Derrida, Badiou, Baudrillard: Three Thinkers of the LLM

How French Theory foreshadowed the Rise of the Machines

You may have, or think you have, an aversion to postmodern gibberish: dense, foreign-sounding verbiage that spiels on and on, playfully, self-referentially, never connecting with reality; caustically insinuating that no such connection is even possible, that “reality” itself is nothing more than a consensual hallucination. What could such patently irresponsible nonsense have to do with the modern technological marvel that is the Large Language Model? Oh, wait.

In this essay I’m going to outline how each of three French thinkers from the late C20th illuminates, and is in turn illuminated by, the LLM. The three I’ve chosen are Jacques Derrida (1930-2004), Alain Badiou (1937-) and Jean Baudrillard (1929-2007). From each I take a single key term: “trace” from Derrida, “encyclopedia” from Badiou, and “immunodeficiency” from Baudrillard. I will give a brief account of what these terms are doing in each thinker’s work, and show how they anticipate with uncanny foresight the nature of the LLM and the problems it throws up.

Derrida: the structure of the trace

It is not at all by chance that Derrida talks of Joyce’s books in terms of supercomputers, nor that his thought should communicate in an essential way with certain discourses on so-called artificial intelligence. Nor should we have conceived this book a little on the model of a “hypertext” program which would allow, at least in principle, an almost instantaneous access to any page or word or mark from any other, and which would be plugged into a memory containing all of Derrida’s texts…such a machine would suspend reading in an open system, neither finite nor infinite, labyrinth-abyss, and would thus also retain the memory of the traversals tried out, following their nose, their flair, by all its readers, these being so many texts to plug back into the general network.¹

A Large Language Model is, as the name suggests, a model — a large model — of (some) language. A vast corpus of digitised text is surveyed, and a statistical digest of observed sequential relationships is built up: this word often comes before that; the likelihood of this occurring is modified in this way by these preceding contextual markers. The model does not literally store any of the training corpus itself: it encodes only observed statistical relationships, and relationships between relationships, to the nth degree, across billions of parameters. (That said, it is sometimes possible due to overfitting for generation from the model to reconstitute parts of the corpus more or less accurately, which can give rise to awkward issues around copyright).

“Trace”, in Derrida, names not a fixed entity but a differential relation: a manifold of differences between things that are themselves manifolds of differences, and so on, to the nth degree. A trace is structurally a trace of something, or multiple somethings, but the somethings of which it is a trace are also traces (and hence never really positive “somethings” at all). The trace thus marks an indefinitely deferred presence, something that isn’t here — we have a trace of it instead — but that also isn’t there, in the place of the original. Substitute of a substitute, copy of a copy, sign of a sign, without an “authentic” originating presence from which the whole sequence sets off. (If this seems unbearably abstract, or too much like postmodern gibberish, then pause here and re-read the paragraph before this one).

The usual motivating example of this is the way linguistic meaning arises from differential relationships between words (“cat” is not “rat” or “bat”), and associations between words and bundles of other words (for example, between a word and the definitions one might give of its meaning, or between that word and other words commonly linked with it by customary usage). But in talking about the structure of the trace Derrida is not simply advancing a theory of what words are, or how linguistic meaning works. Instead he’s showing how theories of this kind encounter an aporia, a theoretical limit or self-undoing, when they consider language as a structural phenomenon.

If the goal is to produce a stable, “synchronic” picture of language as an arrangement of discrete elements into a single and total structure, the intrinsically differential nature of these “elements” undermines that goal: wherever you try to pin one down, it turns out to be differentially constituted by other differential constitutions — and so on, to the nth degree. Derrida calls this elusiveness différance, a neologism indicating both spatial difference and temporal deferring, and argues that any structural theory of language (or anything else pictured as an organisation of synchronously present elements) will find itself haunted and unsettled by it.

What many people find unsettling about LLMs is that they, like language itself in Derrida’s account, are haunted by an absence of origin or ground. There is no stabilising relationship to truth or presence, no reliable mapping into a domain of reliably bounded and interconnected concepts, just differential relationships between (differential relationships between, and so on) signifiers. And yet, this phantasmagoria of difference emits a plausible apparition of sense. The model seems able to reason and argue and evaluate, operationally mimicking the surface forms of human intelligence. Its epistemic unreliability, its tendency to drift into unmoored improvisation or outright confabulation, is taken as evidence that something essential to the workings of human intelligence is missing; but human beings do not contain truth oracles that reliably connect their utterances to reality either. (Unlike LLMs, however, we can refer to first-person experience in verifying claims, and our handling of second-hand sources is regulated by the social technology of trust and reputation: we operate under an informal regime of accountability, because our perceived reliability is at stake in our assertions.)

Readers of Derrida may recognise in common complaints against the unanchored errancy of LLMs a longing for what Derrida calls the “transcendental signified”, a presence that sits outside the play of differences, bounding and securing it. But Derrida argues that signifieds — semantically stable concepts — are stabilised by and within the play of differences: their stability is contingent upon the system, not exterior to it as a prior condition. The process of meaning-making, through language use, is what makes concepts cohere. If the LLM seemingly has a grasp on the conceptual domain, on the meanings of words, it is because it models a deeply sedimented record of that usage.

Derridean “deconstruction” wagers on the primary process of meaning-making, and the ongoing play of difference within that process: not only is the signifying process responsible for making concepts cohere, it also by the same token exposes them to contingency, to new situations or figurations which can shift the stakes of signification. Thought can be made to “tremble” — indeed, it is already vibrating internally — and deconstruction is in a sense nothing but attunement to that vibration, to its barely-audible resonances, traces within the unfolding moment of a future yet to come. By contrast, the tendency of an LLM is to double down on coherence, to seek the most historically plausible continuation of a prompt. It is a performative, facilitative, reassuring machinery, but also a fundamentally conservative one, smoothing away contradiction, vibration, the dissonance of the real.

Which brings us to:

Badiou: The Encyclopedia

For Derrida the differential structure of the trace means that a structure is never simply the synchronous presentation of a totality of positive elements. However comprehensive a system may appear, and no matter how stable, sedimented and reinforced it may be, it is always inhabited by something elusive and virtual — a constitutive deferral or spacing that prevents closure. It is difficult to reconcile this picture with the reality of the LLM which, stochastic though its generative activity may be, is essentially a single very large (but finite) data structure: a frozen parameter space trained on a fixed corpus. The trace’s differential movement is arrested at the border of the model — or rather, it meets a thermodynamic threshold, where the physical operation of the computational system instantiating the LLM converts the endless lateral slippage of meaning into power consumption and water draw-down.

By contrast, for Badiou the structure of any “being”, or of any “situation” in which beings might appear, is inherently static: an instance of structured presentation, or what he calls the count-as-one. (His mathematical template is Zermelo-Fraenkel set theory). The fundamental condition of Being as such is “pure inconsistent multiplicity” — a non-totalisable field of presentation — and in this sense Badiou agrees with Derrida that the totality cannot be presented synchronously, or as he puts it, that “the One is not”. But beings themselves, once presented in a situation, are nevertheless to be regarded as consistent multiples. The LLM’s parameter space would be just such a being: a structured multiple, counted as one.

The “encyclopedia” of a situation consists of the entirety of the resources available within that situation for identifying, classifying, and naming what appears. It encompasses all the linguistic tokens, conceptual operators, syntactic structures, and recognised forms of inference that the situation makes available — in short, the systematised regime of knowledge. It is very similar to what Foucault calls the episteme of a historical period.

We can see immediately how this corresponds to a certain fantasy of the LLM as an all-knowing oracle, possessing a fully-encompassing model of everything the encyclopedia can say about the situation to which it belongs. In fact it is more like a sampling of the encyclopedia, via a training corpus comprising more or less representative (although in practice endemically biased) statements sanctioned by the encyclopedia. From this sampling it is then able to generate new statements as if it were following the encyclopedia’s own rules for statement production. It thus feigns to reconstruct the encyclopedia from its digital archive by statistical interpolation.

In the Derridean story about the LLM we encountered an anxiety about the lack of an authorising presence, the touchstone of reality, without which meaning-production (or token-emission) might float free, unanchored by conceptual and referential constraints. Derrida shows us is that this anxiety is misplaced: such a presence is not reliably available in everyday human speech situations either. We don’t arrive at truth by ensuring that signifiers are properly yoked to signifieds, but through an ongoing social process of meaning-creation, negotiation, and contestation. In the Badiouian version, we similarly notice that the LLM behaves like an apparatus of knowledge, whose limits are the limits of the sayable, but that knowledge and truth are not the same thing: the encyclopedia furnishes the means for saying what can be said, but not for determining which sayable things are true and which are not.

Badiou decisively rejects the idea that establishing the truth is a matter of sorting true statements from untrue ones. There is no guarantee that the encyclopedia of a situation will contain semantic resources adequate for talking about everything that appears in that situation: we may lack names for things, or the conceptual armature needed to bring particular relationships or mechanisms into focus. In fact, Badiou argues (via a mathematical analogy I won’t get into here) that there must be aspects of the situation that are in excess of its power to name and address them. There will always be some structured multiplicities which are opaque to knowledge, appearing as “black boxes” from the point of view of the state apparatuses which seek to manage them. Badiou calls such a multiplicity an “evental site”, a site “on the edge of the void”, from which a surprise might suddenly erupt.

A truth, on Badiou’s account, is the working out of the practical and epistemic consequences of such a surprise, an “event” which ruptures the organised consistency of the situation and forces a reshaping of the world and what we think we know about it. Scientific discoveries, artistic innovations, political movements erupting from the unseen margins of society, and even erotic epiphanies can all carry the force of an event; each can serve as the precondition of a “truth procedure” which cuts diagonally across the structured multiplicities presented in a situation to compose a new being, insisting on the existence of that which the encyclopedia does not know how to name.

Truth here is radically disjoint from knowledge, and it is also definitionally inaccessible to the LLM, whose entire domain is that which can be named, encoded, and compositionally summoned from the depths of latent space. If it is hoped that “AI” in the form of LLMs will deliver startling new scientific discoveries, efficient solutions to problems in engineering or social organisation, or even the transformative power of love to the chronically lonely, then Badiou is here to dash that hope. All the LLM has to offer us is repetition — in disorienting, permutive profligacy — of the same.

But it’s worse than that. Corruption lurks in the heart of the apparatus of knowledge. Baudrillard knows.

Baudrillard: Immunodeficiency

Up until this point we have been relatively sober, philosophical even, perhaps disappointingly given our opening promise of postmodern gibberish. But now it’s time to reckon with French theory’s Great Accelerator, the one who leapt giddily from the invention of the automobile to the erotic spectacularisation of the car crash²: Jean Baudrillard.

Already we hear murmurs that, as the sophistication of LLMs increases, their hallucinations worsen. Every mechanism introduced to tether them to veracity becomes merely another vector along which their unhinged stochastic performativity propagates. As their outputs flood into the public sphere they pollute future training corpora, replacing situationally apposite human utterance with the calcified fixpoints of the algorithmic ingest-and-secrete cycle. The very totalising enthusiasm of the AI rollout in its imperial phase, racing up the steep slopes of the hype curve, degrades the system’s ability to regulate itself, exposing it to viral subversion and runaway metastasis.

Baudrillard had a name for this: immunodeficiency. It was, at the time, an unquestionably tasteless riff on the AIDS crisis, which Baudrillard depicted with scabrous cynicism as infecting all of civil society with “that other mental AIDS that is the AIDS-a-thon, the Telethon and other assorted Thanatons — expiation and atonement of the collective bad conscience, pornographic orchestration of national unity”. One thinks of the strange rite, at the height of the Covid lockdown in the UK, of people coming out onto their doorsteps to clap for the NHS.

Immunodeficiency, in Baudrillard’s lexicon, is not merely a pathology of individual bodies unfortunately afflicted by disease. It’s a form of systems failure, caused by the suspension of external regulation and the runaway acceleration of frictionless reflexivity. Under these conditions, the slightest mote of subversion instantly explodes into galloping rot. The simulation of emotion and action in social spectacle in due course gives way to the simulacrum, a simulation without original, which runs on its own autogenerative logic. “Irrational exuberance” in financial markets, the tulip-mania of the NFT bubble: energy feeding on energy, signs feeding on signs.

Baudrillard observed the explosion of financialisation and the media sphere through the 80s and 90s with a mixture of horror and exhilaration. He would, I believe, have seen the present AI boom as the inevitable intensification of the same dynamics: a terminal spiral of simulation without exteriority, meaning without referent, sense without ground.

We sometimes treat the LLM as if it were reading the map of meaning laid down in the archive of our speech; as if it were voicing a sort of digitised anima mundi. But it is not merely a reciter of received verities, a glorified search engine indexing a snapshot of human culture. Its generative power, running free of any referent save its own higher-dimensional Calder mobile of statistical weightings, carries it rapidly away from any such relatable origin, into the domain of the simulacrum, whose strange logic of wild feedback loops, sudden reversals and fatal strategies Baudrillard documented as if calmly observing the end of the world. Will the AI kill all humans? On a Baudrillardian account, it already has.

1. Bennington, Geoffrey, Jacques Derrida (London: University of Chicago press, 1993), pp. 314-315.

2. As mentioned in my previous essay, Frederik Pohl famously observed that the task of the science fiction author was to predict not the invention of the automobile, but the traffic jam.