harbor protections, investing heavily in automated copyright enforcement. LibGen, on the other hand, is an illicit aggregator specifically designed to distribute pirated textbooks and novels. Conflating the two looks like a deliberate attempt to blur the lines between accidental infringement and systematic scraping of pirate hubs.

This dispute cuts straight through the "fair use" defense that Big Tech has relied on to train generative AI. If the plaintiffs prove that Meta employees knew datasets were pirated, compared them against licensed catalogs to skip paying, and actively masked the training trails, courts are unlikely to treat this as good-faith transformative use.

For developers and enterprises relying on Meta's open weights, this is an alarming development. Llama is the bedrock of open-source AI today. If a court rules its underlying training data unlawful, Meta could face crushing financial damages or even be forced to retrain from scratch, creating massive legal uncertainty for the entire downstream ecosystem.