The Pasticheur’s Machine

The most accurate description of what a large language model does was written in 1950 by a philosopher who had never seen a computer. In his essay “La décadence,” Vladimir Jankélévitch describes how creative consciousness, once it becomes aware of its own operations, enters a regime of self-imitation: “the creative invention imitates itself and exploits itself and dilutes itself and lives off its own royalties before the studious epigones have even had the time to copy it.” Replace “creative invention” with “training data” and “studious epigones” with “transformer architecture,” and you have a serviceable account of GPT.

This is not a coincidence. The questions that LLMs force upon us — What is originality? Can a machine have a voice? Is there a difference between imitating a style and understanding it? — are the same questions that literary critics have been asking about pastiche for two centuries. The fact that we are asking them again, this time about silicon rather than ink, does not make them new. It makes them urgent.

The anatomist and the lover

There are two ways to learn a style. You can analyze it: count the sentence lengths, tabulate the vocabulary frequencies, map the syntactic patterns, build a statistical profile of every measurable feature. Or you can absorb it: read so much of an author that his rhythms become your rhythms, his instincts become your instincts, his way of thinking-in-sentences becomes indistinguishable from your own. The first method produces knowledge about; the second produces knowledge of. The scholastics called the difference notitia versus sapientia. The computational linguists call it feature engineering versus end-to-end learning.

The history of NLP is, in a precise sense, the history of this distinction. Rule-based systems and early statistical models were anatomists: they dissected language into parts of speech, dependency trees, n-gram frequencies, and operated on these explicit representations. They knew everything about language and understood nothing of it. The transformer revolution was a shift from anatomy to absorption: instead of engineering features, the model soaks in text — billions of words, trillions of tokens — until the patterns of language have been absorbed into the weights of the network with an intimacy that bypasses explicit representation entirely. The model does not “know” that English prefers SVO word order; it has never been told. But it has absorbed SVO order the way a child absorbs it: by immersion, by saturation, by the statistical equivalent of what I have elsewhere called “the lover’s knowledge.”

This is what makes LLMs so uncanny. They reproduce the surface of human prose with an accuracy that no rule-based system could approach — not because they understand the rules better, but because they have bypassed the rules entirely and gone straight to the texture. They are, in the exact sense of the word, pasticheurs: they have entered the voice of another (the collective voice of their training data) and can produce sentences that sound as if they belong to that voice. The question is whether this reproduction constitutes understanding or merely resemblance — whether the LLM is the lover or only the anatomist who has, by some computational miracle, achieved the lover’s results without the lover’s experience.

The ventriloquist’s residue

In my recent essay on pastiche, I argued that every act of stylistic imitation oscillates between two impossibilities: total coincidence with the model (which would erase the imitator) and irremediable distance from the model (which would erase the imitation). The pasticheur lives in the space between — too present to be transparent, too absent to be author. This formulation applies to LLMs with an exactness that is almost embarrassing.

When ChatGPT produces a paragraph in the style of Hemingway, whose voice are we hearing? Not Hemingway’s — Hemingway is dead, and the model has never read Hemingway in the way that a human reader reads Hemingway, with the biographical context, the emotional resonance, the sense of a particular man writing in a particular room at a particular moment in history. But not the model’s own voice either, for the model has no “own voice,” no native idiom from which it departs in order to enter the foreign idiom of the imitation. The LLM is a pasticheur without a self — a ventriloquist who is nothing but ventriloquism, a mask with no face beneath it. And this is precisely what makes it so philosophically interesting: it is the limit case of pastiche, the thought experiment that human pasticheurs have always implicitly posed, now realized in hardware. What would a pastiche be if there were no one performing it?

The answer, I think, is: exactly what we see. A prose that is fluent, coherent, syntactically impeccable, and yet faintly hollow — not because of any identifiable deficiency, not because of any grammatical error or logical inconsistency, but because of the absence of what Jankélévitch calls the “irreducible residue,” the grain of sand in the mechanism of identification that betrays the presence of a living consciousness behind the imitation. The human pasticheur, however skilled, always leaves this residue — a trace of his own voice, his own concerns, his own mortality — and it is this residue, paradoxically, that gives the pastiche its life, its pathos, its value as a human document. The LLM leaves no such residue. Its imitations are technically superior to anything a human pasticheur could produce, and for exactly this reason they are less interesting: they are perfect masks with nothing behind them, perfect echoes of a voice that no one is speaking.

What the machine reveals

But the LLM also reveals something that literary theory has long suspected but never been able to prove: that “originality” is, at bottom, a statistical phenomenon. If a model trained on a sufficiently large corpus can produce prose that is indistinguishable from human writing, then the patterns of human writing — the patterns that we experience as “style,” as “voice,” as the irreducible signature of an individual consciousness — are, in principle, extractable from the aggregate. The individual voice is a deviation from the collective norm; style is the residual after the statistical average has been subtracted. This is not a reductive claim. It does not diminish the achievement of the original writer to observe that his originality is, from a certain angle, a kind of noise — a departure from the expected that registers, in the consciousness of the reader, as surprise, as freshness, as the unmistakable mark of a particular mind. On the contrary: it makes originality more precious, not less, to know that it is rare in the precise mathematical sense, that it occupies the tails of the distribution, that the machine can reproduce everything about a writer except the thing that makes him irreplaceable.

This has practical implications for AI alignment, for content generation, for the future of writing as a profession. But it also has implications for literary theory — implications that, as far as I can tell, the literary theorists have not yet grasped. For if the LLM is a pasticheur, then the entire critical apparatus that literary studies has developed for thinking about pastiche — about influence, about imitation, about the relationship between originality and tradition, about the anxiety of inheritance and the burden of the past — becomes suddenly and urgently relevant to the most consequential technological development of our time. The humanists have the concepts. The engineers have the systems. What is missing is the bridge.

I am trying to build that bridge. My doctoral research at the Sorbonne — on the aesthetics of decadence and the critical rhetoric of style in nineteenth-century French and Chinese literature — trained me to think about exactly the problems that LLMs now pose in a new key: the problem of style as both individual and collective, the problem of imitation as both parasite and preserver, the problem of the relationship between what a text says and how it says it. My work as an AI researcher has trained me to think about these same problems in the language of transformers, attention mechanisms, and embedding spaces. The two vocabularies are not as far apart as they seem. The nineteenth-century critic who diagnosed the “constitutional illness” of self-conscious style and the twenty-first-century engineer who fine-tunes a model to avoid “mode collapse” are working on the same problem, seen from opposite ends of a century and a half. The question is whether we can get them to talk to each other.

I believe we can. I believe we must. The alternative — a technological revolution that proceeds without the conceptual resources that the humanities have spent centuries developing, a literary criticism that ignores the most powerful reading and writing machine ever built — is a waste that neither side can afford. The pasticheur, Proust told us, understands the original better than the critic does, because the pasticheur’s knowledge is carnal, intimate, performative rather than analytic. Perhaps the machine pasticheur, for all its hollowness, for all its absence of a living voice behind the mask, will teach us something about our own voices that we could not have learned in any other way. Perhaps the best reason to build a machine that imitates human writing is to discover, at last, what it is about human writing that no machine can imitate.