AI and Copyright Law

This week’s coverage focused on a high‑profile copyright suit filed May 5, 2026, by Scott Turow and major publishers accusing Meta of training its Llama model on copyrighted books and journal articles taken from pirate sites (LibGen, Anna’s Archive), seeking classwide statutory damages, an injunction and destruction of infringing copies; Meta says courts have found training can be fair use and will fight the case. Reporting linked the complaint to earlier revelations about Llama’s use of Books3 and prior lawsuits, and noted a June 2025 federal decision that training on copyrighted books can qualify as fair use while related distribution and storage disputes continue.

Mainstream coverage largely omitted important legal and technical context that’s visible in alternative sources: prior rulings (e.g., Kadrey v. Meta and Bartz v. Anthropic) have already drawn fine distinctions—courts have in some cases deemed training fair use while finding storage/distribution unlawful, and Bartz produced a reported $1.5 billion settlement and per‑work payouts; Meta has said Llama 3 was pretrained on over 15 trillion tokens and outside datasets like LibGen contain millions of books (est. 7.5M). News reports also lacked deeper analysis of dataset composition, the proportion of copyrighted vs. public‑domain material, licensing economics, or quantified risks to authors — gaps that make it hard for readers to gauge legal precedent, technical scale, and the economic stakes; no organized contrarian viewpoints appeared in the coverage, though the existing court decisions and industry defenses that favor a broad fair‑use interpretation are a crucial counterpoint readers should know about.