The writers' lawsuit against Meta has revealed new details about the use of pirated copies of books, including from the Russian library LibGen, to train the company’s artificial intelligence models.
The writers' lawsuit against Meta has revealed new details about the use of pirated copies of books, including from the Russian library LibGen, to train the company’s artificial intelligence models.
Last month, Meta admitted that it used torrents to download a large dataset known as LibGen, which contains tens of millions of pirated books, Ars Technica reports .
Now, for the first time, unredacted Meta emails have been released, indicating that the company downloaded «at least 81.7 terabytes of data from multiple shadow libraries via the Anna’s Archive website, including at least 35.7 terabytes of data from Z-Library and LibGen,» the authors’ lawsuit states. It also states that «Meta also previously downloaded 80.6 terabytes of data from LibGen.»
«The scale of Meta’s illegal torrenting scheme is staggering,» the writers note. They argue that «much smaller acts of data piracy—just 0.008 percent of the number of copyrighted works copied by Meta—led to judges referring the case to the U.S. Attorney’s Office for a criminal investigation.»
The company had previously sought to block publicity about using pirated books to train its AI models. However, the judge rejected Meta’s request, saying the company insisted on redacting the materials not to protect its business interests but to «avoid negative publicity.»
Meta revealed in a research paper that it trained its large Llama language model on snippets from Books3, a dataset of about 196,000 books scraped from the internet. However, it had not previously publicly indicated that it was loading the data directly from LibGen.
