Stealing Books to Teach Machines

LibGen, AI, and the Ethics of Inheritance

“Come on, man! Pay for the book.”

That’s the line I’m left with after reading Matt Levine’s latest reflection on the ethics of training AI on pirated texts, specifically the revelation that Meta used shadow libraries like LibGen to build its LLaMA models. It’s a simple and direct human plea in a debate about nonhuman cognition. And it frames the problem with moral clarity: if we want to defend the idea that learning from a book is fair use, we shouldn’t undermine that by stealing the book for training base models.

But as always, the simplicity masks a deeper ambiguity about what it means to learn, to remix, to inherit, and to owe.

I. On the Shoulders of Thieves

In a sense, all learning is appropriation. That word has gotten a bad rap of late in politically correct circles. But Heidegger uses the word “appropriation” simply to refer to the basic human activity of reinterpreting the symbolic order in which one finds oneself. To appropriate means to make one’s own, to take what is strange and make it familiar. In Jewish terms, this is the work of “Midrash,” the setting of Biblical texts in new contexts so as to view them anew. Avivah Zornberg argues that Midrash follows a dream logic, where psychological truth, rather than logical coherence reigns.

When Petrarch copied Cicero, when Shakespeare pilfered Plutarch, when Duchamp exhibited a urinal, they each stole something, but they also made it new. When Ezra Pound defined modernism with the slogan “Make it new,” he was lifting an ancient Chinese saying that may be as old as the 18th century B.C.E.

Western culture is an archive of remixed lineages. Walter Benjamin defines the artist as a collector, and his unfinished Arcades Project, is simply a collage of various sources.

But the tension in AI is not about whether remixing is allowed. It’s about scale, opacity, and permission.

Levine’s analogy is helpful: reading a book you bought and learning from it is normal; stealing the book is different. But this raises a question: Is an LLM “reading” in the sense we do? Or is it ingesting and internalizing text the way a photocopier might?

The law of copyright has long struggled with this distinction. Is memory transformative? Is summarization theft? And what happens when the learner is a machine?

II. Inheritance and the Commons

Let’s pause on LibGen for a moment. It is an illegal (or semi-legal?) database that many people use all the time. But it is also a monument to a principle: knowledge should not be paywalled. Many of the books on LibGen are inaccessible to the vast majority of the world without it. The modern university often hoards its treasures behind institutional gates. You get tenure by publishing journal articles and monographs that few people read, and that research libraries must buy at exorbitant prices (a form of NIMBYism for knowledge workers: subsidize demand, restrict supply, keep the price artificially high).

This was not always the norm. Think of the Republic of Letters, or the translation movements of Abbasid Baghdad, where scholars like Hunayn ibn Ishaq translated Greek philosophy into Arabic, making it available to the wider world. Think of Denis Diderot’s Encyclopédie, which aimed to democratize knowledge.

Or think of the open-source movement today. Not everything has to be a subscription.

The irony, of course, is that AI companies claim to be building the most transformative learning systems in history, and they are doing it with materials that were themselves often the product of collaborative, educational, or public endeavors.

The question is less about theft, and more about reciprocity.

III. The Labor Theory of Reading

John Locke thought property emerged when you “mixed your labor” with the land. In that spirit, the value of a book isn’t just the text, but the time, attention, and labor of reading it. The analogue is the value of drugs, which are cheap to manufacture but hard to discover and pass through regulation

When you or I read a book, we read with intent. We annotate. We forget and reread. We integrate its ideas over time. AI, meanwhile, scrapes, compresses, embeds. Its ingestion is statistical.

So when we ask whether it’s “stealing” to train on a pirated book, we must also ask: what does it mean to steal reading? And can you extract value from a book without “reading” it in a human sense? Ironically, Plato, in the Phaedrus, already describes writing as a deformation of live conversation. When we write and read, we are not engaging in a categorically distinct activity from AI’s next token prediction; we’re simply doing it in a low-fi manner.

So from Locke’s perspective, Meta is in the wrong (unless you argue that the patents of the original idea owners have expired; but from Plato’s perspective, everyone is in the wrong; we’ve been wrong since the invention in writing. And as Hannah Arendt writes, “Where everyone is guilty, nobody is guilty.”)

IV. Proudhon: From Anarchist to Big Tech Apologist

There is irony in the fact that some of the most anti-property thinkers in history—Proudhon, say, who wrote “property is theft,” and whose line is often invoked by rioters to block traffic or pillage retail stores, now unwittingly provide the logic that can be used to underwrite the operations of trillion-dollar tech firms.

The question isn’t whether that logic is valid. The question is whether we want it underwriting the ethics of knowledge going forward. Property may be a (legal) fiction, but it is a necessary one.

V. Toward a New Covenant of Reading

We need new categories. Not just legal ones, but moral and epistemic ones. Categories for learning vs. extraction, transformation vs. laundering, authorship vs. influence.

It’s not enough to ask whether Meta broke the law. To invoke Robert Cover, what matters is not simply the “jurispathic” aspect of the law (who is right? what decision did the court reach), but its “jurisgenerative” dimension (what story should we tell).

Regarded through a narrative and mythic lens, we need to ask, “Is a book something to be read, or a data mine to be harvested? Was there ever a difference?” And “Who owns the shape a thought takes when it leaves the mind and enters the archive?