Author Who Beat Anthropic Warns Unlicensed Training Will Drive Writers Out

The piece doesn't read as a victory lap. Graeber argues the settlement hasn't actually resolved anything. If unlicensed training destroys the writer ecosystem, he warns, it eventually leads to an "AI famine" — a shortage of human writing for AI itself to learn from. He's invoking model collapse, the scenario where training data gets contaminated by AI-generated output until quality degrades, and turning it back on the AI companies as a warning from a writer who actually won his case.

$3,000 a book against a $2 trillion valuation

Start with the numbers. The $1.5 billion settlement is the largest copyright payout in U.S. history. But with roughly 480,000 works covered, once legal fees and administrative costs are subtracted, it comes out to about $3,000 per book — and under the default distribution rules, authors split that in half with their publishers. For Graeber, who spent 14 years writing two books, the math doesn't add up to "life-changing money." Legal observers call it a major win, four times the typical damages for illegal downloading, but that's not how it feels to the person who lived it.

The op-ed sets that figure against Anthropic's valuation. According to the Financial Times, investors expect Anthropic's IPO, anticipated for October, to value the company at more than $2 trillion — which would make it the largest public offering ever. From that contrast, Graeber draws his sharpest line: "theft was a smart cost of doing business" for Anthropic. He points to court records showing the company was trying to build an AI that could write like something "an editor would approve," and that CEO Dario Amodei wanted to avoid the "legal/practice/business slog" of buying and licensing books properly.

The court drew a line between training and theft

Here's how the case unfolded. Graeber, along with Andrea Bartz and Kirk Wallace Johnson, sued Anthropic in August 2024. According to court records, Anthropic had downloaded more than 7 million books from pirate sites like Library Genesis (LibGen), and separately bought used books in bulk, stripped their bindings, and scanned them to build a digital library.

In June 2025, Judge William Alsup drew a distinction between those two acts. Training a model on legally purchased books, he ruled, is "quintessentially transformative" fair use — but downloading and storing pirated copies is a separate act of infringement. As the case grew into a class action with statutory damages of up to $150,000 per work, Anthropic faced theoretical liability in the tens of billions of dollars and opted for the $1.5 billion settlement that September. The court gave final approval on July 20 of this year. Judge Alsup had since retired, so Judge Araceli Martinez-Olguin signed off on it.

It's worth noting this ruling isn't a blanket pass for the industry. Because the case ended in settlement, no binding appellate precedent was set, and a different judge in the same court, in the Meta case, rejected outright the argument that AI training is comparable to how a student learns. Last month, major publishers including Hachette and Elsevier sued Google over Gemini's training data, and the New York Times' lawsuit against OpenAI is still ongoing.

When writers disappear, AI starves

The op-ed's core argument isn't a victim's plea — it's a supply-chain warning. Graeber cites a working paper released this month that analyzed roughly 14,000 books sold on Amazon between 2023 and this past March. It found evidence of "market dilution": AI-generated content flooding bookstores, confusing readers, and distorting bestseller lists. With the median income for full-time authors already around $20,000 a year, added sales losses would push full-time writers out first.

That's where the piece pivots. Fewer writers means fewer new human-written books, which leaves AI feeding on its own output — a copy of a copy. Graeber frames this feedback loop as a future both writers and AI companies should want to avoid, and calls for legislation mandating licensing for AI training. He goes further, suggesting AI companies should act like modern-day Medicis, investing directly in human writers and publishers to cultivate the quality writing their own models will need to consume.

Editor's view

What gives this op-ed its force is where it moves the argument. The copyright debate has mostly been framed as creators' rights versus technological progress. Graeber reframes it as a raw-material supply problem for the AI industry. The moment you talk about it in terms of supply chains rather than morality, the audience shifts from regulators to AI company boardrooms. If the model quality behind a $2 trillion valuation depends on human writing as raw material, then protecting the system that reproduces that raw material isn't charity — it's capital expenditure.

The same question arrives in the Korean market on a delay. Because the Alsup ruling recognized training on purchased books as fair use, the cheapest legal path in the U.S. is now bulk-buying and scanning used books rather than negotiating licenses. But the Korean-language corpus is far smaller than English, which gives Korean publishers and content holders relatively more leverage. When a licensing market for Korean-language text — including web novels and webtoons — eventually forms, the real thing to watch won't be the price per book, but the price per retraining cycle. In that sense, $1.5 billion is just the opening bid.