Meta is pushing a bold new defense in its ongoing copyright lawsuit, arguing that uploading pirated books via BitTorrent during LLM training qualifies as fair use. The company claims that BitTorrent's peer-to-peer nature makes uploading unavoidable and technically necessary for downloading the massive datasets needed for AI training.

This argument comes after Meta previously won on the core issue of using pirated books for training Llama models. Authors like Sarah Silverman and Christopher Golden sued in 2023, but courts ruled the training process itself was transformative fair use. Now Meta faces claims about distributing those books through BitTorrent, which automatically uploads data to other users during downloads.

Meta's attorneys argue the BitTorrent uploads were simply part of obtaining training data that courts already deemed fair use. They point to author depositions where writers admitted no harmful outputs from Meta's models. The authors counter that Meta delayed raising this defense until after discovery deadlines, calling it an improper tactic. Judge Vince Chhabria must now decide whether this 'fair use by technical necessity' argument can proceed, a ruling that could impact many AI companies using shadow libraries for training data.