The distribution of Anthropic's $1.5 billion runs into publishers and agents claiming more than their due

By Zendoric · September 7, 2026.

This week, several US authors received an uncomfortable notice: someone else had claimed their share of the $1.5 billion settlement Anthropic reached for training its models on pirated books. That someone, in many cases, was their own publisher or a literary agency that no longer held any rights to the work.

The settlement, granted final approval in July, closes a class action Anthropic resolved after a court ruling that established a key distinction: training an AI model on copyrighted material falls under "fair use" (legitimate use the law permits without asking permission), but obtaining that material through pirated copies does not. Under those terms, the authors of nearly 500,000 titles will be paid $3,000 for each pirated work. If the book is still published by a traditional publisher, the payment is split 50-50 between author and publisher; if the author self-published it, or if the publisher let the book go out of print and reverted the rights, 100% goes to the author.

The problem has surfaced in the fine print of that split. Crime novelist April Henry said HarperCollins claimed as its own a book whose rights had been reverted to her at least 17 years ago — on the same day, by her account, that a credit alert informed her the publisher was listed as her employer, something that never happened. Victoria Strauss, of the well-known industry blog Writers Beware, wrote that she has spent days receiving complaints from authors that fall into two patterns: publishers claiming books to which they no longer hold any rights, and publishers claiming 100% of the payment when only 50% is theirs.

Added to those claims are ones from several literary agencies, something that surprised even close industry watchers: as Strauss notes, agents do not hold rights to the books they sell, so they should have nothing to collect. Author Courtney Milan (the pen name of Heidi Bond, a former judicial clerk and former law professor) summed it up bluntly on social media: "some agents are trying to claim percentages of the Anthropic settlement, and I don't think they should be doing that at all".

Neither Strauss nor Authors Guild chief executive Mary Rasenberger attributes this to bad faith. Rasenberger told the New York Times that she does not see it as "an assault by the publishers" nor does she believe they are trying to "deliberately cheat" any author; for her, it is the predictable result of poor record-keeping and a confusing claims process. Some publishers, according to Strauss, have already asked Anthropic to correct their own erroneous claims. But Strauss herself qualifies her caution: she says she has seen only "a small crack in an enormous wall", and even so, the volume of complaints and the repetition of the same errors suggest to her something "broader and more systemic" than isolated glitches.

There is also a technical detail that will decide many of these cases: for an author to claim 100% of the payment, the reversion of rights must have occurred before August 10, 2022, the "download date" the settlement sets as its reference point. Milan and the Authors Guild have already begun explaining publicly how to dispute an incorrect payment allocation.

Our reading is that this episode matters beyond the publishing world: it is the first real large-scale test of how the money gets divided when generative AI finally pays for what it used without permission. And the test is going only so-so. When a class action puts billions on the table, the advantage lies not only with whoever holds the legitimate right, but with whoever has the administrative apparatus — legal departments, contract databases, dedicated staff — to file the claim first and most forcefully. The individual author, who in many cases does not even know precisely when they regained the rights to a book published two decades ago, starts at a disadvantage against a publisher or an agency that can automate hundreds of claims at once.

This connects with something we have already pointed out in other contexts, such as facial image licensing markets in Asia: when AI turns a previously intangible asset — a face, a pirated book — into something that generates a concrete payment, the real dispute is not over whether that payment should exist, but over who controls the verification infrastructure that decides whom it belongs to. Anthropic did not design a granular, auditable copyright registry; it inherited the documentary chaos of a publishing industry that for decades did not need to keep those records precisely, because no court had ever before put a per-unit price on each pirated book.

In the short term, this is a boring but real governance problem: opaque claims processes, tight deadlines and the burden of proof falling on the weakest party. It is exactly the kind of transition friction that should not be downplayed, because at the scale of half a million titles, a "systemic" error of 1% means thousands of authors being underpaid. But it is also a useful precedent: this settlement will not be the last. Similar litigation and negotiations are already under way among news publishers, image agencies and AI developers, and all of them will need, sooner or later, a distribution mechanism. If the industry learns from this first stumble — clearer rights records, independent audits of claims, reasonable dispute windows — the next settlements could become the template for how to fairly compensate whoever created the content a model was trained on, instead of another arena where whoever has more lawyers wins. That, ultimately, is the question that really matters as AI distributes more and more value generated by human work: not just how much is paid, but who actually receives that payment.