The books then arrived. Any undergrad with a good VPN could locate millions of them in less than a minute by pulling them from shadow libraries. And now the biggest copyright settlement in American history.

Judge William Alsup of the Northern District of California gave preliminary approval to a settlement that many Silicon Valley residents had been quietly dreading on a Friday in early September. At least $1.5 billion is a significant amount. It’s a message. The authors whose books were used to train the models that power Claude received about $3,000 for each work that was pirated. The irony that a chatbot named after a French mathematician owes part of its fluency to novelists who were never asked is difficult to ignore.

A group of fiction and nonfiction writers filed the lawsuit, Bartz v. Anthropic, in August 2024. It developed as these cases typically do: slowly, with procedural back-and-forth that would bore anyone who isn’t a copyright lawyer. Anthropic relied on fair use, that antiquated legal safety net that permits restricted copying for things like research, teaching, and criticism.

Judge Alsup also gave them something in June. He decided that using legally acquired copyrighted content to train LLMs could be considered fair use. Anthropic prevailed in that section. However, the copies that were pirated and taken from repositories that the court deemed to be “inherently, irredeemably infringing” were a different matter.

The billion-dollar issue began there. It’s one thing to train on a legally acquired copy of a book. Another is to feed your model from a torrent website. When you put it that way, the difference seems clear, but for years, the industry treated large scraped datasets like an open buffet. Not many questions were posed. The answers were not wanted by investors.

There are three key terms in the settlement itself that are important to consider. The money is paid by Anthropic. The pirated copies are destroyed by Anthropic. Additionally, Anthropic is only released from allegations regarding prior training behavior, which is the point that most people seem to overlook. Anything pertaining to outputs—that is, what the model genuinely generates—is still acceptable. The next wave of lawsuits seems to have just begun.

The implications for the remainder of the industry are more complicated than the headlines imply. The cost of entry for smaller startups is suddenly much higher. Real library licensing is expensive, and once-free data now has a cost and a paper trail. The calculation has changed for authors as well; five years ago, no one could have predicted that their third book would end up feeding a neural network in a San Francisco server farm. Now, it is important to register work.

If willful infringement had been established, statutory damages in this case might have exceeded $15 billion, possibly reaching $75 billion. From that perspective, $1.5 billion nearly seems like a steal. Nearly.

Whether this settlement establishes a true standard or merely buys Anthropic some peaceful space before the next lawsuit is filed is still up for debate. The wars over copyright are just getting started.

The post The AI Copyright Wars: Why Anthropic’s Latest Settlement Could Break the Internet appeared first on Creative Learning Guild.