Court documents unredacted in Kadrey v. Meta reveal that Meta trained its Llama language model using data downloaded from Library Genesis (LibGen), a Russian-origin piracy archive, with evidence of deliberate internal obfuscation and executive-level approval up to Mark Zuckerberg. Meta had initially disclosed use of Books3 (196,000 scraped books) but not LibGen torrenting; discovery reveals engineers hesitated to access pirated data from corporate devices, internal discussions escalated to CEO, and Meta seeded pirated files back to torrent sites. The judge called Meta's redaction efforts "preposterous" and noted Meta sought to conceal the information to "avoid negative publicity" and protect "negotiating position with regulators." This directly contradicts Meta's prior public positioning that it used only publicly available, legally-obtained materials, and exposes training-data provenance liability.
309 other tracked developments touching the same claims.