Finn Brooks: Clara, hey — okay I have been genuinely stewing on something since last Tuesday and I need you to tell me if I'm being contrarian or if this is actually real.
Clara Bennett: Stewing how? Like "I found a thing" stewing or "I stayed up until two" stewing?
Finn Brooks: The second one, definitely the second one. So the Anthropic settlement — Bartz v. Anthropic — Judge Araceli Martínez-Olguín signed off on it July 21st, called it "meaningful relief," plaintiff attorney Justin Nelson called it the largest known copyright recovery in history, billion and a half dollars, everyone's celebrating — and I cannot shake the feeling that AI companies just got handed a price list, not a punishment.
Clara Bennett: No, I think you're — wait, let me push back on that, because the deletion order is real. Anthropic has to actually destroy the pirated files. That's not a fee structure, that's a consequence.
Finn Brooks: Okay but here's what's making my brain itch — Judge William Alsup ruled in June 2025 that training on *legally acquired* copyrighted books can be fair use. So the message isn't "you can't do this." The message is "you can do this, you just can't use LibGen and PiLiMi to skip the part where you pay for the books."
Clara Bennett: That's the fair use split, and it's actually more precise than you're making it sound — the ruling distinguishes legally acquired copies from pirated ones, and that distinction has real teeth. Seven million pirated books, Finn. Anthropic stored seven million.
Finn Brooks: Wait — seven million? I knew the number was big but — okay that's insane, specifically because Anthropic is the safety-first, alignment-focused company. These are the people writing papers about responsible AI development and someone in their org is pulling files from Library Genesis and going "yeah, this is fine."
Clara Bennett: Right, and that's the contradiction worth sitting with — because the settlement covers 482,000 books at roughly three thousand dollars each, and 91% of eligible authors already claimed their share. That's not symbolic. So the question I keep turning over is: does this create real deterrence, or does it just tell OpenAI and Meta exactly what a clean licensing deal should cost them going forward?
Finn Brooks: That deterrence question is exactly what I think about, and I think the answer is no — and here's why. The music teacher analogy is actually perfect for this. If you legally buy sheet music, you can teach from it. If you photocopy it from a bootleg site, that's the crime. Alsup didn't say teaching is illegal. He said the bootleg copy disqualifies you. That's the whole ruling.
Clara Bennett: That's exactly the click. The lesson isn't illegal. The illegal copy is.
Finn Brooks: Which means Anthropic isn't paying $1.5 billion because building Claude on books was wrong. They're paying because someone chose LibGen over a licensing desk. That's — I mean, that's a procurement failure dressed up as a moral reckoning.
Clara Bennett: I'd push back on 'just' procurement — the deletion order is structural. They have to destroy the files, not just pay for them. That's not a fee.
Finn Brooks: Hold on. Does deleting the files un-train Claude?
Clara Bennett: No — and that's the uncomfortable part. The weights are already baked. So the deletion is real as a legal remedy but, in practice, the model trained on those 7 million books still exists. Now — only 350 authors opted out. Three hundred and fifty, out of a class covering 482,000 books. I want to sit with that number because it means either the $3,000 felt genuinely fair, or authors looked at the litigation math and decided they had no leverage.
Finn Brooks: Wait — 350. That's like, what, 0.07% of the class? That's not a negotiation, that's a surrender.
Clara Bennett: Or it's clarity. Andrea Bartz, Charles Graeber, Kirk Wallace Johnson filed this in August 2024, Alsup drew the piracy line in June 2025, and by the time Martínez-Olguín signed off July 21st, 91% of eligible authors had claimed. The legal path was mapped and the payout was real. That's not surrender — that's authors reading the room accurately.
Finn Brooks: Reading the room accurately — okay, but reading the room accurately and having no real options are the same outcome. Like, a group of authors actually tried to block the settlement in May. They argued lawyer fees were too high, payouts too low. A handful tried to opt out past the deadline to pursue separate lawsuits. And they still only got to 350 total. That's not authors making a rational choice, that's authors hitting a wall.
Clara Bennett: That's a real point — but Judge Alsup rejected an earlier version of the settlement before the case was reassigned to Martínez-Olguín. The court applied real scrutiny here. It wasn't just rubber-stamped.
Finn Brooks: Wait — Alsup rejected it first?
Clara Bennett: Rejected an earlier version, yes. Which means the $3,000 figure — four times the statutory minimum for copyright infringement, by the way — went through actual judicial pressure before it landed. That's not nothing.
Finn Brooks: Four times minimum — okay that sounds big until you actually, no wait — sit with the science fiction novelist for a second. She's got a backlist, sold steadily for a decade, files the claim form, gets $3,000. Real money. And then she opens Claude and it's generating prose rhythms, plot structures, the specific cadences she spent years building. The $3,000 is a one-time payment. The extraction — that just keeps running.
Clara Bennett: The ongoing use isn't priced in. That's the actual gap, and the settlement doesn't close it.
Finn Brooks: Right — and that's what makes $3,000 feel like a retroactive license, not restitution. Anthropic didn't get punished for what they built. They paid a one-time invoice for how they sourced the raw material.
Clara Bennett: I mean — I don't entirely disagree, but the framing shifts completely when you realize Alsup's fair use logic may end up helping OpenAI, Meta, Google more than it ever hurt Anthropic. That's the part I want to get into, because the precedent problem here is actually worse than the settlement math.
Finn Brooks: No, I don't buy that it's worse — but yeah, that is the next terrifying question.
Clara Bennett: But that 'worse' framing I threw out — I want to walk that back one step, actually. Not the substance, just the scale. Because Alsup's ruling isn't binding appellate precedent. It's a summary judgment from one district court judge in the Northern District of California. OpenAI's cases, Meta's cases — different judges, different fact patterns, different training data sourcing. No court is obligated to follow it.
Finn Brooks: Wait — so you're saying other courts can just... ignore it?
Clara Bennett: Ignore is strong, but — yes, in practice, they're not bound by it. It's persuasive authority at best. And the cases pending against OpenAI and Meta involve different sourcing, different alleged conduct. Alsup's logic doesn't automatically transfer.
Finn Brooks: Okay, no — I love that you're conceding that, because that's exactly what I've been arguing. Justin Nelson called it the largest copyright recovery in history, which, fine, the number is real. But 'largest recovery' and 'binding legal constraint on the industry' are completely different things. No rule compels OpenAI to license anything because Anthropic settled.
Clara Bennett: That's the narrow point and you're right on it. I'll give you that fully. But — and this is where I'm holding ground — the incentive shift doesn't require a legal mandate. Every AI company's board just watched Anthropic absorb a $1.5 billion settlement. That's a board-level number regardless of whether a judge in a different circuit is obligated to cite Alsup.
Finn Brooks: Hmm. I want to believe that but — the flip reading is that those boards look at Alsup's fair use logic and go 'license your data upfront, claim the same defense, you're clean.' That's not a deterrent. That's a roadmap.
Clara Bennett: It might be both simultaneously, and I think that's the uncomfortable landing here. The precedent that shakes out could genuinely be: license your data, training is fair use, you're untouchable. Which helps OpenAI and Meta more than it ever constrained Anthropic.
Finn Brooks: So the $1.5 billion is real, the deletion order is real, and it still might be the case that the net effect on the industry is — AI companies got handed the legal architecture for doing this cleanly. That is a wild outcome.
Clara Bennett: And that's — I mean, that's actually where I want to land, because you opened this whole thing saying you couldn't shake the feeling that AI companies got handed a price list. And I kept pushing back. But Andrea Bartz, Charles Graeber, Kirk Wallace Johnson filed this case, they won the largest copyright class action in U.S. history, and the ruling underneath the win might be the thing that lets OpenAI and Meta license upfront and claim exactly the fair use defense Anthropic lost. That assumes their sourcing is clean, though — and it's not obvious it is.
Finn Brooks: No, it's not. But that's almost beside the point now. Anthropic paid $1.5 billion for downloading from LibGen and PiLiMi. The industry now knows what the shortcut costs. That's — I don't know, it's a strange thing to call a landmark.
Clara Bennett: The authors won the settlement. The AI companies may have won the law. That's probably where we have to leave it.
Finn Brooks: Yeah. Uncomfortable place to stop.
Clara Bennett: The good ones usually are. Thanks for the two a.m. stewing — genuinely.
