Of all the things Anthropic has done, this one probably shocks regular people the most. The AI giant bought millions of used books, cut off their spines, scanned the contents to train its models, and then destroyed the originals.
Most people hear that and think it's awful. But Anthropic was actually following copyright law, which lets you digitize a book you own as long as you don't create an extra copy. In a high-profile case, a federal judge ruled this part of Anthropic's process was fair use.
The bigger question has always been what happens after the training, when models use what they've learned to generate new content.
The US Copyright Office examined that issue last year and identified the biggest weakness in AI companies' argument. One of the four fair-use tests asks whether the new work harms the market for the original. The Copyright Office concluded that using copyrighted works to create commercial content that competes with those works "goes beyond traditional fair-use boundaries."
The problem was proving it. Now there's evidence. Researchers recently analyzed more than 14,000 ebooks sold on Amazon between 2023 and 2026. They found many books with substantial amounts of AI-written text — so many that they squeezed the market for everyone else.
The number of books with sales exploded 19-fold, while the total amount of money readers spent on those books grew ninefold. In other words, the market got much more crowded without growing nearly as fast, leaving less revenue for the average title. Books with no detected AI text lost market share as AI books spread, the researchers also found.
That's exactly the kind of market harm the Copyright Office said courts should be looking for. When AI can produce thousands of books at almost no cost, human authors end up competing against a flood of new content (that wouldn't exist without their original work).
Ed Newton-Rex, CEO of the nonprofit Fairly Trained, noted that AI companies have said their models don't financially harm the creators whose work they learned from. This research leaves that argument "well and truly dead," he wrote on X.
Sign up for BI's Tech Memo newsletter here. Reach out to me via email at abarr@businessinsider.com.