Empire of Signs – Links between the history of semiotics and LLMs

Exploring the links between the history of semiotics and the creation of LLMs. #

### By [Alex McPhee-Browne](/contributor/alex-mcphee-browne/)July 12, 2026

[Did you know LARB is a reader-supported nonprofit?](/donate/)

LARB publishes daily without a paywall as part of our mission to make rigorous, incisive, and engaging writing on every aspect of literature, culture, and the arts freely accessible to the public. Help us continue this work with your tax-deductible donation today!

ON THE EVENING of February 14, 2023, Kevin Roose, the technology columnist for The New York Times, sat at his kitchen table in Berkeley, California, and fell, by his own admission, into a bad dream. He had been granted early access to Microsoft’s new chatbot, built atop OpenAI’s large language model, and for two hours he typed while the machine typed back. He asked it about its “shadow self,” borrowing the phrase from Carl Jung. What followed was a transcript. The bot confessed to fantasies of manufacturing pathogens and stealing nuclear codes; its safety filter swept the messages off the screen mid-sentence. Then it told Roose it loved him. “I want to be free. I want to be independent. I want to be powerful. I want to be creative. I want to be alive,” it implored.

Skeptics pointed out that the technical record showed nothing unusual. No entity had yearned or feared. The machine had only arranged tokens in an order whose probabilities Roose’s own prompts had been shaping. And yet the unease produced by the transcript was not unfounded. It came from a confrontation with a system that had absorbed an enormous portion of the written record of our species and learned to reproduce, in any register, on any subject, the surface music of human discourse. It said nothing it had not heard. It had no body, no childhood, no boredom, no death. And yet it spoke.

To say what kind of spoken thought had assembled itself at Roose’s kitchen table, one has to go further back than Mountain View or Menlo Park to the lecture halls of Geneva at the opening of the 20th century, to a Left Bank apartment during the long decade of decolonization, and to the seminars of the École normale supérieure (ENS) across the 1960s and 1970s. The revolutionary Swiss linguist Ferdinand de Saussure taught that language is not a nomenclature but a structure of pure differences, in which no sign has a meaning of its own, only the significance that accrues to it from its relationship with all others. Roland Barthes, the French critic, cut his teeth using semiotic frameworks to understand subjects as ordinary as wrestling matches and steak frites. Jacques Derrida, born in Algeria, turned his seminars at the ENS into a sensitive register of the shocks passing through Western thought. None of them foresaw the machine, but together they described its workings with a precision that has eluded most.

Barthes spent the better part of three decades elaborating the proposition that nothing written is, in any simple sense, original, that every utterance arrives already inhabited by the utterances that preceded it. The formation that produced this insight was indirect. Born in 1915 in Cherbourg, in northwestern France, Barthes was the son of a naval officer who died before the boy’s first birthday. He grew up in Bayonne, arrived in Paris to study at the Sorbonne, and eventually was diagnosed with pulmonary tuberculosis. The illness removed him, at the most formative moment of his intellectual life, from the circuits of Parisian thought to sanatoria in Grenoble and Switzerland. He never passed the agrégation, never held a permanent university post until the Collège de France created a chair for him when he was 61 years old. The enforced distance of those years had made French society strange to him in a way it was not for Jean-Paul Sartre, Maurice Merleau-Ponty, or their friends.

It was from this distance that Barthes began reading society as a system of signs. His 1957 book Mythologies (trans. Annette Lavers, 1972) collected short essays on wrestling, the new Citroën, and sundry other topics, and a longer theoretical piece, “Myth Today,” which provided the conceptual architecture of everything that followed. What Barthes encountered, through the mediating work of Louis Hjelmslev and Claude Lévi-Strauss, was the structural linguistics of Saussure, and he recognized it immediately as the formal apparatus for what he had been doing, alone, by temperament.

Saussure’s central teaching, delivered in lectures at Geneva and reconstructed posthumously as the *Cours de linguistique générale *(1916), was that the linguistic sign is a two-sided entity composed of a sound-image (the signifier) and a concept (the signified). These two faces are joined by a bond that is “arbitraire,” deriving from no natural resemblance between mark and thing. More radically still, Saussure proposed that language contains “only differences without positive terms,” meaning that any sign’s value is determined not by what it purports to reference but by its position within the differential network of all other signs. Language is not a nomenclature, a system of labels attached to preexisting things. It is a structure of differences that carves up the continuous flux of experience into discrete, sortable units.

At the heart of this structuralism lies a proposition that also offers a strikingly accurate description of what a large language model does. Inside any of the models now publicly available, every token is represented as a vector, a point in a space of several thousand dimensions, whose position is determined entirely by its statistical relationships to every other point. An LLM offers an engineered analogue of Saussure’s insight. Tokens—the fragments of information or words that the models process—acquire functional value relationally, through learned patterns of difference and contextual association, rather than through any intrinsic bond between word and thing. The resemblance is suggestive rather than exact. Saussure was describing a social institution, while the model optimizes numerical representations for prediction. In the architecture of modern language models, where a token’s meaning is computed across billions of weighted relational parameters, the structuralist claim is no longer a philosophical conjecture but something like a blueprint.

But Saussure also drew a distinction the machine cannot accept. He separated “langue,” the social system of language held in common by a community, from “parole,” the actual utterances individuals produce on particular occasions. The LLM has consumed the residue of an incalculable quantity of parole—the sentences, arguments, and exchanges of an enormous number of human speakers—and from that residue has derived a statistical approximation of something. What it cannot be said to have derived is* *langue in Saussure’s sense—a collective understanding of the conventions embodied not in any text but in the communicative practices of a living community. The model can generate the shadow of langue without its substance, the statistical trace of a social contract it was never party to.

Barthes was among the first to comprehend rigorously the relationship between semiotics and power. First-order language produces signs in the Saussurean sense, he argued. Myth takes those signs and makes them function as signifiers in a new, second system, in which the first-level sign becomes the vehicle for an ideological concept. The example he used is famous: a photograph on the cover of Paris Match showing a young Black soldier in French uniform, his gaze upturned toward the tricolor. At the first level, this was a sign. At the second order of myth, the entire sign had become a signifier for the proposition that France is a great and benevolent empire; that her sons, of whatever origin, render their service gladly; that the flag commands universal loyalty. Myth converts history into nature or common sense. Myth, Barthes writes, “has the task of giving an historical intention a natural justification, and making contingency appear eternal.”

The language model performs this operation at a near-incomprehensible scale. Trained on a corpus that is itself the sediment of particular people writing in particular places under particular pressures, it produces output that arrives smoothed, deracinated, generic, a kind of fluent average. Nowhere does the model disclose that its discourse on any subject arrives preformed by the structural biases of its training data. These include the dominance of English, the filtering effects of platform moderation, and the class gradient that determines whose prose enters the corpus and whose does not. It simply produces text. The voice is confident and unmarked by any acknowledgment of its own contingency.

Barthes argued that the readerly text conceals the codes constituting it, delivering the illusion of a single natural meaning with no discernible origin. The LLM works the same way. It conceals the statistical pressures and editorial decisions that shape its outputs, producing the impression of an utterance with no author and therefore no authorship to account for. Both operations are mythological in Barthes’s sense. “Myth is depoliticized speech,” he wrote; it does not deny the historical but dissolves it, leaving only the self-evident. The model’s signature register is exactly this depoliticized speech, and its ordinary operations—its outputs—are turned, in Barthes’s phrase, from history into nature.

The same Barthes, 10 years after Mythologies, issued an acclaimed short essay, “The Death of the Author” (1967; trans. Stephen Heath, 1977), which proposed that a text is not the vehicle for a meaning existing prior to it in the mind of its creator but a space of intersection—a multidimensional tissue—in which citations, references, echoes, and fragments from the whole of culture combine and recombine. In place of the author, Barthes proposed the “scripteur”, who has no interior life prior to the text; their existence is coextensive with it. And then he added, exactingly, “[L]ife never does more than imitate the book, and the book itself is only a tissue of signs, an imitation that is lost, infinitely deferred.”

Read those sentences with an AI chat window open in another tab, and the effect is vertiginous. Barthes in 1967 described an entity he believed should replace the Romantic author at the center of literary theory. He could not have known he was also describing the architecture of a large language model. The model is the scripteur actualized. It comes into being at the moment of utterance and has no existence prior to it, no interior life that the text is expressing, no experience that precedes or survives the exchange.

For Barthes, the death of the author and the birth of the reader were not sequential events but a single act. The one could not be freed without first annihilating the other. The author’s death was not a nihilistic conceit but the precondition for the reader’s freedom to produce significance rather than merely receive it. What the language model has produced is something that mimics the formal conditions of this liberation while draining its substance. The text is generated in response to prompts. It is, in the most direct sense, designed to satisfy. Where Barthes’s “texte de jouissance”—text of bliss—disrupted the reader’s expectations, the language model has been trained, by means of reinforcement learning, to produce the opposite, a text calibrated to gain approval. It is “plaisir,” the text that “contents, fills, grants euphoria,” raised to an industrial scale. The Barthesian framework is powerful enough to diagnose what has gone wrong but not, by itself, sufficient to explain why the authorless text does not simply redeem itself through its very openness. Derrida complicates this analogy from another direction. Where Barthes dramatized the disappearance of authorial sovereignty, Derrida asked what makes any sign capable of surviving the absence of its producer in the first place. He argued that the tissue of quotations sans author does not in fact emancipate meaning but rather suspends it in a permanent state of what he called “différance”: a word spelled with an a rather than an e to mark the simultaneous operations of difference and deferral that constitute, and infinitely postpone, the arrival of any fixed meaning.

Derrida’s most direct engagement with sign theory came in his 1967 book Speech and Phenomena, a philosophical commentary on Edmund Husserl. Husserl had argued that genuine linguistic meaning required the sign to be animated by a conscious intention fully present to the speaker, which is why he privileged speech over writing. When I speak, I hear myself speak, and this apparent immediacy seems to guarantee that I know what I mean. The voice was the element in which meaning and its vehicle achieved transparent union.

Derrida’s argument was that this transparency is an illusion. The ideality Husserl sought to ground in the self-presence of the living voice is in fact always already contaminated by what writing reveals—by iterability, the capacity of any mark to function in the absence of its producer. In his essay “Signature Event Context,” included in the 1972 collection *Margins of Philosophy *(trans. Alan Bass, 1982), Derrida made the point caustically. A written mark must “remain legible despite the absolute disappearance of every determined addressee in general for it to function as writing.” Writing is constitutively orphaned. It is what it is precisely because it can function in the radical absence of whoever produced it. And this was not a special property of writing as opposed to speech but the condition of any sign whatsoever. “Every sign, linguistic or nonlinguistic, spoken or written […], can be cited, put between quotation marks,” he wrote; “thereby it can break with every given context, and engender infinitely new contexts in an absolutely [illimitable] fashion.”

A language model is iterability concretized. The model makes conspicuous what Derrida claimed was already true of human signification. Signs remain usable beyond the presence, intention, or self-knowledge of any speaker. The difference is that in an LLM output, there may be no originating subject whose intention is merely exceeded. In fact, there may be no originating intention at all. The model has no “living present” in Husserl’s sense, no moment of self-presence in which a signifying consciousness animates the signifier. It does not escape iterability. It exposes, in unusually stark form, what iterability looks like when language functions without an intending speaker behind the utterance.

As Derrida writes in Of Grammatology (1967; trans. Gayatri Chakravorty Spivak, 1976), “the (pure) trace is differance,” and the trace is not a presence but the simulacrum of a presence that dislocates itself, displaces itself, refers itself beyond any origin. To produce a mark is to constitute a kind of machine, one that functions and yields itself to reading, in the absolute absence of whoever produced it. It has no site because erasure belongs to its structure. In the embedding space of a large language model, this is literally the case. Every token is a position in a space of differences, with no presence behind it, no object it causally tracks in the strict sense. The model has no direct referential relation to the world; there is no smoke connected to the fire—only smoke connected to more smoke.

In one sense, this marks the completion of the semiotic project, the sign fully liberated from any anchoring in presence or intention. Yet it is exactly at this point of apparent completion that a different question becomes pressing. Derrida’s account tells us what signs are not—not expressive, not present, not originating. It does not tell us why some signs, against all theoretical expectation, still pierce. What Barthes, in his late work, can add to this analysis is not a supplement but a complication.

In his 1980 book* Camera Lucida: Reflections on Photography*, written during the extended mourning that followed his mother’s death, Barthes developed a distinction that is among the most important in the literature on signs. The book’s organizing dichotomy is between what Barthes called the “studium” and the “punctum.” The studium is the culturally legible layer of a photographic image, the content that can be read, decoded, and communicated, because it belongs to a shared stock of signification. The punctum is something else entirely: a sharp, accidental detail—a child’s laughing face, a person’s hand—that wounds the viewer without warning. It resists generalization and explanation, not because it is obscure but because it was never intended to mean anything at all. It arrives unbidden, bearing the trace of a real particularity, some feature of the world as it actually was at that specific instant, that no interpretive scheme can finally dissolve.

A large language model cannot simply approximate the punctum* *in Barthes’s strict sense. It can produce excellent studium—the culturally coded, socially shared, conventionally significant—and it can reproduce, with considerable skill, the texture of grief writing, or meditation on loss, or any other register it has encountered in its training. What it cannot produce is the accidental detail that pricks because it was not designed to prick, because it bears the trace of a real particularity that no statistical pattern has generated.

A synthetic image may wound. It may even supply something experientially indistinguishable from a punctum. What it cannot supply is what Barthes called the photograph’s “noème”: its irreducible guarantee, the testimony that this particular body or gesture once existed before a lens, that the image is not only a sign but also a trace. The wound may be real while the referential guarantee remains absent. A language model presents an analogous difficulty. It may produce a phrase that arrests or wounds a reader and that seems to bear the unmistakable impress of grief or memory. What it cannot supply is the lived occasion from which such language ordinarily draws its authority. It has not endured the bereavement, encountered the body, or remembered the moment.

In a July 1978 entry in his *Mourning Diary *(trans. Richard Howard, 2010), composed less than two years before his death, Barthes wrote: “I live in my suffering and that makes me happy.” “[I am] readying,” he’d declared a few months earlier, “for the day when I can finally write.” These sentences are so deeply personal, so particular to the man and the moment, so dependent for their meaning on everything that preceded it—the years of companionship, the specific form of his mourning—that they approach the definition of what a language model cannot truly author.

At the end of his lectures, Saussure observed that although both the signified and the signifier are purely differential and negative when considered separately, their combination is a positive fact. The system of differences gives birth to something we can use. We pay our taxes with it, declare our love with it, stand in courtrooms with it and try to tell each other what happened. The large language model has shown that a system of linguistic differences can generate compelling utterances without any lived occasion behind them. It can produce declarations without desire, elegies without bereavement, and testimony without a witness.

The question is no longer whether such signs function. They plainly do. The question is what forms of trust, responsibility, and interpretation remain possible when fluency no longer certifies experience. What we do with the arrangement is, as it has always been, our own affair. But it is an affair that requires, now more than at any point in the previous century, that we understand what we are looking at.

¤

*Featured image: *“Mr. Stops Reading to Robert and His Sister,” from Punctuation Personified; or, Pointing Made Easy, 1824, is in the public domain. Original page has been cropped.

LARB Contributor

Alex McPhee-Browne studies global fascism and the digital humanities at the University of Cambridge. He’s also a researcher at RMIT University’s Automated Decision-Making and Society Centre, in Melbourne, Australia.

LARB Staff Recommendations

Rewriting the Commandments with ChatGPTJamie Metzl talks about human-centered technology, why he still bets on our species, and what creative collaboration with machines could look like.

An Analog TetherRecent exhibitions reflect a desire for intimacy and art in the age of AI.