📰 New article from TorrentFreak
Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’
It turns out the secret sauce behind some of Meta’s AI models might have been a little… unauthorized.
In a series of ongoing lawsuits, authors and publishers are accusing Meta of using pirated books to train its Llama AI models. While Meta previously played it cool—claiming they weren’t sure how the “Books3” dataset (a massive collection of nearly 200,000 pirated titles) was obtained—new testimony has thrown a wrench in that narrative.
A former Meta researcher, Stephen Roller, reportedly admitted in a deposition that he actually torrented the dataset and distributed it. While Meta argues he might have just been downloading it for personal use, the plaintiffs are digging deeper. They aren’t just looking at AI training; they’re investigating whether Meta employees were using “shadow libraries” for internal book clubs.
The key takeaways:
- The “Smoking Gun”: Plaintiffs claim the researcher deleted his copy of the data specifically because of the “contentious” litigation.
- Internal Chatter: Court docs reveal employees joking about sharing pirated books in internal folders.
- The Long Game: This isn’t settling anytime soon—the trial isn’t expected until 2027.
For now, Meta is stuck in a high-stakes game of “he said, she said” over whether they were building cutting-edge tech or just running a very expensive, very illegal digital library.
