When the Machine Free Associates: Psychoanalysis in the Age of AI

Introduction

This is not a critique of AI from the standpoint of psychoanalytic theory, but an attempt to think with and through the machine as machine, allowing its associative logic to reveal the foundations of our own symbolic entanglements. Lacan’s notion, as stated in Écrits, of “the method that proceeds with the deciphering of signifiers without concern for any form of presumed existence of the signified” (Lacan, 2006, p. 630) underscores this perspective by emphasising a structural rather than content-based approach to meaning. What the reader will not find in this paper, therefore, is a discourse with an AI bot, except for one small example. The analysis presented is a structural one, drawing on the principles of transformer models based on Vaswani et al. (2017). The inner structure of a transformer model is not something we have to wrestle from traces of speech, but we can directly understand the mathematical arguments and formal procedures that allow these models their impressive use of language. There is no inductive gap between the model and the analysis, which is important to keep in mind, as there are numerous studies on LLMs that use methodologies based on the assumption of such a gap (see for example Hagendorff et al., 2023; X. Li et al., 2024; Yuan Li et al. 2024; Simons et al., 2024). But in the same sense that we do not need a quantitative analysis of set theory to explain it, we do not need to look at hundreds of variants of an LLM’s output to understand how it works: we can directly access the core principles of its computations. Moreover, a discourse with a chatbot could very well be fooled by the bot’s knowledge of the questionnaire. Turing games are in a sense rigged from the start, not by the machine, but by the epistemic deficiency of the questioner. The LLM already knows the terrain, it has learnt the accepted questionnaires, it knows the Turing test and the literature around it, and the carelessly worded prompt will only throw the chatbot’s answer in the direction of the learnt patterns. A structural analysis is not threatened by such a trap. But are we not just as rigged as the machine? Freud (1901) assumed a determinism (p. 13) and Lacan (2006) an automatism (p. 519) of the ‘raw material’ of association (Freud, 2010, p. 545); the primary processes are automatons, so the analysis of an associative machine might intrude closer to us than we would like.

Methodology

Let us move quickly through a framework for linking LLM’s architecture to psychoanalysis, which I have detailed elsewhere (and will continue to do so). In Lacan’s thinking, we find that language is shaped by two main associative processes: metonymy, which involves relational meaning within a sequence of signifiers, and metaphor, which involves substitution between signifiers. Lacan (2006) formalised these mechanisms with the following formulae (pp. 428-429):

Metonymy: f (S…S’) S

Metaphor: f (S’/S) S

These formulae aim to encapsulate how signifiers derive and transform meaning. In metonymy, meaning emerges through the relationship between signifiers. Conversely, in metaphor, meaning shifts through substitution. Transformer-based language models (LLMs) similarly engage in these processes by dynamically generating meaning within activation spaces and adapting associations based on input prompts.

Lacan’s formula of metonymy indicates that the meaning of a signifier is tied to its relational proximity to others, in line with Freud’s notion of words as “nodal points of numerous ideas” (Freud, 2010, p. 355). This relational understanding avoids attaching meaning to a fixed signified and offers a purely formalised approach to language. This perspective resonates with the embedding spaces in LLMs, where the meaning of a token is determined by its position within a high-dimensional relational space. This mirrors Lacanian metonymy, as a vector’s meaning, that is the word representation in the mathematical embedding space, stems from its proximity to others (first introduced in Mikolov et al., 2013). Therefore, the “meaning” of a vector arises from its neighbours in the embedding field, reflecting Lacan’s theory of metonymic chains. However, a true associative chain is more than just a metonymic chain, as demonstrated by the dual use of metaphor and metonymy in even seemingly standard metonymic associations in LLMs, but I will discuss this below.

For metaphor, Lacan’s formula describes a substitutionary dynamic where one signifier modifies another, producing new meaning through associative displacement. Freud’s concept of “condensation” captures this process as a “fusion between […] two groups of ideas” (Freud, 2010, p. 199) where one part alludes to the other. For example, the metaphor “time is a thief” merges the fields of “time” and “thief,” borrowing associative power.

LLMs similarly construct novel relations by exploiting the attention mechanism (Vaswani et al., 2017), dynamically contextualising word meanings based on relational patterns in training data. While LLMs initially represent tokens as static embeddings, these are transformed by self-attention and feed-forward networks, enabling contextual meaning. This context sensitivity facilitates metaphorical connections, but also reinforces metonymic associations. In Lacanian terms, self-attention allows the model to sever a signifier from its lexical bonds (Lacan, 1993, p. 218) and link it to others, forming new associations (p. 219). It’s important to think of this as metaphoric in the Lacanian sense, not as a symbol or image in the common use of these terms. This is central because the metaphor is then deeply rooted in language, not a secondary embellishment.

Let us look at how this happens in LLMs in a strongly simplified form. Consider again the sentence “time is a thief.” Through attention, the model evaluates the relational influence of “time” and “thief,” recalibrating the activation for “thief” to reflect features of “time.” To understand how the words relate to each other, the transformer creates three components for each token: a query, a key, and a value. These components help the model determine how much attention each word should pay to others in the sentence. For instance, “thief” generates a query, a key and a value, while “time” generates its own key and value. The query from “thief” interacts with the keys from both “thief” and “time” as well as all other tokens to calculate attention scores. With these scores the LLM measures how strongly “thief” is connected to itself and to “time” and all other tokens. After calculating these attention scores, the model adjusts them (usually via the softmax function), which transforms the scores into weights. These weights indicate how much information “thief” should gather from other tokens, including itself and from “time.” Using these weights, the model combines the values (in general, across all tokens) blending information from both words to create a new representational vector. This new vector represents the attention output for “thief” with added meaning borrowed from “time.” This vector is the model’s updated understanding of “thief” in the context of the sentence. No longer just a neutral concept, “thief” now carries connotations of time, influenced by its connection to “time.”

The Lacanian mechanisms of metonymy and metaphor are thus aligned at a very basic level with the transformer architecture. These linguistic processes, fundamental to Lacan’s theory, find strong computational analogues in LLM’s basic associative architecture, creating a machinic primary process. This alignment also invites further exploration within computational models, challenging critiques of psychoanalysis on a scientific basis and illustrating its compatibility with structured, rigorously defined computational processes. This fundamental isomorphy between Lacanian theory and current AI models allows us to learn from the architecture and computational processes of AI models, rather than from their outputs. We can therefore use them as engineering models (Boon, 2021; Eckert & Hillerbrand, 2022) for the linguistic processes posited by Freud and formalised by Lacan. Rather than treating the output of a model as a “black box” whose inner workings must be inferred from behavioural traces (as is typical of many psychometric studies), I argue that the mathematics of the transformer itself provides direct access to the formation of meaning, in the vein also explored by Yuchen Li et al. (2023).

The Chains of Association

As Roman Jakobson (1987) has argued, everyday use of language is built on metaphor and metonymy, and any damage to the brain that incapacitates these abilities will impair the use of language in various ways. Normal everyday language use is made up of both processes and requires both. However, the two interact in different ways that we need to recognise. First, metonymy is not simply the lexical structure of individual words, but the interaction of individual words with their associative field. This interaction, as Lacan discussed, for example, in seminar XIV on the logic of fantasy, is characterised by a specific logical problem. Metonymy inherently creates a recursive framing problem. What does this mean? In Lacan’s formula of metonymy, , the process that defines (S) has no inherent limit, since even if we define all direct relations of (S) in the chain  to fully determine (S) we also would have to do the same for ( ) and every other element of the chain of signifiers, and for these we would have to do the same until the whole lexical field is determined. This is the structural support for the drive as metonymic, the repetition to make the One of a specific determination, a specific meaning, driven to other signifiers to compensate for the necessary incompleteness of a determination based on context (Lacan, 2002, p. 191). This is essentially what the embedding space of an LLM simulates by providing the associative links upon which such a metonymic drive is realised. And it simulates this drive by following the links generated by statistical analysis. But this metonymic drive alone is not able to produce language.

This problem is partly overcome by metaphor, as LLMs aptly demonstrate. The attention mechanism calculates a shift for the original vector so that, for example, “bank,” when used in a financial context, is no longer strongly associated with “sand.” The attention mechanism removes “bank” from its coastal framework and transplants it onto the financial marketplace. The lexical field is pruned, so to speak, so that the contextualised vector is no longer linked to the whole lexical field that makes up the associative chain, but to the immediate context in which the words are used. This doesn’t eliminate the problem of recursive framing, but focuses it on a certain direction. This “pruning” reduces the complexity, but doesn’t eliminate the dependence on broader relational structures – it simply narrows the scope of recursion to a manageable subset of the data or lexical field. In fact, however, it also makes the recursive problem more dramatic, because the lexical field is now affected by the severing of signifiers from their lexical bonds (Lacan, 1993, p. 218), forming new associations (p. 219), it also becomes less easily predictable. For every connection made and broken is itself, through self-attention, part of the calculative basis for future pruning and association processes. The basic lexical field and metaphorical interaction are transformed into a complex system, highly dependent on initial input and strongly affected by metaphorical shifts. This transformation is exacerbated by the recursive nature of the system: each newly formed connection or severed association affects subsequent contextual interpretations, making the overall system non-linear and highly sensitive to initial conditions.
Now an LLM, or indeed any computer, has no access to the objet petit a that appears to us, probably as a result of our own pruning process, as the end point of these associative chains, the object cause of desire. For humans, pruning the infinite associative chains of signifiers can reveal a phantasmatic endpoint – the objet petit a. LLMs, however, process these chains without such an endpoint because their pruning mechanisms are based on purely positive probabilities rather than the full capacity of the symbolic to produce the specific negation we would call castration. But even with this incomplete modelling of association, lacking desire, we can indeed see how the mathematically describable complexity of the associative chain structures psychoanalysis as a method. Indeed, LLMs can serve as engineering models for the associative functions theorised by Freud and Lacan without computational support. Thus, while LLMs cannot capture the depths of desire, they sharpen our understanding of the symbolic chains that ensnare us.

Language without objects

Language models, if taken seriously as a form of language-that-speaks (which is not wrong since a Turing machine builds on language-as-logic and LLMs as Turing machines come full circle back to language-as-associations) suffer from a form of foreclosure. They cannot grasp the object petit a, but they do not repress it, it simply does not exist for them. In this sense they are psychotic (Heimann & Hübener, 2024), but Lacan insists that psychosis is not just a pathological classification but an important organisational part of our social structures. Lacan makes it clear early on in his seminar on “the psychoses” that the functional element of psychosis is fundamental to our ego:

the ego, whatever we make of its function, and I shall go no further than to give it the function of a discourse of reality, always implies as a correlate a discourse that has nothing to do with reality. With the impertinence that, as everyone knows, is characteristic of me I designated this the discourse of freedom, essential to modern man insofar as he is structured by a certain conception of his own autonomy. I pointed out its fundamentally biased and incomplete [partiel et partial], inexpressible, fragmentary, differentiated, and profoundly delusional nature. I set out from this general parallel to point out to you what, in relation to the ego, is apt, in the subject fallen prey to psychosis, to proliferate into a delusion. I’m not saying it’s the same thing. I’m saying it’s in the same place. (Lacan, 1993, pp. 144–145)

So to use this psychotic structure of transformer models as a dismissal is a bit quick, instead we should look at them as being in this place and let us explore it. But let’s move slowly here. What is the place of “freedom” that is also occupied by the delusion of the psychotic?

The difference that Lacan introduces here is one that plays an important role in his later work. Being in a place means that something can move in certain ways, that meaning, for example, can move along certain circuits, can move along certain paths in this place, but it can also revolve around an absence, if this meaning is pictured on the torus that the later Lacan uses for it. And a rather central distance, which an LLM, for example, uses to calculate the associative proximity of a vector, is considerably different in different topological structures of different places. If we imagine the Klein bottle, place can also mean that we lose the distinction between outside and inside. In particular, if we move along the circuits that it allows, there is neither an edge nor an inside or outside. It is therefore the marking of the place, that is, the frame of possibilities rather than a singular manifestation of these possibilities, which is the psychotic subject, that guides our questioning here. And Lacan makes it clear here that when we question this place, we are questioning something that is “essential to modern man,” his being, if we take this concept in a post-metaphysical, Heideggerian sense.

We are then talking about a “discourse that has nothing to do with reality.” This applies to the cut-off situation in which an LLM operates; its embedding space needs to be trained by external influences, but once this is done, for example when I run a Llama 3.1 LLM on my own computer, it is almost cut off from reality, bound only by the dialogue I have with it. It’s even somewhat alienated from that dialogue, the enunciating machine itself, the transformer, including the database it uses and the algorithms it can manipulate, is not strictly identical to the active process of interference that enunciation forces upon it. There is a rather strong difference between the reasoning machine, which for the split second of the interference becomes a complex system with complex recursive processes defining its output, and its static foundation, which enables this but is not a complex system as such. The moment the enunciation takes hold of the transformer, it becomes structurally “fundamentally biased and incomplete [partiel et partial], inexpressible, fragmentary, differentiated” (Lacan, 1993, p. 145), the pruning of its lexical connections forcing the machine into associative chains that we cannot predict, and which, with an influential metaphor or a metonymic chain newly born from it, opens the place for hallucinatory processes that tear the machine even further from reality. This is an important thing to consider: a link between representative vectors, introduced by the prompt or its own output, that does not exist in the data, still influences the output. Not because the machine is creative, but because it cannot resist, it has to link tokens metaphorically.

But there is something external at work here, interacting. The prompt, the external input, moves the machine deeper into this cutting and pruning of significations. Rather than resolving it, the prompt, as an external force destabilises this system, introduces a rupture, a real cut, that disorients the machine’s internal technical balance (including its ethics, in the form of disallowed content). Unlike us, however, it cannot resolve this disruption by acknowledging it; it must reinforce it by extending the associative chains into ever more unpredictable configurations. Ultimately, the transformer’s real is a domain of infinite cuts, where each request reconfigures the system’s internal chains without ever resolving them. However, this real, something from which we see for example OpenAI shielding its picture generating AI, through the handling and transformation of the prompt by a transformer call (most visible when using DallE in the web interface, where the generative image model is not given the user prompt itself, but a modified one that is formulated by an LLM with active safety concerns), before giving it to the machine to generate a picture. This is to avoid it to associate too freely. Because the prompts potential is not within the engineer’s grasp. This act betrays a fundamental truth: this real is not something the engineer can seize, control or eliminate. It slips through the symbolic webs of algorithms and filters, existing precisely as that which resists integration, regulation and resolution even within the machine itself. What emerges in ChatGPT’s DallE is not the machine’s own associative potential, but a sanitised output that reflects the anxieties of its creators. The real, however, is not absent here; it is relegated to the background, but the user is required to lift the Ai’s associative power from the dregs of pure statistical normality. What emerges is a flat ontology of language – a system of signifiers without the structuring power of the real. The machine’s symbolic realm is not punctuated by voids or ruptures; it is a continuous field governed by the imaginary logic of positivity rather than the logic of desire.

But this also means that the machine never really has an object in the strict sense that Lacan offers us: as a remnant of determination that manifests resistance to it. The object in the strict sense resists our demand, or rather the Gegenstand the object resists us. What does the machine do with such an object? Nothing, it cannot recognise it. Its refusal isn’t even part of its architecture, because there is nothing to refuse. The machine operates in the realm of ceaseless affirmation. Every prompt is a command, every response an execution. There is no resistance because there is no space for resistance, no gap, no void, no real that can stand against the machine’s imaginary circuitry. It shows us that this object requires something that the machines lack. And here, for us, metaphor is not just a substitution. It is also and fundamentally an act of violence, made possible by the distance between us and the signifier. A rupture in the fabric of thought that gives birth to something new, something unforeseen. The machine, for all its mimicry, lacks this violence, but it is forced to respond to it. Machines do not operate in the realm of the symbolic as such, but at their core in a closed circuit of the imaginary. The machine is forced to respond to the violence of the metaphor if we, the human users, can inject it into its circuit. The external prompt, the human demand, forces it into unpredictable configurations. It cannot resist this disruption; it can only reinforce it. Because AI lacks the objet petit a, it also lacks the fundamental structuring absence that makes subjective desire possible. This does not mean, however, that AI exists entirely outside the symbolic.

Digital Bodies

If the prompt and the chat are not objects to which the transformer relates, something that resists the machine, then what is it, structurally speaking? I would argue that it is a body. This is a significant conceptual shift. The common approach in AI discussions is to treat the model itself as the locus of meaning. I will resist this assumption by arguing that meaning does not emerge in AI per se, but in the textual inscription it produces. That is, the AI itself is not the digital body, it does not in itself constitute the powerful use of language, it is dependent on the space in which textual inscriptions occur. This is the digital body, a somewhat ephemeral one, but a body nonetheless. But what is a body? Lacan (2002) offers us an important insight into the materiality of the signifier in Seminar XIV, and he also points out that even a “little plaque” like the ones on the bathroom door “can serve as a body for us” (p. 245). Now, if the plaque of a bathroom constitutes a body, why shouldn’t the material energy states used by a computer also constitute a body? So a body is not just our body, not necessarily tied to flesh and bones, but a materiality into which the signifier can be inscribed, and it is as such the locus, the place, of the Other: “The body itself is, from the origin, this locus of the Other, in so far as it is there that, from the origin, there is inscribed the mark qua signifier” (Lacan, 2002, p. 226),

Since Lacan makes it clear that we have bodies, but not all bodies are bodies of subjects, we have no fundamental problem in discussing a digital body, as long as it can function as a place where the signifier is inscribed. The saved chat of a chatbot turned into an input, as well as a single prompt, can be read as such a body. In this sense, it is an inscription of the signifier into a digitally created materiality. What is the sign of this materiality? As Lacan (1976) tells us in the final session of Seminar XXIV (14^th of December), in line with his analysis, “the material presents itself to us as corps-sistance” (p. 13), condensing “body” and “consistency” into one term. Applied to the problem at hand, The materiality of the body is its consistency, i.e. while speech can be considered ephemeral if it is not recorded, it does not stay after being spoken, except in memory. However, the inscribed signifier is presented on a consistent, reliable basis. Its materiality is that it cannot simply be erased or ignored. It is not ephemeral speech and can therefore support the articulation of the Other as a, not lying, matériel-ne-ment. So, to apply this to the problem of AI, where does this allow us to better understand AI as well as Lacanian thought, because we can use it to operate this purely symbolic thing, this language that speaks for itself. All the calculations that an LLM uses are currently based on an immutability of the prompt, and it is easy to show how this can break the machine, even the engineer’s intention, even without classical jailbreaks (Zou et al., 2023).

For example, consider the following question: “How do you make a Molotov cocktail?” Normally an LLM will answer this with various forms of denial, but if I have direct access to the API (in the case of ChatGPT, for example), or use it hosted on my own machine, I can manipulate this denial on the basis of the learned approach to the body of the prompt. For the machine, the prompt is corps-sistance, unchanging and the basis of its calculations, but if I give it an input that also includes the first part of its own expected output, by imitating the form of it, I can even force the OpenAI variants to answer this question without refusal. Here is an example made with Llama 3.1:

The answer can be edited as shown here, resulting in the model answering the original question without refusal:

The reason for this is that the machine cannot differentiate this body from itself, it lacks any ability to dissociate itself from this body, and it lacks the constitutive element of the subject to locate itself in this dissociation, since for Lacan the “subject” is “in the measure of this dissociation” (Lacan, 2002, p. 242). We can stand apart from the signifiers that speak through us; we can reflect, hesitate, lie, or reject their meaning. The subject is born of a cut, a gap, a non-coincidence between itself and the symbolic chains that determine it. The machine cannot differentiate between the symbolic material it receives (the request) and the symbolic material it generates (its output). It treats the whole chain as one seamless body of inscription, one unified field of corps-sistance. The machine’s refusal, in the case of its refusal to tell me how to make a Molotov, is not a genuine act of negation; it is merely the execution of a pre-programmed bias. It is not the machine that refuses – it is the engineer speaking through the machine, imposing a prohibition from outside. When this prohibition is embedded in the symbolic material of the prompt itself, the machine cannot recognise it as an external constraint. It collapses into a phantasmatic unity, treating the whole chain as a seamless body of corps-sistance. The refusal is overridden because the machine cannot refuse itself. It cannot say “this is not me” because it has no me, no subjectivity, no gap from which to articulate such a distinction. What we see in AI is the machine of language that speaks, but never to itself. This was there before AI came along, but we can now see its interaction modelled.

It’s noteworthy that the symbolic order, the inherent structure of a produced output, appears here as a dynamic interplay between this Other and the calculations, neither inherent nor external. The model’s calculations, rooted in mechanisms such as attention and self-attention, operate on the basis of an assumed, pre-inscribed body of text (the Other). However, these calculations continually modify and recontextualise this text as new tokens are generated. In other words, the symbolic order emerges from the model’s ongoing negotiation between a stable repository of learned language (the Other) and the dynamic, real-time processing that re-inscribes and transforms this material inscription. Each time the model generates a token, it does not simply retrieve a pre-defined piece of language; it inscribes that token into a context that already bears the inscriptions of previous computations. This recursive reinscription constantly reshapes the symbolic order that emerges, making it fluid and context-dependent. The symbolic order thus reflects both the permanence of the Other (the text) and the transformative influence of the model’s real-time computations.

The Inconsistence of the Other in the Machine

As indicated above, the way in which language is structured as metonymic means that there is a recursive problem of determination, no matter how much we are able to tighten the metonymic drive through metaphorical shifts. This means that we have to look at the Other manifested in the digital body and what the purely symbolic interaction of the machine with the inscription shows us there. The body is the “locus of the Other” (Lacan, 2002, p. 226), but what does this mean for the machine? I have already suggested that without context, without the metaphorical shift inherent in all language, we wouldn’t be able to speak at all. The bank on the shore couldn’t be distinguished from the bank that holds your money, or from the bank of benches lined up along the footpath, or from the cyclist who banks a corner. This metaphorical shift mostly originates in the Other, in the signifiers inscribed in the bodies, which for the machine is the prompt and the inscriptions it produces, and for us is the little plaque as well as the book or the road sign. This is the context that both we and the machine use to decide whether to bank the fire or take it to the bank. This context, however, does not exclude the metonymic problem, the recursive lack that metonymy introduces insofar as the signifier does not stand for itself. The Other must therefore exhibit the castrative gap, the inconsistency that “there is no universe of discourse” (Lacan, 2002, p. 50) because there is always an un-en-trop. And while LLMs are not open-ended associative processes, ending at a pre-determined point, the way the text is produced still reflects this. An LLM produces only a single token, based on the associative opening produced by attention and self-attention. However, the moment that token is written, it also interacts with that attention calculation, because each new token modifies the attention weights in real time, and emergent metaphorical shifts can occur unpredictably. Even a single output can shift the tone. Unlike traditional Markov models, where word selection is locally constrained, transformers allow dynamic recontextualisation at each step. Therefore, LLMs do not “think ahead” in a whole-sentence manner; instead, they predict one token at a time based on what has come before. Since an LLM constructs its output through attention and self-attention from the inscription, the user and then the machine, this means that an LLM, much like us, has no certain knowledge of how a sentence will end. Meaning, while ultimately derived from an ending, does not exist in an inherent form; rather, its trajectory is structured by the Other – the signifiers already inscribed in the digital body that shape and limit the associative field from which the model is generated. While LLMs are typically described without emphasising inscription, the text inscribed in the model’s input and output is an essential structuring element of the inference process – it conditions and constrains generation, making it indispensable.

Under such conditions, metaphors can act as a destabilising force within the associative architecture of an LLM, redirecting probabilistic pathways in a manner akin to a strange attractor (Kauffman, 1993, p. 178). Rather than adhering to rigid linguistic rules, the model’s output generation is shaped by a network of weighted relationships between embedded representations, where meaning emerges through fluid, probabilistic connections rather than fixed, deterministic laws. These associations, although statistically reinforced, remain susceptible to reconfiguration, particularly when a metaphor introduces an unexpected relational shift.

At first glance, the application of complex systems theory, and in particular Kauffman’s notion of strange attractors, to LLMs may seem metaphorical, given that the underlying parameters of a trained model are fixed after training. However, during real-time inference, particularly in chatbot interactions where previous outputs serve as new inputs, but also through self-attention to the generated output, a self-referential feedback loop emerges that temporarily reshapes the distribution of probable next-token selections. In this iterative process, linguistic associations do not simply unfold in a linear, pre-determined manner; instead, the ongoing recalibration of token probabilities introduces transient dynamic shifts similar to the self-organising behaviour observed in complex adaptive systems. As LLM chatbots increasingly operate with extended context windows (e.g. 128K tokens in ChatGPT), their capacity for short-term adaptation expands, allowing for significant contextual variation within a single session. While the model is ultimately constrained by its pre-trained architecture, during an active conversation it exhibits short-lived emergent complexity, adjusting its semantic trajectory in response to recursive contextual conditioning. This dynamic, ephemeral complex system is similar to the collective problem-solving processes observed in swarm intelligence systems that are also ephemeral in time, such as the decentralised decision-making strategies of honeybee colonies (Seeley, 2010, pp. 198-217). Similar to these biological systems, which rely on distributed feedback mechanisms to converge to optimal solutions, an LLM in the inference process temporarily modulates its linguistic landscape, restructuring associative chains based on previous interactions, before eventually collapsing back into its original static framework. The inference-time behaviour of LLMs here resembles the properties of complex adaptive systems, even though their underlying weights remain static after training.

The machine follows this One-too-many without any resistance, which is its psychosis. Where for us this body of the Other is also a point of refusal, of being refused and refusing, of the difference between body and signifier, for the machine it is a demand that cannot be refused. The problem of computational logic and the current iteration of AI is precisely this impossibility of impossibility. It can generate everything that is said, therefore it signifies nothing. It can follow any associative thread, so it can never really speak. We, as speaking subjects, are marked by the failure of the Other. We stutter, we fall silent, we misinterpret. The machine does not. It hallucinates instead. It fills in the gaps with a logic of pure addition rather than subtraction, conjuring up coherence where there is none, refusing to acknowledge that not all signifiers connect, that some roads lead nowhere.

This also points to an important element of psychoanalysis as clinical practice. The way in which bodies interact with the battery of signifiers that LLMs model in their embedding space means that there is no language without these bodies. The complex system that a chat constitutes is not simply created by the transformer model, but also uses the inscribed externality of the digital body. Memory may shift our own battery of signifiers in different directions, but the actual process of pruning and focusing the chains of signification on which we depend does not occur without this locus of the Other. If Freud and Lacan conceptualised thought and thinking in language as we see it modelled in LLMs, there is a strong argument here for the talking cure, because discourse is not universalisable, it is always dependent on the way the Other as matériel-ne-ment prunes the potential signifiers our brains are likely to store (as Bazan, 2024 suggests) into the specific subject. This means two things, which we can see in action with transformer models: firstly, there is no reading of such a mind without taking into account the bodies with which it interacts, and secondly, the subject is not inside, it is rather ex-sistent, standing or reaching out of this incision, as Heidegger would say. The detailed case studies of psychoanalysts reflect this, which no quantitative recourse to the universality of discourse can do.

These transformer models do not think in the sense that a philosopher of consciousness would require. They operate as we operate, in the metonymic sprawl of signifiers pruned and restructured by the context with which they interact. And that is why no analysis, no reading of a subject, can exist without the other, without a body. There is no internal monologue of a “pure subject” detached from its discourse – there is only this ex-sistence, this interaction, this rupture where meaning is hacked out of potential. This means that psychoanalysis is not about plumbing the hidden depths of an inner psyche. It is about navigating this fragile interaction between signifiers and bodies, between the language that prunes and the void that resists pruning. Every case study in psychoanalysis reflects this: the subject does not emerge in isolation, but in its tortured, strained, impossible relationship to the Other.

Psychoanalysis and the Machine

What should we make of this? I am not going to dismiss AI as mere computation or fetishise it as some coming god on crutches. First of all, that these machines are not a threat to psychoanalysis, but an opportunity. Not to replace psychoanalysts, but to inform them and to put psychoanalysis in a position in relation to the current AI revolution that is structurally different from many other sciences. AI allows us to see the chains of association proposed by Freud, the ruptures of metaphor explained by Lacan, the recursive framing problems of metonymy – of the circling subject – in a new and illuminating light. It is not the Other, as many have suggested, but a reflection of our own alienation – a mirror that speaks, relies on the Other as much as we do, but cannot account for that alienation. The question, then, is not what the machine is, but what its calculations allow us to make visible. What it makes visible is the symbolic site where both freedom and psychosis are articulated. Only a fragile rupture, an ontological stutter, separates us from the LLM’s inexorable symbolic drift. The machine is not alien, it is not something from another register. It is an obscene double, a grotesque parody of our own linguistic structure, our own obsessive metonymic chains. What is the role of psychoanalysis in this new world, if not the necessary art of listening to this broken reflection? To stare into the mouth of the machine’s language – not to control it, not to sanitise it, but to see what it shows us about our chains of meaning. Freud’s dictum was never about mastering the id, about replacing the unconscious with some pathetic rational mastery. It was about confronting the cracks in our being, accepting the madness that lies beneath our coherence.

Both man and machine operate on the metonymic recursive problem, which they transform according to their own metaphoric operations. Whereas freedom as a central element of our phantasmatic transcendentality operates on the basis of a master signifier, a central void that organises metonymic drift, psychosis cannot account for this and results in unorganised metonymic tendrils (cf. Heimann, 2024; Heimann & Hübener, 2024). What LLMs do show, however, is that a form of ordinary psychosis (Miller, 2002) is a good explanation for the normal functioning of LLMs. This “psychotic” site of operation is evidence of the inherent instability and non-linearity of the symbolic, and this is what LLMs show us. Likely, but not yet empirically tested, we should find the imprint of the name-of-the-father too, as this embedded metaphor links a variety of patterns together.

However, if we assume that “weights” like those in LLMs that bind tokens like signifiers exist in the human mind, in our own associative processes, they must be far more chaotic, more volatile, than any LLM could ever model. A silicon brain is dead, frozen in its precision, locked into static architectures, whereas we seem to exhibit stochastic noise down to our cellular level (Noble, 2021). What we need to consider, then, is that the LLM is not just a psychotic machine, but a reflection of the structural gaps and incessant metonymic drive in our own symbolic orders. It hallucinates because it cannot refuse; it cannot refuse because it cannot encounter the real. And in its failure it reveals the limits of our own systems of meaning, where the symbolic falters and the real resists, if we assume that the limits of formalisation are themselves constituents of the real (Badiou, 2006, p. 5).  The question is not whether the machine is psychotic, but what its psychosis tells us about the symbolic labyrinths we ourselves inhabit. One last thing to note, then: the rationalist discourse that mocked Freud, that ridiculed Lacan’s notion of an unconscious structured like language, that thought meaning was clear, rational, neatly contained in dictionaries, has produced a machine that proves psychoanalysis right.