Artificial Intelligence Daily edition
ARTIFICIAL INTELLIGENCE AI GOVERNANCE
Anthropic Settlement Enters Awkward Phase: Deciding Who Gets Paid
Payout notices for the settlement over books used by Anthropic have uncovered conflicting claims among authors, publishers, and agents. The debate is no longer about an abstract figure: it hinges on contracts, rights reversions, and who owned the rights when each work was obtained.
The $1.5 billion settlement that concluded the class-action lawsuit against Anthropic for obtaining pirated books has reached its least photogenic stage: distributing the money. In early September, the administrator sent claimants a summary for each title, listing the individuals or companies that had requested a share and the percentage each had chosen. This cross-referencing has revealed disagreements that the headline figure of the settlement concealed. Some authors are claiming one hundred per cent because they recovered their rights, while some publishers are requesting half, and some agents have noted their own stake.
The scale of the situation explains why this administrative conflict matters. The fund covers approximately half a million eligible titles out of the millions of copies Anthropic allegedly downloaded from pirate libraries. The baseline calculation is at least $3,000 per title before costs and fees, but this amount does not automatically equate to a cheque for a single person. It can be divided among the author, co-authors, publisher, and other rights holders, and could increase if there are fewer valid claims. The aggregate figure was simple; the chain of rights for each book is not.
Conflicting Percentages, Incorrect Forms, and a 30-Day Deadline to Agree
The standard distribution planned for trade and academic works starts with a fifty-fifty split between the publisher and the author, unless the contract states otherwise. Educational texts do not even have this default split, so discrepancies are more common. Added to this are form-filling errors: several publishers reported that they had mistakenly marked a one-hundred-per-cent allocation, and the administrator is now correcting these percentages. Not every discrepancy indicates bad faith, although none should be accepted without verification.
The procedure grants co-claimants thirty days to try to resolve their differences. The portal accepts reversion letters, contracts, and other documentation; as long as a dispute exists, the payment for the title is withheld. If the parties fail to reach an agreement, the administrator will attempt to facilitate one, and ultimately, a special master appointed by the court may decide the outcome. The sequence is reasonable, but it shifts a documentary burden onto authors and small publishers, many of whom did not expect to have to reconstruct it so quickly.
10 August 2022 Determines Who Held the Right
The decisive date is 10 August 2022, used as the download reference date in the settlement. The Authors Guild maintains that if the rights had reverted to the author before that day, the author should receive one hundred per cent, depending on their contract. If the reversion occurred afterwards, the publisher that held the rights when the work was obtained may retain a valid claim. This is an important distinction: the current owner and the person who held the right at the time of the event are not necessarily the same. Backlists, changes of imprint, and automatic reversions further complicate the proof.
This episode is important beyond the publishing industry because it shows where the abstraction of training data ends. A mass licence or compensation only works if there is a chain of title that can be audited by work and by date. AI systems operate at the scale of millions of files; rights, however, still reside in individual contracts. For providers, buying a corpus does not eliminate the duty to know its provenance. For content creators, keeping contracts, addenda, and reversion letters becomes economic infrastructure, not just dead-file storage.
It is also important not to confuse this distribution with a general ruling against training. The lawsuit distinguished between the use of legally acquired works, which the judge deemed transformative in the plaintiffs' case, and the obtaining of pirated copies, which left Anthropic exposed and led to the settlement. What must now be monitored is how many corrections are resolved without arbitration, what criteria the special master applies, what the net amount per title is, and whether AI companies start demanding more detailed rights inventories from their suppliers. The useful precedent may not be how much was paid, but how much it costs to prove who should be paid.
Tags
- Anthropic
- Copyright
- AI governance
- Training data
BOLDERROR Daily edition Rubén Campoy