Ex-Meta Researcher Admits Torrenting Books Dataset in AI Case

A former Meta researcher admitted in a recent deposition that he downloaded and distributed the pirated Books3 dataset while employed at the company, according to a joint court status report reported by TorrentFreak.

Stephen Roller, who worked at Meta from 2017 to 2023, confirmed during his deposition last week that he acquired the material via BitTorrent. The Books3 archive consists of roughly 200,000 copyrighted titles scraped from the private torrent index Bibliotik and used across the tech sector to train artificial intelligence systems.

Roller testified that he could not recall the exact date he downloaded the files, though he acknowledged that the dataset was created in 2020 and that he was not employed by any other company during that window. The admission contradicts Meta's statement to plaintiff Entrepreneur Media in August, when the company claimed it did not believe it had used BitTorrent to obtain the collection.

The litigation against Meta is consolidated before U.S. District Judge Vince Chhabria in California federal court. High-profile plaintiffs, including authors Sarah Silverman and Ta-Nehisi Coates alongside publishers such as Chicken Soup for the Soul, allege that Meta infringed on their intellectual property to train its Llama family of AI models.

For rights holders across entertainment and publishing, the method of data acquisition has become a pivotal legal vulnerability for tech companies. While the court previously found that training AI models on text can fall under fair use, the actual act of reproducing and distributing copyrighted files through BitTorrent networks remains a live infringement claim. As Hollywood Tomorrow tracks the legal fallout of generative AI scraping, the deposition exposes tech firms to renewed scrutiny over how their engineering teams acquired training material.

The disclosure follows a recent court order requiring Meta to identify internal staff who authorized downloads from unauthorized digital libraries. Meta has previously acknowledged that it used portions of the Books3 dataset in developing its Llama architecture.