Quick correction to how I'd normally frame one of these: the lawsuit I want to talk about was actually filed December 23, 2025, not this week[1]. I'm writing about it now because the settlement it's reacting to is still actively being administered — notices went out to claimants on September 4, 2026 — and because I think the math behind it is worth walking through slowly.

Start with the settlement these six authors said no to. Anthropic agreed to pay $1.5 billion to resolve a class action over training its models on pirated books, and a judge found that while training on pirated material was unlawful, the trained models themselves could still be used legally[1]. Eligible authors and publishers get roughly $3,000 per work out of that pool, spread across an estimated 500,000 titles.

A group led by John Carreyrou — the journalist who broke the Theranos story and wrote "Bad Blood" — looked at that number and opted out[1]. Opting out of a class action means giving up your guaranteed share of the settlement in exchange for the right to sue on your own. It's a real bet: you're trading a certain $3,000 for the chance at something bigger, with real legal fees and real odds of getting nothing.

What they're actually asking for is $150,000 per title, per defendant, under the Copyright Act's statutory damages provision — and they didn't stop at Anthropic. The new suit, filed in the Northern District of California, names Anthropic, OpenAI, Google, Meta, xAI, and Perplexity, and as far as I can tell it's the first copyright case either xAI or Perplexity has faced over how they trained their models[2]. The complaint accuses all six of pulling books from the same pirate libraries — LibGen, Z-Library, OceanofPDF — to train their systems without permission or payment[2].

The line from their filing that stuck with me: the earlier settlement "seems to serve [the AI companies], not creators[3]." That's a pointed thing to say about a $1.5 billion deal, and I think the underlying argument is actually simple. $3,000 per work, spread across hundreds of thousands of authors, adds up to a huge total number, but it's a number the AI companies can treat as a fixed, absorbable cost of doing business — a line item, not a deterrent. $150,000 per title, times six separate defendants, times however many books each of these authors has published, is a completely different order of magnitude. It's the difference between a settlement that closes the book on a practice and a number big enough to actually change what a company decides is worth doing next time.

I don't know yet whether a court will actually award anything close to $150,000 a title — statutory damages ranges are wide, and juries and judges land all over them. But I think the fact that six authors, one of them a Pulitzer winner with real leverage of his own, looked at a $1.5 billion class settlement and decided it undervalued what happened to their work is itself the story. It's a genuinely open question whether the system built to handle mass copyright harm at scale — the class action, the negotiated settlement, the flat per-work payout — is actually built to make anyone whole, or just built to make the problem go away for a fixed price. These six authors are betting it's the second one, and betting they can do better outside it.