A Lossy Library

One and a half billion dollars is the number sold to you as a win for authors. It's the value of the final settlement in Bartz v. Anthropic, approved on 20 July 2026: about three thousand dollars a book across half a million works, for training Claude on copies pulled from LibGen and Pirate Library Mirror. The press release reads like justice. The documents unsealed last week, more than four thousand pages, read like something else.

They describe Project Panama: the operation in which the same company bought, scanned, and then physically destroyed between five hundred thousand and two million books in six months. A hydraulic cutter to strip the bindings, production scanners, a recycling firm to haul off the paper. An internal document puts it plainly: "Project Panama is our effort to destructively scan all the books in the world," and adds that a codename is used "because we don't want it to be known that we are working on this." The codename is the confession. You disguise only what won't survive daylight.

The same case that put a price on theft blessed the destruction. Downloading for free from pirate sites cost a billion and a half; buying the volume at retail, guillotining it, and digesting it had already been ruled transformative fair use in 2025. Read the two decisions together and you see the incentive they write as a pair: the clean way to extract a book's value without infringing anything is to own a lawful copy and make the carrier disappear. Copyright protects the copy, not the object. So the object becomes a cost to dispose of.

Anyone who has worked with files knows this move. It's the difference between a backup and a lossy re-encode you throw the master away from: you keep the compressed version and destroy the original. Nobody charges you with theft; the negative simply no longer exists. The text, formally, survives. It lives in the model's weights, queryable on command. What doesn't come back is the edition with its annotated margin, its provenance, its typographic variant, the object that let a third party reread and collate. What goes is the ability to check the source against anything but another run of the model.

Digitizing is not the crime. Google Books scanned millions of volumes and left them on the shelf; the Internet Archive exists to multiply copies, not to thin them to one. The line Panama crosses runs between copying and replacing: here the scan doesn't sit beside the original, it stands in for it and disposes of it. A preservation method run backwards, subtracting redundancy instead of adding it.

That's where the project betrays the promise these same companies keep repeating. A machine that opens a book should extend how far a reader gets: linking it, translating it, making it answerable to people who will never walk into that library. A key is for opening. Panama uses it to wall off: it ingests the thought and shatters the support that kept the thought contestable. The machine doesn't decide this; the incentive the ruling sets around it does. A tool stops being neutral the moment its outcome touches the hand that holds it.

For anyone working with these models daily, the stake is concrete, not nostalgic. You trust an output because you can, in principle, trace the source and verify it. When the source is an original sent to the pulper, the chain of verification breaks at a point you can't see: you can only compare the model's answer with another of the model's answers. A lossy library doesn't take away the books you had; it takes away your right of appeal against their compression. And a professional who loses the appeal on the source hasn't lost an archival detail: they've lost the way to say "this is wrong" with something in hand.