An artificial intelligence company bought used books by the tens of thousands, cut away their spines, scanned every page, and sent the paper to be recycled. The words survived. What may not have survived is harder to name, and it concerns who we become when memory-keeping passes from many hands into one.
In early 2024, under an internal effort the artificial intelligence company Anthropic called Project Panama, a hydraulic blade began coming down on the spines of books, and it kept coming down until millions of them had passed beneath it. Project Panama was defined in the company’s own planning document in a single unhedged line: “Project Panama is our effort to destructively scan all the books in the world.” The loosened pages were fed through high-speed scanners, the text became machine-readable files for training the company’s Claude models, and a recycling company was scheduled to collect what remained. The same document noted that the company did not want the project to become known. It became known anyway, when a copyright lawsuit forced more than 4,000 pages of internal records into the open.
Much of the public reaction reached for the oldest image we have for this kind of loss: the burning of books. That is not quite what happened, and the difference matters. Nothing was burned, and nothing was banned. What was lost is subtler than paper, and it says something uncomfortable about how a civilization remembers, and how it forgets.
What Project Panama actually did
Project Panama was not Anthropic’s first answer to where to find books. According to Judge William Alsup’s June 2025 ruling in Bartz v. Anthropic, the company had earlier assembled a library that included millions of pirated books downloaded from shadow libraries, such as Library Genesis. As the legal risk of that approach became clear, it changed course. In February 2024, it hired Tom Turvey, a former Google executive who had helped create Google Books two decades earlier, with a mandate to find a way to obtain all the books in the world.
The filings show the company weighed buying from libraries and from landmark used bookstores, such as New York’s Strand, before settling on bulk purchases, often tens of thousands of books at a time, from used-book retailers, including Better World Books and the U.K.-based World of Books. One scanning vendor described the job as converting between 500,000 and two million books over six months. That is thousands of books every working day.
Why books at all, when the Internet seems endless? Because the Internet is not where the best writing lives. Company executives wanted books that could teach the model to write well, rather than imitate the hurried, low-grade prose that fills much of the web. Books written before 2022, before generative AI began seeding the Internet with its own output, were considered especially valuable.
The process itself was industrial. A hydraulic cutting machine sliced the bindings away, production-grade scanners captured every page, and workers picked up the stripped books for recycling. Alsup’s ruling summarized the workflow in plain judicial language: the books were stripped from their bindings, cut to size and scanned, “discarding the paper originals.”
Here the record needs careful reading, because it corrects the most common assumption about the story. The court documents give no indication that rare manuscripts or one-of-a-kind volumes went under the blade. These were mostly used copies, bought in bulk from mainstream sellers, of books that still sit on library shelves and in private homes worldwide. The works were not erased. The copies were. That distinction does not make the loss trivial. It relocates it.
Why cut the books instead of turning the pages?
Two reasons, and neither requires assuming malice. One is physics. The other is law. A bound book resists being scanned. Its pages curve toward the spine; they must be turned one at a time, and every turn costs seconds. Cut the spine away, and the book becomes a stack of loose sheets that a feeder scanner can swallow as easily as office paper. Across millions of volumes, that difference can mean the difference between a project measured in months and one measured in years. Non-destructive scanning at enormous scale has been done, however. Google has digitized more than 40 million books in over 500 languages without destroying them, while partner libraries kept the originals and the public gained a searchable index.
The second reason is more revealing. When Alsup ruled on the scanning, he treated it as a change of format rather than the making of a new copy. Anthropic had bought a print copy, replaced it with a digital one for internal use, and removed the print original. Because no extra copy entered the world and the file stayed inside the company, the judge found the conversion to be fair use. He found training the models on the books to be “exceedingly transformative,” and therefore fair use, while the earlier use of pirated copies was not. Anthropic later agreed to pay $1.5 billion to settle the piracy claims.
Read that reasoning the way an engineer reads a specification, and the incentive becomes visible. Under the ruling’s logic, destroying the physical book was not an unfortunate side effect of the copy. It was part of what made the copy defensible. Legal commentators quickly pointed out that the decision effectively rewards AI companies for destroying the print books they buy once those books have been scanned.
This pattern underlies most large outcomes in technology. Systems do not need villains to produce troubling results. They need incentives, and they follow them faithfully. A blade, a scanner, and a court opinion together formed an architecture, and that architecture favored a one-way conversion: from paper that anyone could hold, lend, or inherit, to files that only one company could read. Systems do not need villains to produce troubling results. They need incentives, and they follow them faithfully.
In fairness, books have always been mortal. Publishers have pulped unsold stock for generations, and libraries routinely discard worn or little-used copies to make room. Paper was never eternal. What sets Project Panama apart is not that books were destroyed, but the scale, the secrecy, and, above all, where the text went afterward.
The end goal is a library that never forgets
Alsup’s ruling records the ambition plainly. Anthropic wanted to build a central library of “all the books in the world” and to retain it “forever,” drawing sets and subsets from it to train each new generation of its models. Forever. The word deserves a pause, because it shows that Project Panama’s end goal is, in its own way, memory.
There is a quiet irony here. Underneath the whole debate sits a simple question: what knowledge could a machine learn from, other than ours? There is no other source. Every sentence a language model can produce, it learned from people who wrote, argued, mourned, prayed, joked, and explained things to one another across centuries. The machine’s hunger for books is, in effect, a confession that everything it knows comes from us.
And the most coveted human writing is the older writing. Physicists who build extremely sensitive instruments prize steel made before the first atomic tests, because it carries no trace of the fallout that has seeped into newer metal; some of it is salvaged from sunken ships. Pre-2022 books have become the low-background steel of language: human thought recorded before machines began writing back into the record. The AI industry is mining the past because the present is filling up with its own reflections. So the goal is not destruction for its own sake. It is a machine that remembers how humanity wrote and reasoned. The real question is narrower and harder: whose memory does it become, and who gets to visit it?
A book is a token of memory, not a container of data
Anyone who builds AI agent systems learns a hard lesson early. An agent works inside a context window, a limited span of attention. When that window closes, everything the agent understood but did not write down is gone. So we teach our systems to externalize what they know into artifacts: specifications, decisions, test results, a record of what was tried and why. The next agent reads those artifacts and carries the work forward. An agent that loses its artifacts does not merely forget. It re-decides settled questions, contradicts its earlier choices, and repeats failures already paid for.
Human civilization runs on the same principle, only slower. A generation is a context window. A lifetime is a context window. Books, monuments, graves, festivals, and family heirlooms are the artifacts we leave so that the next window can open with something already in it. Human beings are forgetful by nature, which is precisely why we raise stones and print pages. They are physical tokens that hold memory beyond any one mind.
But an artifact carries more than content. In a well-built system, it carries lineage, the record of where it came from and which decisions produced it. And it stays open to whoever comes next. A trained model fails both tests. The text of a million books dissolves into billions of numerical weights. The model may tell you, more or less, what a book argued, but it cannot reliably tell you which book, which author, or which edition. The knowledge survives. The lineage does not.
The physical copy held something else, too. A used book is rarely just a text. It may carry a name written inside the cover, a dedication from a grandmother, a pencil line under the sentence that once changed a stranger’s mind, a train ticket left as a bookmark decades ago. That copy witnessed its readers, and in the used-book trade it would have passed on to the next one.
None of the works scanned under Project Panama vanished from the world. But each individual copy left public circulation for good, and its text entered an archive that no reader, librarian or author can enter. That is the true shape of the loss. Memory did not burn. It moved from many hands into one. Memory did not burn. It moved from many hands into one.
Who we become when we forget
Why should this matter to anyone who has never set foot in a used bookstore? Because identity is built on memory, and the chain runs deeper than most of us notice. Our experiences result from our actions. Our actions are guided by intention. Identity guides intention, and intention guides our actions. Identity itself forms through our dealings with other people against a shared backdrop of history: the stories, the names, the debts of gratitude, the mistakes no one wants to repeat. Remove that backdrop and identity loses its frame of reference. Intention loses its compass. Action becomes reaction.
We can already see this happening, without any machine’s help. In many families, each generation inherits fewer stories than the last. Grandparents’ recipes go unwritten, great-grandparents’ names go unspoken, and the roots that once tied people to a shared cultural identity wither, not because anyone cut them, but because no one watered them. A Chinese proverb, 饮水思源 (yǐn shuǐ sī yuán), says it in four characters: when you drink the water, remember its source. A people that forgets the source still drinks. It simply no longer knows what it is drinking, or whom to thank.
The ancient puzzle of the Ship of Theseus asks whether a ship whose every plank has been replaced is still the same ship. One answer is that it remains the same ship for as long as the shipwrights remember its form. The planks are material. The form is spiritual. A culture renews its planks with every generation, new people, new tools, new words, and that renewal is healthy. What keeps it the same culture is that its form, its values, its sense of what is noble and what is shameful, is handed down intact beneath the new timber. Books and monuments protect that form by keeping it visible. When the tokens disappear, and the teaching stops, the ship is rebuilt from a memory that no longer exists, and it becomes a different ship without anyone noticing the moment it happened.
The keepers who taught the past
Many traditional societies did not leave this to chance. They appointed keepers. Among the Zulu, the imbongi, or praise poet, carried the izibongo, poems that recorded the deeds and lineages of chiefs and ancestors. Scholars of Zulu oral literature compare the imbongi to the historian of a literate culture; he was at once historian, counselor, and social critic, reciting the past before the living and, at times, reminding rulers of what their forebears would have expected of them. Beside him stood the izangoma, diviners understood to be called to their work by the amadlozi, the ancestors, whose name comes from a word meaning to care for and keep watch over. Between them, memory was never simply stored. It was performed, taught, and spoken again in each generation, before people who could answer back.
Chinese history offers a sharper lesson about custody. In 213 BCE, on the advice of his chancellor Li Si, the First Emperor of Qin ordered privately held copies of the Classic of Poetry, the Classic of History and the writings of the Hundred Schools to be surrendered and burned, sparing works on medicine, divination and agriculture. According to the Records of the Grand Historian, the court’s own scholars were permitted to keep their copies. Then, in 206 BCE, the rebel general Xiang Yu set fire to the palaces of Xianyang, and much of what the state had kept was lost. Historians still debate how complete the destruction was, but the story’s structure has endured: the single custodian proved the most fragile.
What survived came back through people. The scholar Fu Sheng is said to have hidden his copy of the Classic of History inside a wall. When the Qin dynasty fell, he recovered 29 chapters and taught them to a new generation, and the text lives on because one man kept it and then gave it back.
To be clear, Project Panama is not the Qin edict. No book was banned, no reader was punished, and every work it scanned still exists elsewhere. The parallel is not one of intent. It is one of architecture: what happens to a culture when the most complete copy of its memory comes to rest in a single house, and that house does not open its doors.
Keeping is not remembering
The strongest defense of Project Panama deserves a fair hearing. AI models do give something back. Every day, millions of people ask them questions and receive answers shaped, in part, by the books they learned from. In one sense, the model is a new kind of imbongi, reciting humanity’s accumulated knowledge to anyone who asks, in any language, at any hour. For a student with no library nearby, that is nothing.
But the imbongi spoke names. He recited lineage, not just content, so that listeners knew whose deeds they were hearing and to whom they owed their inheritance. Knowledge returned without its source is knowledge without lineage, and a culture that receives its past that way may grow better informed while becoming less rooted. It can know what was said without knowing who said it, or why it mattered to them.
The difference between keeping and remembering is the heart of the matter. Keeping is storage. Remembering is the living act of bringing the past into the present so that it shapes who we are becoming. A sealed archive keeps. A teacher remembers. Project Panama built, at great speed and expense, one of the most complete keepers of human writing ever assembled. Whether it ever becomes a rememberer depends on choices still being made: whether companies scan without destroying, whether they share what they hold with libraries where the law allows, and whether the answers their models give carry their sources with them. A sealed archive keeps. A teacher remembers.
Some of those choices belong to companies and courts. Many belong to us. Keep a shelf of physical books at home and let the children see them in use. Write the family recipes down. Tell the stories of the great-grandparents, with their names. Visit libraries, and lend books to friends. In a world where machines are learning everything, the old role of the keeper has quietly passed to every parent, every teacher, and every reader who chooses to hand something on.
The blade in that scanning facility could only ever reach the paper. The rest of memory lives where it has always lived, in the next person we choose to tell.
Follow us on X, Facebook, or Pinterest
