Sarah Silverman sues OpenAI and Meta over pirated book training data

If you ask ChatGPT to summarize a contemporary novel and it spits out chapter-by-chapter details in seconds, that might look like magic to you. To the author who wrote it, it looks like a smoking gun.

Comedian and author Sarah Silverman, alongside writers Christopher Golden and Richard Kadrey, just filed dual copyright infringement lawsuits against OpenAI and Meta in US District Court. The trio alleges that both companies used their copyrighted books without permission to train their flagship AI models.

Lawyers Joseph Saveri and Matthew Butterick are handling the suits. The duo is already leading high-profile class actions on behalf of coders and visual artists, and they are now zeroing in on text datasets.

The shadow library trail

The complaints hone in on how these massive datasets were allegedly scraped in the first place.

According to the filings, OpenAI's ChatGPT and Meta's LLaMA were trained on datasets sourced from notorious "shadow library" websites like Bibliotik, Library Genesis, and Z-Library, which circulate copyrighted titles in bulk via torrents.

In the OpenAI lawsuit, the plaintiffs included exhibits showing that ChatGPT effortlessly generated detailed summaries of their books upon request, including Silverman's The Bedwetter, Golden's Ararat, and Kadrey's Sandman Slim. Crucially, the suit points out that the chatbot reproduced this material while omitting all original copyright management information.

Meta faces an even more awkward paper trail. When Meta released its LLaMA models in February, its research paper openly listed ThePile as one of its training data sources. The lawsuit notes that EleutherAI, the creators of ThePile, previously documented that part of the dataset was compiled directly from a copy of the private tracker Bibliotik. The complaints level six counts each against both companies, demanding statutory damages and restitution of profits.

Why it matters

This lawsuit goes straight after the AI boom's biggest dirty secret: the provenance of training data.

For authors and creators, this is about drawing a hard line before AI models swallow their entire life's work for free. Fair use has long been Silicon Valley's default shield, but pointing directly to illicit shadow libraries makes that defense a lot harder to sell in court. If training on pirated repositories is deemed wholesale copyright infringement, AI giants could face crippling statutory damages or even be forced to purge models and retrain from scratch.

For everyday users and builders, the downstream ripple effects are real. If courts decide scraping books requires explicit licensing, the era of dirt-cheap, effortlessly trained open-source models could take a massive hit, and data acquisition costs will skyrocket.

We think the legal strategy here is razor-sharp. Instead of getting bogged down in abstract philosophical debates over whether machine learning mimics human reading, the lawyers are focusing on basic, traceable piracy. Tech giants cut corners to move fast, and now that paper trail is catching up with them. It looks like the free ride on copyright-protected content is finally coming to an end.