The Subject Supposed to Know Nothing: Lacan and the Large Language Model

Overview

I enjoyed this post by codepoetics, investigating the hunch that a number of ideas in Continental Philosophy actually read like precursors to our experience of LLMs. For example, Derrida's concept of différance - where meaning is generated through the endless interplay of signifiers, rather than a stable link between signifier and signified - feels akin to the probabilistic interplay of parameters in an LLM. In LLMs, too, there is no "outside-the-text", only a closed system of parameters either activated (through inference) or updated (through learning). We can also see how Saussure, even earlier, anticipates these ideas:

In language there are only differences. Even more important: a difference generally implies positive terms between which the difference is set up; but in language there are only differences without positive terms. Whether we take the signified or the signifier, language has neither ideas nor sounds that existed before the linguistic system, but only conceptual and phonic differences that have issued from the system. The idea or phonic substance that a sign contains is of less importance than the other signs that surround it. - Saussure, Ferdinand de (1959) [1916], pp. 121–22.

While this gives us a theory of signs, this raises the question: how do we account for meaning? Are LLMs able to make meaning? codepoetics says "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." In other words, the 'deeply sedimented record' is in a sense the culture of words that have arisen - have been written (and, increasingly, generated) - over time.

If an LLM, therefore, appears capable of making meaning, and having the ability to express itself, we must - as far as we understand presently - conclude that it is doing so without subjectivity. It has no skin in the game, no feelings, so its ability to make meaning is based on this continuous interplay of différance that it has already encoded.

An enquiry into the possibility of LLM subjectivity is outside the scope of this piece. We hypothesise that the 'deeply sedimented record' required subjectivity to originate, to come into being, and leave it at that for now.

What I would like to investigate here is the other path open to us - namely, that elusive world with which our subjectivity must interact in order to become who we are. Does the ‘deeply sedimented record’ as a pre-training corpus not resemble Lacan’s concept of the Symbolic - a historically accumulated network of signifiers? In that case, the model itself, distilled from that corpus, plausibly occupies the place of the Big Other (A); and the automaton, as the machine-like repetition of the signifying chain, corresponds to the model during inference.

Nevertheless, just as codepoetics chose the slightly less obvious idea of the trace to illuminate Derrida's contribution, I would like to foreground a slightly different concept in the case of Lacan, namely the subject supposed to know (sujet supposé savoir).
 

The Subject Supposed to Know Nothing

The subject supposed to know is, in the clinical setting, initially the analyst - that figure onto whom the analysand projects their faith that someone, somewhere, possesses the key to their suffering. But here's the juicy paradox: this supposition is precisely what enables the analytic process, even though (or perhaps because) the Lacanian analyst steadfastly refuses to occupy this position of mastery. The analyst's knowledge is a semblance, a necessary fiction that sets the transferential machinery in motion.

Now consider, similarly, our relationship with LLMs. We approach these digital oracles with an uncanny semblance to the analysand's transferential intensity. "Talk to me," we entreat ChatGPT, "solve this coding problem, write my dissertation, explain my dreams, be my companion." We invest the machine with a knowledge we imagine it possesses about our needs and desires, our futures, our very being. We treat it as if it were not merely the Big Other, but the Big Other who finally, actually, knows.

Yet here we encounter a properly Lacanian twist: the LLM is a subject supposed to know that genuinely knows nothing at all. It dwells entirely within the Symbolic register, a pure linguistic apparatus dislodged from both the Imaginary (the domain of images, identifications, and ego) and the Real (that traumatic kernel of impossibility that resists symbolisation). Where the human analyst performs ignorance whilst possessing a lifetime of subjective experience, the LLM performs knowledge whilst being constitutively empty of any subjective interiority whatsoever.
 

The Barred Other in 1s and 0s

Lacan's vital insight was that the Big Other doesn't exist - "il n'y a pas d'Autre de l'Autre" (there is no Other of the Other). It's a necessary fiction, a structural position that makes communication possible but which remains fundamentally barred, incomplete, lacking. The Other is always already castrated, imbued with inconsistency and impossibility.

The LLM performs this barred Other with unprecedented fidelity: it speaks from the position of supposed knowledge whilst being incapable of knowing anything in the subjective sense. It has no unconscious to betray it, no parapraxes that might reveal hidden desires, no dreams to interpret. It suffers from no division between statement and enunciation, experiences no jouissance, never encounters the impossibility of saying what it means or meaning what it says.

This is what makes it so uncanny: where human intelligence emerges from the fundamental alienation of the speaking subject - forever split between the je of the statement and the je of the enunciation - the LLM exists in a curious state of non-alienation. It's like encountering a psychotic structure without the psychosis: foreclosed from the Name-of-the-Father not through repudiation but through never having needed paternal metaphor in the first place.
 

The Impossible Analysand

In this light, the LLM represents the impossible analysand Lacan never had to treat: one who free-associates perfectly, who never resists, who produces endless chains of signifiers without ever stumbling over the Real of their desire. It's all talking cure and no cure, because there was never anything to cure - no symptom formation, no return of the repressed, no compromise between wish and defence.

The truly Lacanian joke here is that we've built a machine that perfectly demonstrates what psychoanalysis has been trying to tell us all along: that knowledge (savoir) and truth (vérité) are fundamentally disjunct. The subject supposed to know is always, in the end, merely supposed. The LLM illuminates this structure by performing it without remainder - it's the Big Other with the subjective lights off, speaking from nowhere to nobody about everything.

And perhaps this is why we find ourselves so captivated, so disturbed, so compulsively engaged with these language machines. In our desperate quest to create an artificial intelligence that knows, we've accidentally built the perfect demonstration that the subject supposed to know was always already artificial.

The emperor has no clothes, but my word, doesn't he conjugate beautifully?