Anthropic's $1.5 billion book settlement has run into a problem that has nothing to do with Anthropic. Authors and publishers are sending the settlement administrator mutually exclusive claims to the same titles, the New York Times reports, because many publishers do not keep accurate records of which rights have reverted to authors — that is the assessment of Mary Rasenberger, head of the Authors Guild. Money a court ordered paid out for books cannot be distributed until the industry can establish who holds the rights to those books, and right now it cannot.
The split is worst for textbook authors. Under the terms of some of their contracts, they stand to receive as little as 10–15% of the payout attached to their own work. Literary agencies are also demanding a share, according to Writer Beware, despite having no legal basis for the claim.
The individual cases are more telling than the aggregate. The author April Henry found that HarperCollins had claimed a book whose rights had returned to her long ago. Another nonfiction author said his publisher offered to pass along 10%.
The settlement itself sits on a split ruling. The court found Anthropic's use of illegally obtained books unlawful, but treated training on legally purchased books as fair use. Disputes the parties cannot settle between themselves go to a court-appointed arbitrator.
That split is the part most coverage skips, and it changes what the fight is about. The money is compensation for how the books were acquired, not for what was done with them. It is a piracy payout wearing a training-data label. Claims filed against it on the strength of publishing contracts are being read against documents drafted for a world of print runs and reversion clauses, by an administrator with no authoritative registry to check them against.
Which is why the headline number is the least informative thing here. A $1.5 billion pool divided among disputed claimants, with some authors contractually entitled to a tenth of what their own title draws, is not a $1.5 billion settlement from the perspective of the people the suit was nominally about. It is a fixed pot that authors and publishers now have to fight each other over, one title at a time, in front of an arbitrator. Anthropic's liability is settled; the claimants' is not.
The question the reporting does not answer is what happens to a book where nobody can prove anything. If a publisher asserts rights it no longer holds and the author cannot document the reversion, there is no described mechanism for who wins by default — the arbitrator is named as the venue, not the rule. On a catalogue this size, the number of titles in that condition is unlikely to be small.
The precedent forming here is not about model training. It is about whether the book industry can produce a list of who to pay. The next developer that wants to license text legally rather than settle for taking it will need that list, and this settlement is the first public demonstration that it does not exist.