Marcus Kline: I want to start with a very small thing. Last week I was re-reading The Feather Thief — Kirk Wallace Johnson, bird skins, the Natural History Museum heist — and I remember thinking: this is exactly the kind of book that ends up in a LibGen catalog. Nonfiction, backlist, the sort of title a reader finds years after publication.
Ben Okonkwo: Hm — and Johnson is one of the three lead plaintiffs in Bartz v. Anthropic.
Marcus Kline: He is. Along with Andrea Bartz and Charles Graeber. And the settlement that Judge Araceli Martínez-Olguín just approved covers four hundred eighty-two thousand works. One point five billion dollars. Paid by Anthropic — the company that makes Claude.
Ben Okonkwo: Right — and the mechanism is the part I want to make sure we don't gloss over. Anthropic didn't just scrape text from the web. They sourced from Library Genesis and Pirate Library Mirror. LibGen and PiLiMi. Those are shadow libraries — pirate infrastructure, named and known.
Marcus Kline: LibGen operating since 2008. Millions of titles. And Anthropic went there directly.
Ben Okonkwo: Judge William Alsup — he was on the case before he retired — put a number on it. Over seven million pirated books stored in a central library. And the storage is what he called out. Not just training on the works. Storing them. Napster-style, his words.
Marcus Kline: And now plaintiff attorney Justin Nelson is standing in federal court calling this 'the largest known copyright recovery in history.'
Ben Okonkwo: One point five billion dollars. And I genuinely want to sit with what that number does and doesn't tell us, because the per-book math... okay, we'll get there.
Marcus Kline: But that per-book math cuts against the headline, and I think that's worth naming before we move on. The number that actually explains why Anthropic settled isn't one point five billion. It's what Alsup said about *how* the books were obtained.
Ben Okonkwo: Right — and this is the thing I want to be precise about, because the headline version is 'Anthropic trained on pirated books' and that's... not wrong, but it's not where the legal line actually fell. Think of it this way. A chef reads stolen recipe books to learn technique. The cooking itself might be perfectly legal. Walking out of the bookstore without paying was always the crime. That's the split Alsup drew.
Marcus Kline: So the one point five billion is about *where the books came from*. Not what Claude learned.
Ben Okonkwo: Exactly that. Alsup found that training on legally acquired books — potentially fair use. Downloading seven million works from LibGen and Pirate Library Mirror and storing them in a central corpus — infringement. The storage and acquisition is what he denied summary judgment on. Training was the open question they never got to answer at trial.
Marcus Kline: Which means Anthropic paid one point five billion dollars to avoid a jury deciding whether their fair use argument actually held.
Ben Okonkwo: And here's the consequence that I think lands hardest — the settlement requires Anthropic to destroy all the pirated works covered by the ruling. The corpus itself has to go. Now, wait, that's... the court is still worried about what it means for those books to *exist* inside Anthropic's infrastructure. Even post-settlement.
Marcus Kline: The destruction requirement.
Ben Okonkwo: Right — because if training were the only harm, destroying the corpus wouldn't be necessary. Claude already exists. The model's been trained. But the court ordered destruction anyway, which tells you the stored books themselves are the live legal wound. The storage never stopped being the problem.
Marcus Kline: And Google, Meta, OpenAI — they're all still in litigation. Watching this. The question they're now pricing is whether their own acquisition methods survive that same Alsup distinction.
Ben Okonkwo: But that pricing structure for the parallel cases — actually, before we get there — the '$1.5 billion' framing is where I think the circulating take goes wrong. Because $1.5 billion divided across 482,000 books is $3,000 per book. Flat. Regardless of whether your book sold twelve copies or twelve million.
Marcus Kline: That's the number I keep returning to. Three thousand dollars. Picture Andrea Bartz — her debut thriller, commercial sales, real market value — and she's getting the same figure as a backlist academic title nobody's bought since 2011.
Ben Okonkwo: No, that's — right. And the attorneys received over $101 million of the $187.5 million they requested. Martínez-Olguín said the full request was unreasonable. But $101 million is still... that's not a rounding error against $3,000 per book.
Marcus Kline: Some authors and publishers named that directly. Objected. Said the settlement was insufficient. Some opted out entirely and filed separate lawsuits.
Ben Okonkwo: And Martínez-Olguín overruled the objectors. Her logic being — if this goes to trial, zero recovery is a genuine possibility. Fair use was still an open question. So $3,000 is better than nothing.
Marcus Kline: Now, that logic is... I mean, it's pragmatic. But telling an objector they're wrong about risk when the underlying legal questions were never resolved — that's a stretch. The fair use question is still open. We don't actually know what a jury would have found.
Ben Okonkwo: And roughly 45,000 books went entirely unclaimed. Ninety-one percent of the covered works had claimants — which means about 45,000 didn't. Those authors presumably get nothing.
Marcus Kline: Though — and this is the part that signals nobody believes this is actually settled — the agreement preserves authors' rights to future litigation if Anthropic misuses their works. The settling parties kept a door open.
Ben Okonkwo: Which is maybe the most honest signal in the whole document. And the question that follows — whether Anthropic's $1.5 billion looks like a catastrophe or a bargain — that depends entirely on what happens next with Meta, Google, and OpenAI. Their cases don't have a settlement. Yet.
Marcus Kline: And that hanging question — catastrophe or bargain — the answer lives entirely in a courtroom that has nothing to do with Bartz. Because the settlement explicitly doesn't establish binding precedent. Anthropic paid one point five billion dollars and bought... nothing that binds a judge presiding over Meta. Or Google. Or OpenAI.
Ben Okonkwo: Which is — okay, that's the structural irony. The non-precedential posture was probably Anthropic's condition for settling. They didn't want a ruling that hands plaintiffs a loaded weapon in the next case. But the consequence is that they're now standing outside the building while the fire is still burning inside.
Marcus Kline: And every other company watches.
Ben Okonkwo: Right — but watches what, specifically? Here's what I think is the live variable. If a court in the Meta case or the Google litigation rules that training on copyrighted material is fair use — full stop — then Anthropic just paid one point five billion dollars for a legal answer they could have gotten for free at trial. That's the catastrophe scenario.
Marcus Kline: Now consider the inverse. A judge looks at how Meta acquired its training data, draws the same Alsup distinction — acquisition versus training — and finds infringement. Suddenly one point five billion looks... modest. Against what Meta might owe.
Ben Okonkwo: And the settlement amount itself becomes a data point even without precedential force. Meaning — a plaintiff's attorney in the OpenAI case walks into court and says, look, a comparable defendant priced this at one point five billion. That's not binding. But it's not nothing either. Judges aren't operating in a vacuum.
Marcus Kline: So the non-precedential clause insulates Anthropic and simultaneously leaves every other company exposed to... whatever a fresh judge decides. With Anthropic's number floating in the background.
Ben Okonkwo: That's — yeah. The thing to watch is whether Alsup's storage-versus-training distinction gets adopted, rejected, or complicated by the next ruling. That's the load-bearing question the Bartz settlement deliberately left unanswered. And whoever's litigating against Meta or Google right now is essentially running the experiment Anthropic paid not to run.
Marcus Kline: And that experiment... I don't know if it settles anything, actually. Because whoever runs it — whoever gets a verdict first, whether that's the Google case or the Meta litigation — they're going to answer a question Anthropic specifically paid not to answer. And then we'll know. If training is fair use, one point five billion was, and I mean this precisely, the most expensive insurance policy in copyright history. Bought to cover a risk that didn't exist.
Ben Okonkwo: Or, the inverse — a jury finds against fair use, and Anthropic's $1.5 billion turns out to be the number every other company should have settled for two years ago. I genuinely don't know which of those I think is more likely. The data doesn't tell us yet.
Marcus Kline: No. It doesn't. And that's — I keep thinking about Andrea Bartz and Kirk Wallace Johnson and Charles Graeber. They filed the lawsuit that produced the largest known copyright recovery in history, got three thousand dollars each, and the central legal question is still open. The case closed. The question didn't.
Ben Okonkwo: Yeah. That's a strange place to land. Thanks for working through this one.
