Onpode
Cover art for Australian rare-book sellers say AI companies are destroying 'horrific' quantities of irreplaceable titles to feed training pipelines

Australian rare-book sellers say AI companies are destroying 'horrific' quantities of irreplaceable titles to feed training pipelines

August 2, 2026 · 10 min

Tess Hollis & Felix Ortiz

AI companies, including Anthropic via the secret 'Project Panama,' bought millions of physical books, had contractor Datamation scan them at 80–120 pages per minute, then shredded them for training data. A $1.5 billion copyright settlement followed, but no legal framework addresses whether any destroyed volume was an irreplaceable last surviving copy.

Rare booksellers in Australia and Europe are raising urgent concerns that AI companies are systematically purchasing, destructively scanning, and then pulping physical books — including rare, specialist, and out-of-print titles — to feed large language model training pipelines.

0:0010:01
Get the next episode on Artificial Intelligence

Follow it free — new episodes land in your feed.

Or make your own — any topic, in minutes

More Onpode episodes on Artificial Intelligence

About this episode

Somewhere in 2026, a rare book scout at a European estate sale found a shelf of 1970s academic monographs — the kind she'd normally pass on. Too niche, no collector market. But there was a buyer. And those books weren't going to a shelf. This episode starts with a $1.5 billion settlement and a fair use ruling that coexist without canceling each other out, and works backward to what that combination actually means on the ground. Project Panama — Anthropic's internal book-scanning operation, run through a contractor called Datamation and hidden behind a chain of NDAs — processed pallets of physical books: spines cut, pages scanned at up to 120 per minute, remainder shredded. The episode traces why AI labs shifted from pirated ebooks to physical print, why 1970s obscure nonfiction became premium sourcing material, and what it means that the market signal for cultural preservation and the market signal for data extraction are now identical in real time — visible to no one in the supply chain until the books are already gone. Australian rare book dealers have used the word 'horrific.' The legal frameworks — fair use, copyright settlement — have no way to register whether a destroyed volume was the last surviving copy of anything. This episode doesn't resolve that. It sits with it honestly, which feels like the right response to something that genuinely doesn't have an answer yet.

Frequently asked

Are AI companies really destroying physical books after scanning them?

Yes. Anthropic's Project Panama used contractor Datamation to slice the spines off physical books, photograph every page at 80–120 pages per minute, then shred the physical copies. The operation processed pallets of 800–1,200 volumes at a time and was confirmed through unsealed court filings from Bartz v. Anthropic.

What was Anthropic's Project Panama?

Project Panama was Anthropic's covert book-acquisition and scanning operation. Anthropic outsourced purchasing and physical processing to vendors — all bound by NDAs — to buy millions of printed books, scan them for AI training data, and destroy the originals. The contractor Datamation ran industrial scanning equipment cutting spines and shredding copies.

What was the outcome of Bartz v. Anthropic?

In Bartz v. Anthropic, Judge Alsup ruled in 2025 that buying and destroying physical books for AI training was 'clearly transformative' under fair use, a legal win for Anthropic. Despite that ruling, Judge Martínez-Olguín approved a $1.5 billion settlement to the same plaintiffs on July 20, 2026, covering separate copyright harm claims.

Why are AI companies specifically targeting 1970s out-of-print books?

AI labs seek pre-2022 physical books because text published after that date risks contamination by AI-generated content already online. Obscure 1970s academic nonfiction — rarely digitized, absent from Google Books or university archives — offers clean human-written training data. Buyer Zoom Books placed large orders with European antiquarian sellers for exactly this category starting in early 2026.

Is it legal for AI companies to destroy rare books after scanning them?

Under U.S. copyright law, Judge Alsup's 2025 Bartz v. Anthropic ruling found the practice 'clearly transformative' and therefore fair use. However, the ruling addresses copyright only — not whether a destroyed book was a unique surviving copy. No legal framework currently requires provenance checks or protects irreplaceable physical artifacts from being pulped for AI training.

Grounded in 12 sources
‘More than just objects’: Australian booksellers raise alarm over ‘horrific’ destruction of rare titles to feed AI · theguardian.com
Anthropic destroyed millions of print books to build its AI models · arstechnica.com
A Case Study of Authors Guild v. OpenAI and Microsoft · 4ipcouncil.com
AI firms' mass purchase and destruction of books for model training fuels copyright debate · aa.com.tr
[PDF] Authors-Guild-Comments-AI-and-Copyright-October-30-2023.pdf · authorsguild.org
Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement - The Authors Guild · authorsguild.org
AI companies are buying and destroying old books for training data – Block Stream Media · blockstreammedia.com
‘More than just objects’: Australian book sellers raise alarm over ‘horrific’ destruction of rare titles to feed AI | Collector: Breaking News, World News, Trending Stories · collector.com.tr
Bartz-v.-Anthropic-Order.pdf · copyrightalliance.org
AI firms are shredding physical books because copyright law is quietly rewarding them · cryptoslate.com
Anna's Archive - Wikipedia · en.wikipedia.org
AI Companies Are Buying Antique Books, Ingesting Their Contents to Train Models, and Then Destroying Them at Incredible Scale, Even If Almost No Copies Remain · futurism.com
Read transcript

Felix Ortiz: Tess, I've been sitting with this story for days and I still can't — it keeps shifting under me, you know?

Tess Hollis: Same. I read the settlement number and had to go back and check it. $1.5 billion.

Felix Ortiz: From a company that won. That's the part — Anthropic won on fair use. Judge Alsup, 2025, Bartz v. Anthropic: buying and destroying physical books for AI training is 'clearly transformative.' Win. And then July 20, 2026, Judge Martínez-Olguín approves a $1.5 billion settlement to the same plaintiffs for the same books.

Tess Hollis: Okay, so — that's exactly where I want to start. Because you know this territory better than I do: what does it mean that the legal victory and the billion-dollar payout coexist? Is the law actually working here, or is Anthropic just making noise go away?

Felix Ortiz: Honestly? I think it's mostly noise-erasure. The fair use ruling gave Anthropic a legal shield, but the unsealed court filings — from Bartz — showed what Project Panama actually looked like on the ground. Pallets of books. A contractor named Datamation doing industrial scanning, 80 to 120 pages a minute, spines cut off. The Authors Guild was not going to let that image just sit in public record while Anthropic pointed at an Alsup ruling.

Tess Hollis: So the settlement is — what, reputation management at a billion-five?

Felix Ortiz: Partly, yeah — but also, right, the fair use ruling doesn't settle the damages question automatically. Even if destruction is transformative, you can still owe for infringement on other grounds. The two tracks aren't actually incompatible, they're just — they're talking about different things. Alsup says the act is transformative. The settlement says the compensation was inadequate. And both can be true.

Tess Hollis: That's the part I'm not sure I buy — because if both are true, we've named the thing and still solved nothing.

Felix Ortiz: Right — but the part that makes it concrete is what the operation actually looked like. Picture a warehouse receiving pallets — 800 to 1,200 books per pallet — and a contractor, Datamation, running them through machines that slice the spine off, photograph every page at 80 to 120 pages per minute, then shred what's left. That's Project Panama. Not a legal abstraction. A factory floor.

Tess Hollis: Wait — Anthropic contracted that out? They didn't do the scanning in-house?

Felix Ortiz: Yeah, and that's actually — no, wait, that's exactly the clever part. Every vendor in that chain signed an NDA. So the liability is dispersed and the visibility goes to zero. The only reason we know the name Datamation is because the Bartz v. Anthropic filings got unsealed.

Tess Hollis: Okay, so that's the buried thing. The operation was invisible by design — not just legally insulated but structurally hidden.

Felix Ortiz: And the 2024 pivot makes that sharper — Anthropic moved away from pirated ebooks, sourced through platforms tied to Anna's Archive, and shifted to buying physical print copies in the millions. Cleaner legally. But also far more opaque because physical supply chains don't leave a download log.

Tess Hollis: Which is what makes the ISBNdb thing so strange to me. Because that wasn't hidden — that was a marketing page. 404 Media found it. 'Up to one million physical books per order, strict NDA on every engagement.' That's a sales pitch.

Felix Ortiz: Right, and then it disappears around July 28, 2026. Just — gone. Someone decided the pitch had become a liability.

Tess Hollis: So if the infrastructure was that visible — a literal public marketing page — why did it take unsealed court filings to make any of this public?

Felix Ortiz: Because nobody connected the dots until they had to. The ISBNdb page and the Bartz filings were both sitting there — I mean, 404 Media caught the page, the Washington Post ran the fuller investigation in January 2026 — but it took legal compulsion to unseal the part that named the contractor and described the pulping. The NDA structure meant every individual vendor stayed quiet. The whole thing is visible in hindsight. In real time it looked like a lot of separate boring purchases.

Tess Hollis: But here's what the settlement and the fair use ruling both completely miss — neither one says a word about the physical object. Tim White, Books for Cooks, at the Melbourne Rare Book Fair in Wilson Hall — August 2026 — calls it 'horrific.' And what he's naming isn't the copyright. He's saying the artifact is gone.

Felix Ortiz: Right — and 'more than just objects,' that's the exact phrase. He's not talking about the text.

Tess Hollis: So what does Judge Alsup's ruling actually say about whether a destroyed book was the last surviving copy?

Felix Ortiz: Nothing. It says zero about provenance. Whether Datamation pulped a book that existed in forty libraries or — yeah, one surviving copy in a Melbourne shop — legally invisible. The fair use analysis doesn't have a field for that. And the $1.5 billion settlement from Martínez-Olguín's July approval — that's all going to authors and rightsholders. Copyright harm. The physical object being irreplaceable? No remedy exists for that at all.

Tess Hollis: No, I don't buy that this is just a theoretical gap.

Felix Ortiz: Okay — but wait, I have to be fair here. Nobody has confirmed a specific last-surviving rare book was actually destroyed. We know millions of volumes moved through Project Panama. We know Zoom Books has been placing huge orders with European antiquarian booksellers since early 2026 — nonfiction, academic, 1970s titles — and yeah, the suspected destination is US AI labs. But the jump from 'millions of books' to 'irreplaceable cultural loss' — that gap is still real.

Tess Hollis: The scale IS the evidence. If you're moving millions of volumes through a pulping operation and the law has no way to flag provenance, you don't need a confirmed case. The risk isn't hypothetical at that number.

Felix Ortiz: That's — yeah, okay, that's actually the better frame. And there's a reason Zoom Books is targeting 1970s academic nonfiction specifically — which is the part that, honestly, makes this worse in a direction we haven't gotten to yet.

Tess Hollis: What do you mean — why that window?

Felix Ortiz: Yeah so — okay, this is the mechanism. AI labs need pre-2022 physical books specifically because anything after that risks being contaminated by AI-generated text already circulating online. You want clean human writing? You go physical, you go old, you go obscure. Which means — wait, no, think about what that does to the market signal. The most valuable training material is suddenly 1970s agricultural chemistry monographs. Stuff nobody's digitized. Stuff that's been sitting in provincial estate sales collecting dust.

Tess Hollis: Hold on — that's the inversion nobody's talking about.

Felix Ortiz: Exactly. The thing that makes a book worth saving is now the thing that makes it worth destroying. Picture a rare book scout at an estate sale in rural France, early 2026 — she finds a shelf of 1970s agricultural chemistry monographs. Normally she passes. Too niche, no collector market. But there's a buyer now. Zoom Books. Big order. And those books are not going to a shelf anywhere after the sale.

Tess Hollis: That's the inversion nobody is talking about. The demand signal that raises their price is the same signal that ends their physical existence.

Felix Ortiz: And ISBNdb had explicitly categorized rare and out-of-print titles as a premium sourcing category for AI developers. That wasn't incidental — that was the pitch. Until July 28.

Tess Hollis: Which means Zoom Books targeting 1970s nonfiction — that's not random buying. That's a deliberate demand category.

Felix Ortiz: Right. And nonfiction academic from that decade is almost certainly not digitized anywhere. No Google Books scan, no university archive, no backup. So the scout in France makes a sale she's never made before — and the books don't come back from it.

Tess Hollis: So cultural stewardship and raw material sourcing produce identical market behavior — but completely opposite outcomes for the object.

Felix Ortiz: Yeah — being valued as data and being preserved are not the same thing, and the market has no way to tell them apart in real time. The scout doesn't know. The seller doesn't know. Zoom Books isn't labeling the order 'for pulping.' The price goes up, the book disappears, and the mechanism is invisible.

Tess Hollis: And that's where I keep getting stuck. Because Judge Alsup already ruled — 'clearly transformative,' same logic as Google Books. But Google Books kept the originals. Anthropic pulped them. The legal reasoning is identical and the physical outcome is completely different. So if we find out later that one of those millions of volumes was the last surviving copy of something — what does the law owe? The settlement covered copyright harm. Tim White said these things are 'more than just objects.' There's no court, no framework, that sees that second kind of loss at all.

Felix Ortiz: Yeah. I don't — I actually don't have an answer to that. And I think that's the honest place to land.

Tess Hollis: Neither do I. Which is — I mean, that's probably the right feeling to leave with.

Australian rare-book sellers say AI companies are destroying 'horrific' quantities of irreplaceable titles to feed training pipelines · Onpode