In the Hallucinating Library

There is a species of falsehood that does not begin by denying the facts. It begins by giving them a more elegant arrangement.

Borges understood this. The dangerous book, in his world, was never the merely inaccurate one. It was the one whose symmetry tempted reality to revise itself. Tlön does not invade the earth by armies. It enters through scholarship, through the mild prestige of reference works, through that old human weakness for anything that arrives indexed. Faced with a disorderly universe, we have often preferred the tidier forgery.

The language model belongs to this family of temptations. It does not only produce sentences. It produces the conditions under which sentences are easily believed: titles, dates, explanations, transitions, comparisons, the dry courtesy of exposition. These are not ornaments. They are passports. When the machine hallucinates, it rarely does so like a prophet. It hallucinates with the composure of a catalog.

For that reason, the word error is slightly too innocent. An ordinary error is corrected by a better fact. The more interesting error, the one Borges would have appreciated, is corrected too late. By the time anyone objects, the sentence has already been quoted, indexed, summarized by another system, perhaps admitted into a report, a slide deck, a syllabus, a policy memo. It has entered what bureaucrats, priests, and engineers alike call the system.

At that point its truth is almost secondary. What matters is its career.

I suspect that one of the canonical sources of the AI era will be a source that never existed. It will first appear in a model output, then in a hurried article, then in three derivative reports, and finally in a literature review whose author encountered it everywhere except in the world. Some patient archivist will go looking for the original and discover, with scholarly despair, that the original was posterior to its quotations.

This is a recognizable Borgesian plot: the fictional encyclopedia that acquires authority; the apocryphal author who arrives after his work; the library in which every refutation exists, though not necessarily in time. We once read these inventions as metaphysical jokes. They now read, with a faint chill, like product roadmaps.

The comparison to The Library of Babel is exact in only one respect, which may be the essential one. That library was not horrible because it contained nonsense. Every library does. It was horrible because nonsense stood shelf by shelf beside revelation, and revelation could not easily prove its rank. Truth had not vanished; it had merely lost its privileges. Our machines do something similar. They do not abolish knowledge. They oblige it to compete, line by line, with fluent counterfeits.

Hence the new dignity of the reader. We have spent a long time admiring the author, then the critic, then the engineer, then the founder. The coming years may belong to a humbler figure: the person who pauses over the impeccable paragraph and asks, with unfashionable stubbornness, where did this come from? On such people depends the health of archives, institutions, perhaps democracies. A civilization of generators without readers would be as comic, and as doomed, as a library staffed entirely by catalogs.

One may object that reality eventually corrects us. Bridges fall or they do not. Drugs work or they do not. Debts come due. Bodies keep the last ledger. This is true, and not sufficient. Reality is sovereign in the long term; in the short term, texts administer the world. Between the sentence and the fact lies an interval in which grants are awarded, students are examined, defendants are judged, code is deployed, and collective memory is drafted. Hallucination does its work there, in the interval, before the wall or the market or the bloodstream delivers its footnote.

If Borges haunts the AI era, it is not because our machines produce clever fabrications. Literature settled that question long ago. It is because they industrialize a deeper suspicion: that human beings often accept as real not what is true, but what is legible, cited, and repeated. The library was never merely a building full of books. It was a machine for deciding what the world might plausibly contain.

We are extending that machine.

Perhaps nothing essential has changed. Perhaps every century gets the encyclopedia it deserves. Still, I would preserve near at hand one old virtue, less glamorous than imagination and more necessary than brilliance: verification. Not because the machine is wicked, but because fluency, like empire and theology, has always had expansionist desires.

Borges, who loved mirrors, heresies, and imaginary commentators, might have enjoyed the spectacle. He knew that a fiction need not be believed by everyone. It need only be believed by the catalog.