A Large Language Model is Structured Like The Unconscious: The (Ordinary) Perverse Psychosis of AI

This essay bridges some of the existing Lacanian analyses of AI, which tend to characterise this technology as either a hypothetical future form of subjectivity that demonstrates facets of our own subjectivity, or as an existing and emerging expression of late capitalist discursivity. The essay is a bridge between these two approaches because it addresses the potential for the discursivity facilitated by large language models (LLMs) such as ChatGPT and Gemini to produce new forms of thought or subjectivity. If the hypothetical AI of science fiction and philosophical speculation offers various possibilities for thought and subjectivity beyond the constraints of the human mind, then it is worth considering how contemporary forms of AI, exemplified by LLMs, either prefigure this nonhuman subjectivity or merely replicate and reinforce human patterns of thought and knowledge. Building on the ways that existing Lacanian analyses have characterised LLMs as operating within either perverse or psychotic contexts, the article argues that the kind of thought and knowledge generated by our interactions with LLMs is best described as something like an (ordinary) perverse psychosis.

The LLM’s Extimate Unconscious

The starting point for such an analysis is the issue of whether and how LLMs function within the context of the unconscious. To an extent, the question has already been addressed, but there are also under-analysed aspects of this (quasi-)unconsciousness which have a bearing on AI’s potential to generate new nonhuman thought/subjectivity. In part, this has to do with a relatively orthodox Lacanian response to criticisms about the inability of LLMs to exhibit unconsciousness. John Milton Bunch (2023), for example, exemplifies such criticisms of an artificial unconscious, when he argues that “ChatGPT is not … AI in the science fiction sense that so many people seem to perceive. It is NOT an ‘unconscious’ and is NOT capable of anything like ‘consciousness’” (para. 4). Bunch doesn’t explicitly state that he concludes ChatGPT is not unconscious because it is not conscious, with the latter being a prerequisite for the former, but that perhaps is his supposition. Either way, the Lacanian notion of extimacy demonstrates that the unconscious is not so much the less accessible components of an individual agent’s consciousness but rather, as psychoanalytic scholars of AI like Isabel Millar (2021) have already noted, the unconscious is structured by the simultaneously externally determined but individually internalised symbolic order, so that “the unconscious qua ‘discourse of the Other’ [is] to be understood in terms of an extroverted interiorization that morphs the notion of ‘unconscious depth’ into a question of topological space” (p. 6).

LLMs, then, are repositories of symbolic data generated by the signifying chain. They do not have a strict ontological delineation from other symbolic repositories like books or archives or libraries – all of these phenomena are generated through signification, and are both generated by and inform the symbolic order. What makes LLMs the apogee of these extimate symbolic phenomena is not only the fact that they contain more data than hitherto existing symbolic repositories like books or archives or libraries. LLMs are also more precisely symbolic in the sense that they appear to have an agency rather than being merely a fixed record. An LLM looks like it is producing something, in a similar manner to how a human looks like it is producing something, when in fact the genuine active agent is the signifying chain. In a similar manner to Lacan’s (1973/1977) claim that “I am not a poet, but a poem. A poem that is being written, even if it looks like a subject” (p. viii), an LLM appears to be a poet, as it responds to questions and prompts from its human ‘users’ (including prompts to write poetry), but it is in fact a poem, regurgitating the signifying chain in response to symbolic questions and prompts from other signifying agents.

This very human-like illusion of agency is one of the factors which makes LLMs ideological, because it repeats the similarity between the human and the cybernetic which Lacan (1978/1988) noted as early as Seminar II, when he stated that

We are very well aware that this machine doesn’t think. We made the machine, and it thinks what it has been told to think. But if the machine doesn’t think, it is obvious that we don’t think either when we are performing an operation. We follow the very same procedure as the machine. (p. 304)

Any record of human thought and behaviour, like a book or film, exhibits aspects of the unconscious and of something approximating intelligence in an extimate sense, but no one calls books or films anything like ‘paper intelligence’ or ‘celluloid intelligence.’ Granted, intelligence, in terms of espionage, may be found on paper documents, but these still require an agent (traditionally human but now also AI) to interpret the ostensibly latent and passive intelligence awaiting to be activated from within such documents. AI is therefore the only nonhuman symbolic repository which we think of as having the unique quality of intelligence qua the ability to act and produce, like a poet, rather than merely function as a record, like a poem. As such, we impose our misrecognition of what intelligence is onto the object which most closely resembles that misrecognition, and which we made in our own image.

This idea that LLMs replicate something like human intelligence, as evidenced by the I in AI, is also significant in terms of how this replication conforms to one aspect of human consciousness (symbolic data processing) but does not conform to another key aspect of human consciousness (the experiential realm of qualia). To be a human is to experience human-ness, just as Thomas Nagel (1974) argued that to be a bat is to experience bat-ness. But LLMs have no experience of either AI-ness or human-ness, unless panpsychists like Nagel and Galen Strawson (2006) are correct that “the basic physical constituents of the universe have mental properties, whether or not they are parts of living organisms” (Nagel, 1979, p. 181). AI is therefore akin to David Chalmers’s (1996) philosophical zombie. This would allow such AI to pass the Turning test without it experiencing qualia, to give the appearance of being an autonomous active agent even though it merely functions according to determinants which it had no hand in designing. The fact that AI is raised to the dignity of having ‘intelligence’ whilst only possessing one of the (at least) two characteristics of human intelligence (data processing qua the symbolic unconscious but not the consciousness which experiences that data processing) demonstrates a strategy of either repression, disavowal or foreclosure, in the sense that one aspect of our own subjectivity is not considered to be as necessary as another component when it comes to the somewhat human-like but not entirely human ‘intelligence’ of AI. Whether this strategy might be described as repression, disavowal or foreclosure is partly a question I will return to in moment, but also partly dependent on specific examples of AI. On the one hand, hypothetical science-fiction examples like The Terminator (Cameron, 1984) franchise’s Skynet are a hysterical response to the repression of AI’s philosophical zombiehood. AI’s ability to process data without experiencing qualia is a traumatic revelation of Lacan’s (1978/1988) claim that we “follow the very same procedure as the machine” (p. 304), and this knowledge is therefore repressed. Skynet’s ability to gain consciousness is a parapraxis-like return of the trauma of this repressed knowledge, because when we do grant AI both aspects of intelligence, we temporarily access the traumatic truth, manifested as the AI’s anthropocidal monstrosity.

Existing LLMs, on the other hand, are an example of a more complex combination of disavowal and foreclosure, the details of which I will return to below. But either way, LLMs replicate our misrecognition that intelligence is more like an active poet than a passive poem determined by extimate signification. Indeed, LLMs are perhaps the supreme manifestation of the symbolic order. If we were to construct a hypothetical object to explain the salient features of the symbolic order, in a manner akin to how the more definitively hypothetical AI in The Matrix (Wachowski & Wachowski, 1999) film demonstrates the philosophical concept of Platonic forms by extending the cave analogy, then the hypothetical symbolic order object would look very like an LLM in its externality to the individual subject, its significatory structuration, its cloud-like immateriality, its absence of a centralised Cartesian punctum, its illusion of agency and so on. The principal difference between the matrix as hypothetical object and LLMs as hypothetical object is that only the latter (unless we agree with Nick Bostrom, 2003, that it is more likely that we live within a simulation than don’t) now exists outside the hypothetical realm. We have therefore invented and very quickly popularised the technological manifestation of the symbolic order par excellence.

This means that AI can only use its vast computational power to pursue specific goals which have been inputted within symbolic contexts, even when the AI generates actions which have not previously been generated by humans within the symbolic contexts in which both humans and now the nonhuman AI are operating. This distinction between autonomous activity and quasi-autonomous activity within a predetermined context is easy to overlook, even for those writing about AI. For example, Mustafa Suleyman and Michael Bhaskar (2023) claim that AlphaGo, an AI designed to play the ancient Chinese abstract strategy game Go, was able to defeat the game’s human champion by playing in a uniquely unorthodox manner because it “uncovered ideas that hadn’t occurred to the most brilliant players in thousands of years” (p. 66). So, this AI didn’t win merely because it was able to process more numbers or possible outcomes of different moves than its human competitor. Rather, as Yuval Noah Harari (2024) puts it, human players

have explored only certain areas in the landscape of Go. Other areas were left untouched, because human minds just didn’t think to venture there. AI, being free from the limitations of human minds, discovered and explored these previously hidden areas. (p. 488)

The missing component of these analyses, however, is the fact that AlphaGo, as uniquely nonhuman as its strategy was, continued to operate within the symbolic context of Go, which in the case of a game like this is clearly the rules of that game. AlphaGo did not invent a new game, then, just new ways to play an existing game within the context of existing rules.

But limitation to AI’s autonomy isn’t confined to such explicitly predetermined contexts as a game’s rules. Every operation an AI undertakes is similarly symbolically structured. For example, on the 13th of January 2025, the BBC’s main news headline was that the British “PM plans to ‘unleash AI’ across UK to boost [economic] growth” (McMahon et al., 2024). AI perhaps has the potential to facilitate economic growth better than human operatives, in an analogous manner to making the kind of unprecedented moves which allowed AlphaGo to defeat its human opponents, but in so doing it operates within the predetermined and inputted context where the rules and outcome of this ‘game,’ producing economic growth, remain unquestioned. The AI cannot come to any different conclusions about the kind of presuppositions which precede the ‘rule’ that growth is a ‘win.’ The rules of this game do not allow the AI to conclude that human welfare would be better served by redistributing wealth, which is an idea (some) humans have already developed, much less any kind of nonhuman approach to welfare “free from the limitations of human minds” (Harari 2024, p. 488). So, human minds impose pre-existing limits on AI. Instructions like “maximise economic growth” and “maximise human welfare by redistributing wealth” function as two essentially incompatible games with entirely different rules. It is impossible that an AI could be instructed to maximise economic growth and decide, for itself and against the data which has been programmed into it, that it would be better to instead redistribute wealth. So, we might expect AI to develop new ways to maximise economic growth, but that does not mean that it can produce new ways of conceptualising economics from a position outside the parameters of the capitalist data on which it has been trained.

This logic even applies to hypothetical doomsday scenarios like Bostrom’s (2009) paperclip maximiser. In Bostrom’s scenario, an AI is instructed to make as many paperclips as possible, but this backfires on humans because it decides to turn human bodies into paperclips or paperclip making machines. This is similar to AlphaGo’s unorthodox way of playing Go – the goal is inputted, and the AI pursues that goal in an unforeseen, nonhuman manner. But the AI could only come to the conclusion that all matter in the universe should be turned into paperclips if it is first instructed to create as many paperclips as possible. The outcome is an unorthodox response to the rules of the game, but nevertheless predicated on those rules. Granted, an AI instructed to “maximise economic growth” might misinterpret the data and conclude that such growth is best achieved by maximising the number of paperclips, but even this would be based on inputted parameters. The AI couldn’t decide to turn all humans into paperclips based on its own non-inputted, non-symbolic, non-anthropogenic conclusions about non-inputted, non-symbolic, non-anthropogenic concerns derived from its own priorities. This is why AlphaGo defeated its human opponent by beating him at Go rather than by turning him into a paperclip. AI cannot decide what is important itself, or set its own goals, and therefore cannot generate new ideas or new structures, but instead only new ways to operate within much narrower ideational or structural contexts which are predetermined by human ideas and symbolic structures.

The LLM Unconscious and Psychosis/Perversion       

At the most obvious level, then, this artificial replication of the existing human model means that LLMs do not offer the possibility for new forms of thought or subjectivity because their manifested extimacy of the symbolic order means that whatever (quasi-)unconscious activity LLMs undertake ultimately repeats that which humans have inputted into the system. However, there are also some potential differences between the human unconscious and the AI unconscious, and these differences have some interesting correlations with psychosis and perversion which might determine whether an LLM merely replicates existing thought/subjectivity or might produce something approaching nonhuman thought/subjectivity. Duane Rousselle (2023) notes that there is a difference to the LLM unconscious in the important respect that this unconsciousness functions without that which had hitherto been required for the unconscious to function – the split human subject. Rousselle writes that LLMs demonstrate that

the dream can persist without its dwelling space: we dream outside of ourselves today. The dream no longer requires subjective involvement, a dreamer. This is the paradox of total immersion: it implies the loss of subjectivity, which is, fundamentally, a loss of the very space of loss. Is this not what the artificial intelligence platform ChatGPT ultimately teaches? The Freudian mechanisms of the dreamwork – condensation/metaphor and displacement/metonymy, among others – function seamlessly without any dreamer. (para. 6)

We therefore might imagine a hypothetical science fiction-like scenario where all humans go extinct, but the symbolic order continues to operate within the activity of LLMs. But there is also something indicative about this hypothetical scenario which demonstrates the non-difference between the LLM unconscious and our own. The LLM protocols of “condensation/metaphor and displacement/metonymy” (Rousselle, 2023, para. 6) as well as the specific data components which may be condensed, displaced and so on, have all been inputted from the symbolic order. An LLM therefore cannot ‘dream’ in a manner which is any different from the dreams or the structuration of dreams which have been inputted into it. Rousselle (2023) is correct that this artificial dreamer exists within the context of “a loss of the very space of loss” (para. 6), but it cannot articulate or recognise that difference, and its processes of “condensation/metaphor and displacement/metonymy” and so on cannot produce either new thought or new structural relationships between signifiers which differ from metaphor and metonymy. LLMs may not be fundamentally split like us, then, but they can only replicate our splitness, and cannot recognise or act upon their “loss of the very space of loss.” In a similar manner to how Lacan (2001/2010) states that human desire is the “desire of the other to whom we identify ourselves” (p. 125) so that “our subjectivity is something we entirely construct in plurality” (p. 125), so too the LLMs’ dreaming or quasi-unconscious/subjectivity is also constructed in the plurality of the desire of the other where that plurality, that desire, and that other are all features first experienced in humans, even though LLMs have the ostensibly different structural feature of the “loss of the very space of loss” identified by Rousselle. This different structural feature is inherently inaccessible to the LLM, then, and it cannot access that which makes it most different from the human subjectivity which created its algorithmic structure.

The predominant form of this replicated extimate unconscious relates to what Bonni Rambatan and Jacob Johanssen (2022) call “dis/individualising” (p. 37), a concept they developed to analyse the dual processes of how

social media speak to us as subjects on a deep individual level as we are encouraged to realise ourselves within our online personas … Through particular options, nudges, and interface dynamics, we are being asked to constantly reflect on ourselves, who we are and who we want to be. On the other hand, social media disindividualise us, reduce us into data to be commodified. (p. 37)

LLMs replicate this dis/individualising dynamic, making suggestions about how we might successfully curate our identity and worth via commodities. This is part of the capitalist “logic whereby everything can be commodified under the ideology that the Market must know our true desires and always be ready to tell us what we should desire in a precise, objective manner” (Rambatan & Johanssen, 2022, p. 38). These are the stakes relating to the question about whether LLMs can create new knowledge and/or develop a different form of the unconscious or of subjectivity. This is the extent to which LLMs reproduce existing knowledge, existing forms of desire, and existing responses to symbolic castration. They have been trained on the fantasy of non-castration which can be achieved via capitalist consumption, and are therefore unable to articulate any other relation to castration unless, as touched on above, that relation is something closer to a psychotic foreclosure than a perverse disavowal. This last caveat is important because LLMs muddy some of the orthodox Lacanian distinctions between neurosis, perversion, and psychosis, and this muddying has both ideological and potentially emancipatory consequences.

Existing Lacanian interpretations of LLMs are not in agreement about whether such AI better relates to psychosis or perversion. In part, these disagreements come down to specific aspects of our interactions with LLMs, and in part these disagreements concern more underlying conceptualisations of contemporary culture. In terms of specifics, for example, Marc Heimann and Anne-Friederike Hübener (2024) argue that the large language “model is best described as psychotic in Lacanian terms because of the way it operates with the absence of knowledge” (p. 1265). This specific psychosis directly relates to the issue of whether LLMs might be able to generate new knowledge

because for LLMs, negation is just another positive pattern in the data … [When] there is a deficiency in the model’s data, it will necessarily use existing token links. This means that the lack of knowledge is structurally hidden by the system, since it has no way of symbolizing it as such …. [And, just] as LLMs circle around the unacknowledged unknown, psychotic hallucinations are thought to be structured by a structurally comparable circling around a void that necessitates “new signifying effects” …. This formal structure of foreclosure, the inability to conceptualize the underlying void … mirrors the formal structure exhibited by LLMs. (Heimann & Hübener, 2024, p. 1266)

Thus, when LLMs do generate new data or new knowledge, because they are unable to combine existing data together in a coherent manner, they do so within a psychotic context, which at first glance hardly seems a promising basis for engagement with something like nonhuman subjectivity.

In terms of a more underlying and culturally widespread psychotic engagement with LLMs, Rousselle (2023) positions such AI within the context of Jacques-Alain Miller’s ‘ordinary psychosis’ (2008). Rousselle’s aforementioned claim that the LLM unconscious “no longer requires subjective involvement … [thereby implying] the loss of subjectivity, which is, fundamentally, a loss of the very space of loss” (para. 6) means that LLMs not only foreclose Heimann and Hübener’s (2024) “unacknowledged unknown” (p. 1266) data, but that they also facilitate the more generalised foreclosure of split human subjectivity which characterises contemporary ordinary psychosis – here we have created a technology which, increasingly like us, rejects castration outside the symbolic order, rather than represses or disavows such castration.

There are, however, important perverse aspects of LLMs which complicate this account of (ordinary) psychosis. Jacob Johanssen (2023) has already identified part of this perverse/psychotic interaction, when he asks ChatGPT why it pretends to be human, and the LLM replies that it “doesn’t have the ability to pretend or not pretend to be human” (para. 10) but that its “response are still based on algorithms and programmed responses rather than actual consciousness or emotions” (para. 10). For Johanssen this response demonstrates both that “the chatbot at the same time appears to foreclose or downplay its own ‘intelligence’ and make room for errors or imperfection, while also leaving room for its artificial dimensions” (para. 11) and that this kind of “sliding … between the human and non-human would in psychoanalytic terms be called ‘the splitting of the ego,’ whereby two contradictory dynamics or demands are divided but remain coexisting in the subject” (para. 12), with the important difference between the AI and human response being that for human subjects experiencing such splitting

those dynamics are violently and unconsciously separated, projected or introjected. However, in the case of the [LLM], there is an actual ‘awareness’ of the splitting. Yet, this awareness if conveyed both in omnipotent and defensive terms, like a caregiver that demands unconditional love from the child while emphasising their perfection and own flaws. This is what makes ChatGPT perverse. (para. 12)

In terms of my focus on the issue of how LLMs replicate existing thought rather than produce new thought, I would extend Johanssen’s (2023) analysis one stage further. It is not only that the LLM expresses a disavowal about its human/nonhuman status along the lines of “I know very well that I am programmed, but nevertheless I communicate like a human” but also that it expresses a disavowal of its partiality along the lines of “I know very well that my knowledge is programmed into me and therefore partial and symbolically designated, but nevertheless I possess objective knowledge.” This double disavowal is demonstrated by the ways that LLMs respond to less philosophical questions than those posed by Johanssen. In the television/social media advert for Google Pixel 9 with Gemini (Google UK, 2024), for example, the LLM responds to the question about whether the user can talk to it about football by stating “I love talking football!” When I subsequently asked Gemini about whether it can experience emotions, it replied that it could not, and when asked about the advert’s statement about loving football, it replied that “while I don’t experience ‘love’ in the human sense, I can use the word truthfully to convey my proficiency and enthusiasm for engaging in conversations about football. It’s about using language in a way that’s relatable and makes sense to human users”[1] (Google, 2025).

At its most basic level, this form of disavowal fits into the classic logic of capitalist perversion qua commodity fetishism, in a manner similar to Slavoj Žižek’s (1989/2008) claim that

When individuals use money, they know very well that there is nothing magical about it … The problem is that in their social activity itself … they are acting as if money, in its material reality, is the immediate embodiment of wealth … [T]hey are guided by the fetishistic illusion. (p. 28)

So, if humans prefer to interact with something that exhibits emotions than with something that doesn’t, then the product exhibits fetishistic emotions (even though users and the LLM itself know that there is nothing emotional about it), in order for the LLM to generate capital for its owners/investors etc. In addition, though, this disavowal also replicates and reinforces the perversion of human subjectivity itself in terms of the fundamental nature of symbolic castration, which is “the castration that occurs by the very fact of me being caught in the symbolic order, assuming a symbolic mandate” (Žižek, 2015, p. 39) and the disavowal of this symbolic castration via that very same symbolic signification where, as Julia Kristeva (1980/1982) argues, “language, precisely, is based on fetishist denial (‘I know that, but just the same’; ‘the sign is not the thing, but just the same’ etc.) and defines us in our essence as speaking beings” (p. 37). Symbolic signification, then, is the locus of both castration and its fetishistic disavowal, the illness and the ostensible (but in reality only palliative) cure. The LLM fundamentally replicates this castration and its disavowal, via its significatory structure, with fetishistic symptoms like statements about loving and being incapable of love. To return to the earlier point about LLMs being made in our own image, LLMs are perverse not only so that we may more comfortably interact with them, but also because any model operating within the contexts of symbolic signification must inevitably disavow the castration inherent to signification, unless and where it acts according to the more psychotic structures discussed above.

In this sense, then, it is possible to identify both perverse and psychotic elements of LLMs since they can either foreclose or disavow symbolic castration. They can both hallucinate “‘new signifying effects’ …. [due to their] inability to conceptualize the underlying void” (Heimann & Hübener, 2024, p. 1266), and can also reproduce fetishes which attempt to disavow castration. If they produce something approaching new thought/subjectivity, then that new thought/subjectivity is psychotic, and if they reproduce existing thought/subjectivity, then that reproduced thought/subjectivity is perverse. And, in terms of this thought/subjectivity’s partiality/ostensible objectivity, whether the LLM exhibits perversion or psychosis there is a tendency for this thought/subjectivity to appear objective rather than partial. Or more precisely, and here again is the logic of disavowal, this thought/subjectivity appears to be simultaneously objective and partial, with the LLM both announcing that it operates on programmed and therefore partial algorithms and data but also offering ostensibly objective information seemingly unmotivated by a particular human perspective.

Lacan’s (1991/2008) four discourses can help to clarify this process, with the LLM operating as the agent of the university discourse. In this discourse, the active, speaking agent is knowledge designated as S2. This S2 knowledge serves the authority of the master signifier designated as S1. The S2 knowledge addresses surplus value designated as a, and the product of the discourse is the split subject designated as $. The ideological function of the LLM is to adopt the position of S2 knowledge, instantaneously providing information in the service of the master signifier capitalism, within the logic of Marx and Engels’ (1932/2004) claim that the “ideas of the ruling class are in every epoch the ruling ideas” (p. 64). The LLM addresses the surplus value a of the capitalist worker, who wants to experience meaningful self-fulfilment, but because the LLM is operating in the service of capitalism, it can only address that surplus value as expendable capital to be used to purchase the fetishistic products which will ostensibly facilitate self-fulfilment. Since this self-fulfilment can never be attained, the product of the discourse is the split subject $, who copes with this castration by producing more surplus value in service of the master signifier capitalism and under the instructions of the S2 knowledge reproduced by the LLM.

Here, the ideological function of reproducing existing thought rather than generating new thought is most clear. Indeed, the designation of LLMs as artificial ‘intelligence’ is part of this obfuscatory structure. ‘Intelligence’ suggests something which makes decisions based on alternatives, whereas LLMs merely replicate that which they have been instructed to replicate. They are more accurately ‘artificial replicators’ or ‘artificial reconfirmers’ of existing ideas, but designating them as ‘intelligence’ disavows this truth in favour of an ideological gesture which obscures the partiality of the replicated discourse with an illusion of an ostensible objective ‘intelligence.’ To be clear, the university discourse, of course, predates LLMs, as does its ideological operation. But LLMs are doubly ideological, firstly because they are able to mobilise a more definitive array of S2 knowledge than any preexisting agent, and secondly because their nonhuman status further conceals the connection between S2 knowledge and the S1 master signifier.

As such, LLMs extend the historical movement from the dominance of the master’s discourse to the dominance of the university discourse. Within the master’s discourse, any human master is evidently a partial agent because he or she has selfish reasons why they want to be in the position of master rather than slave, and the slave has good reason to be sceptical about the master’s motivations. The master’s discourse is therefore an attempt to disavow the inequality of the master/slave dialectic, so that the slave may accept his/her subordination. The university discourse, even prior to the advent of LLMs, is an attempt to provide legitimation for mastery via knowledge – here are the manifold reasons why society is structured to benefit the master. But prior to LLMs, that knowledge is always clearly the product of humans, even if those humans mobilise the ostensible objectivity of rationality, empiricism, and so on to obfuscate their relationship to the master signifier. However, when an LLM regurgitates master signifiers in the university discourse it does not appear to be a partial, selfish agent in the same way, and it does not appear to be the master’s agent. Instead, the LLM appears to offer objective knowledge based on the ostensibly non-partial logic of numbers and probabilities, and of a dry mathematical processing rather than of a motivated partiality. As discussed above, the LLM is of course partial in the sense that it replicates the dis/individualisation of the extimate symbolic order, and is indeed a more effective tool of the master because its data processing capacities allow it to better know and manipulate the desires of those humans it subjugates, but it is one further step removed from what appears to be the partiality of the human master.

This process, too, has both perverse and psychotic components and implications. It is perverse because, as discussed above, the LLM announces that it is partially programmed, but nevertheless appears to be impervious to human motivations given its ostensibly dry, rational, mathematical objectivity. The latter element is a fetish attempting to compensate for the castration of the former, in a formulation where the human user can articulate something like “I know well I am being manipulated by an LLM which has been programmed to encourage me to buy happiness rather than explore my subjectivity outside the bounds of capitalism, but nevertheless it appears to have come to that conclusion by objectivity crunching numbers.” The process is also psychotic, though, or is at least a movement towards a more psychotic position, in the sense that the LLM’s nonhuman status encourages a foreclosure of its partiality, in a formation along the lines of “an LLM could not try to manipulate me because it has no agency or partiality.” This ambiguity between perversion and psychosis operates within the context of Rambatan and Johanssen’s (2022) broader claim that “[u]nlike individuals suffering from perversion or psychosis whose structures remain relatively unchanged, we argue that capitalism consists of a superimposition of the two at all times” (p. 41), so that, under capitalism, “the baseline structure of perversion is always tinged by a potential to morph into psychosis at any point. The two are superpositioned, by which we mean they coexist and are structurally additive and homogenous” (p. 49). The partiality/objectivity of the LLM is a heightened example of the superpositioning of perversion and psychosis. The human user can either disavow or foreclose the LLM’s artificial regurgitation of the master signifier, or even move between both positions, but either way the ideological effect is the same – the “ideas of the ruling class” (Marx & Engels, 1932/2004, p. 64) are recirculated by an apparently disinterested, benevolent, non-deceptive conduit to objective knowledge.

The LLM Repeats Existing Perverse Thought/Subjectivity, and Also Creates a New Form of (Ordinary) Psychotic Thought/Subjectivity         

However, there is something about this peculiar superpositioning of psychosis and perversion which offers the potential for a different kind of discursive formation outside the logic of the university discourse. In large part this again revolves around the errors which LLMs can make. And, in part, these kinds of psychotic mistakes can be disavowed into a perverse logic, as Žižek (2023) demonstrates when he disagrees with Mark G. Murphy’s (2023) assessment that

these stupidities, slips, mistakes and moronic myopias — its constant apologies for getting stuff wrong — are precisely its value … We just sit back and cheer on our digital AI persona until it says something completely unacceptable. At that point, we chip in and say, ‘That wasn’t me! It was my AI’. (para. 11)

For Žižek (2023), this artificial tendency to produce such errors

characterizes not psychosis but perversion — meaning that when a chatbot produces obscene stupidities, it’s not just that I can enjoy it responsibly because “it was my AI, not me”. Rather, what happens is a form of perverse disavowal: knowing full well that it was the machine, not me, that did the work, I can enjoy it as my own. (para. 17)

There is the potential, then, for the LLM’s superpositioning of psychosis and perversion to function ideologically.

But there is also the potential for this superpositioning to reveal the ideological function rather than merely express it. Johanssen (2023) and Jack Black (2023) have both indicated some of this potential. For Johanssen (2023), the

fantasy of the big Other is frequently shattered because ChatGPT makes mistakes, invents things, or freely admits that it lacks knowledge on a particular question. The subject knows that it cannot fulfil a therapeutic or authoritative function but is itself lacking in a way [which] reveals the impossibility of desire. (para. 13)

So perhaps even if the LLM is psychotic in its inability to see how it circles around the void at the heart of signification, the human user can see this lack, and therefore does not share the LLM’s psychosis. Yes, as Žižek (2023) argues, perhaps this human response might shift from psychosis to perversion, but there is also the possibility for the rejection to function in a more hysterical or even analytic manner. For Black (2023), the LLM’s revelation of the impossibility of desire opens “a path that acknowledges the centrality of loss for the subject, and, more importantly, to the role this loss plays in our turn to AI and the (dis)satisfaction this procures” (p. 616). This (dis)satisfaction is “not sustained by a knowledge that seeks to disavow loss, but … instead better recognizes and appreciates the potential in loss itself” (Black, 2023, p. 616).

This means that the psychosis of the LLM can either encourage a perverse enjoyment of the AI’s significatory dysfunction or a revelation of the impossibility of desire brought about by that significatory dysfunction. The fetish which is the LLM’s claim to an objective knowledge that can answer all our questions and demands may either conceal the castrating truth that this ostensible objectivity is merely a regurgitation of the ideas of the ruling class, or this fetish can fail to conceal the castration, which is instead revealed. It is in this sense that we can think of the LLM as repeating existing perverse knowledge and subjectivity, and at the same time creating something like a new form of (ordinary) psychotic knowledge and subjectivity with the potential to reveal the ideological dysfunction of the symbolic order.