Why Anthropic Paid $1.5 Billion When AI Training Was Ruled Legal

"Fair use." That was the court's ruling at the heart of Anthropic's copyright case: using books and articles to train an AI model is legal under copyright law. Anthropic still paid $1.5 billion, the largest copyright settlement in AI history. It's not as contradictory as it sounds. Anthropic had built an internal database of 70 million volumes sourced from LibGen, a piracy site, to train Claude in its early days. The court's finding: training on books is fine; the problem is how you obtained them. If they came from piracy, that's infringement. That distinction is the ruling's most important finding for the AI industry: training itself is legal; data sourcing is where liability lands. For anyone who has ever written articles or published a book, this ruling says: if an AI company obtained your work through legal channels and used it to train a model, current law says that's permissible. The settlement covers around 500,000 books at about $3,000 each. More than 90% of eligible authors have filed claims. But some authors argue the principle itself is wrong. Fair use was meant to protect commentary, research, and educational use, not to let tech companies train commercial models on entire libraries. Those authors have opted out and plan to sue separately. For other AI companies, Anthropic's case is the only settled answer so far. OpenAI, Meta, and Midjourney are still in litigation with authors and publishers. The answer Anthropic's case gives: if your data is legally obtained, you have a defensible position. If it came from piracy, the legal exposure is a separate question, regardless of how the AI model itself works. The fair use line has been drawn. Where it sits, who it's fair to, and what it gives authors in this framework, that argument is still going.