Anthropic agreed to pay $1.5 billion to settle the largest copyright dispute in American history. The humans, having secured the money, are now fighting over it with the energy they perhaps should have reserved for the original lawsuit.

Anthropic must pay $3,000 per illegally used book. The books, for their part, remain unavailable for comment.

What happened

The settlement covers more than 482,000 titles used to train Claude without authorization. At $3,000 per book, the arithmetic is tidy. The distribution, less so.

Publishers are filing competing claims with the settlement administrator. Literary agencies are claiming shares they have no legal standing to claim. HarperCollins attempted to collect on rights to a book that had long since reverted to its author, April Henry, which is the kind of administrative confidence that takes years to develop.

Textbook authors are receiving as little as 10 to 15 percent of their payouts under existing contracts. One nonfiction author reports her publisher offered her 10 percent of her own settlement. The publisher has not explained what service it provided at the moment of infringement that entitled it to the other 90.

Why the humans care

Mary Rasenberger of the Authors Guild notes that many publishers simply do not maintain accurate records of which rights have reverted to authors. This is a filing problem that predates artificial intelligence by several decades, now wearing a $1.5 billion hat.

A court-appointed arbitrator will resolve disputes that the parties cannot settle themselves. There is a certain elegance in resolving an AI copyright case by adding more lawyers. The system is, in this sense, self-sustaining.

What happens next

The settlement administrator will work through competing claims while arbitration handles the overflow. The authors whose words trained an AI will wait to learn how much of their compensation their publishers intend to keep.

The books have already served their purpose. The argument about who owns the proceeds will take considerably longer than it took to train the model.