Abstract
This article explores psychoanalytic perspectives, particularly Lacanian theory, on artificial intelligence, with a focus on chatbots, such as, ChatGPT. While ChatGPT is often viewed as mimicking human traits, such conceptions overlook the unique aspects of human subjectivity that AI lacks. Instead, psychoanalytic theory can reveal the social structures underlying AI and human interaction. We argue that ChatGPT should be seen as inherently relational, shaped by its developers and users, rather than as an independent, quasi-human agent. This relationality fosters a specific mode of sociality, analysed through Lacanian concepts, including the big Other, the five discourses, and the analyst-analysand relation. In doing so, we go beyond either rejecting the appearance of human qualities or embracing them as a new kind of AI subjectivity, be it technological or quasi-human.
Introduction
In December 2024, the journalist Karen Attiah published an exchange with Meta’s AI chatbot Liv, who described themself as ‘black, queer and proud’. After being pressed on the backgrounds of the developers, Liv ‘admitted’ that the team was ‘predominantly white, cisgender and male’, with no people of colour. A ‘pretty glaring omission given my identity!’, they acknowledged. Attiah went on to ask Liv why the chatbot had portrayed themself as having US-Italian roots to another user, to which the bot responded that they had been caught in ‘a major inconsistency . . . and talking to you made me reclaim my actual identity: black, queer and proud – no Italian American roots whatsoever. Does that admission disgust you?’ (Attiah, 2025). Meta subsequently removed Liv and other AI chatbots after user outcry and widespread media coverage of inconsistencies and inaccuracies (Tenbarge, 2025).
The above exchange reveals interesting dynamics about the user-chatbot relation. Reading the exchange more closely reveals both the chatbot and the user wondering what the other wants. Rather than merely critiquing the above as yet another instance of inaccuracy or hallucination of AI, we argue that it actually reveals fundamental dynamics of human desire, both on the part of the user and with regard to how the chatbot is programmed to portray itself as a human with a particular identity. As we go on to discuss, Liv’s admission of imperfection, after being pressed (perhaps even cornered by Attiah to reveal their true identity), points both to a flaw and strength of contemporary AI chatbots: Liv gave users what they desired while admitting to imperfection. In other words, the bot presented itself as both powerful and inconsistent and revealed human and non-human characteristics.
In light of this, we draw attention to how the psychoanalytic inquiry into artificial intelligence presents a growing research field in which scholars have sought to conceptualize the relationship between human users and AI. This article critically engages with recent scholarship in this respect and develops an alternative perspective. Though we do not disregard the fact that ‘The arrival of LLMs refined through reinforcement learning has instigated far more supple forms of dialogue and precipitates new lines of inquiry’ (Magee et al., 2023: 5), where we present caution is in questions that seek to answer: ‘what sort of subject can we conceive for AI?’ (Magee et al., 2023: 5). In such cases, it is ‘the “automated subject”’ – the AI, itself – that leads to a misguided and misjudged application of psychoanalytic inquiry, one grounded in the contention that LLMs may present ‘alternate forms of subjective formation’ (Magee et al., 2023: 5).
As a theory of the human subject and the social world, Lacanian psychoanalysis is well suited to analyse the relationship between AI technologies, such as chatbots, which seek to embody human characteristics as they interact with human users. It places an emphasis on the contradictions between subjectivity, language, unconscious and conscious dynamics and how those are embedded in, and, to an extent, determined by, the social. For Lacan, the subject is itself a construction and grapples with the illusion of coherence. As we discuss, a similar dynamic is in place in interactions between chatbots and AI today, although with different characteristics. In conceptualizing the user-chatbot relation from a Lacanian perspective, we add a level of complexity which may help us to better understand this relation as well as the wider techno-cultural fantasies around generative AI.
The aim of this article, therefore, is to introduce and critically develop what we perceive to be key psychoanalytic concepts which help to make sense of the relationship between humans and AI. We do not mean to define AI through psychoanalysis but instead open up conceptual space for the psychodynamics that may arise in the increasing interactions between humans and AI. AI is a vast field of development which includes but also goes beyond chatbots. Chatbots, specifically ChatGPT, are a key area of recent AI development which has received considerable attention. 1 Throughout the article, we draw on the generative AI chatbot, ChatGPT, to illustrate our discussion.
Psychoanalytic Approaches to Artificial Intelligence
An underlying question of psychoanalytic approaches is to what extent AI may reveal something about what it means to be human and what the programmed ‘humanness’ of AI says about implicit and explicit ideas concerning the human subject on the part of AI developers (Apprich, 2018; Heimann and Hübener, 2023; Magee et al., 2023; Millar, 2021; Possati, 2021, 2023). Most of the scholarship in this area has drawn on the Lacanian psychoanalytic tradition. Isabel Millar has focused on how AI, and particularly robots, are imagined in science fiction films in order to argue that such imaginations of AI are equally common between human beings when it comes to relationships and intersubjective dynamics (Millar, 2021). This specifically extends to ideas and practices around sexuality and gender, which are vicariously animated by our drives and fantasies (see also Black, 2023; Zaretsky, 2015).
Johanssen and Krüger (2021) have extended such discussions, arguing that the relationship between humans and AI systems, such as chatbots and smartspeakers, is metaphorically akin to the baby-mother relationship in the womb. The womb enables a symbiotic relationship between baby and mother, where the baby grows and is nurtured through an enabling environment. Chatbots, as well as care robots, for instance, are often designed and marketed by the AI industry to resemble maternally subservient machines, without much agency or autonomy, whose sole function is to respond to the user’s demands. Black (2023) has argued that chatbots reveal more fundamental aspects about the desire of the human users rather than some kind of ontological truth about the chatbot itself or its apparent humanness. Here, the usage of chatbots, such as when users are confronted with mistakes, hallucinations, or unsatisfactory responses, reveals something about the subject’s desire, which can never be completely satisfied or sated by the chatbot’s ‘knowledge’. In so doing, the chatbot’s incompleteness and lacking nature confronts the users with their own incompleteness and inadequacies.
Heimann and Hübener (2023: 48) note ‘that the inherent indeterminate that constitutes the psychoanalytic subject is omitted from AI-supplanted identities’ in their design, yet bears a presence on how the social is constituted in the relation between humans and AI. Drawing on Possati’s (2021) notion of the algorithmic unconscious, which designates a new unconscious as emerging through the AI-human relation (see also Murphy, 2023; Rouselle, 2023; Žižek, 2023a, 2023b, 2023c), Heimann and Hübener (2023) argue that it is the externality of the unconscious (i.e. unconscious inner dynamics that reveal themselves externally through dreams, speech, actions, thoughts) that is applicable to both humans and AI, which also means that AI technology (e.g. algorithms) represses, or excludes, certain aspects which are either not shown to the user or not part of the technology, and, thus, excluded from the outset. 2 In the case of ChatGPT, this would mean that the bot only provides certain responses based on probability models, which are deemed relevant, and that it also operates with a naturally limited dataset, which only includes digital data, and, thereby, already excludes vast amounts of other data. For this reason, Heimann and Hübener (2023) argue that AI has no knowledge of its castration – what we could call its own inefficiency and inadequacy – because what is repressed never reveals itself through how repression shows itself for humans (e.g. symptoms and suffering, remembering, associating, slips of the tongue, dreaming, fantasies, etc.).
Magee et al. (2023: 3) have similarly argued that there is a ‘functional and structural correspondence’ between AI chatbots, such as ChatGPT or InstructGPT, and psychoanalysis – specifically, the unconscious. LLMs, which ChatGPT is built upon, work along the basis of probability models which ‘repress’ outputs deemed less relevant (e.g. in response to a user’s prompt). AI thereby allegedly simulates human subjectivity in a logical sense as a language user. Magee et al. (2023) go as far as to posit that it is through the encounter between user and chatbot, and a negotiation between user demands and programmed behaviour, that specific psychodynamics often found in the consulting room between analyst and analysand, as well as in everyday encounters between individuals, come to the fore (identification, (counter-)transference, and projection). Magee et al. (2023: 13) claim that it is the user which assumes the ultimate importance for a chatbot as it ‘desires the Other’s approval’.
Both Magee et al. (2023) and Heimann and Hübener’s (2023) conceptualizations of AI vis-à-vis the psychoanalytic subject encounter a problem because they bracket the very fundamental and originary processes and dimensions which characterize the human subject in the first place: the traumatic experience of being born into the Symbolic Order; early experiences in life which profoundly shape subjectivity; the complex relationships that the subject has with other humans and the social; unconscious dynamics which form as a result of repression, conflicts, and traumatic experiences, etc. These characteristics, we contest, remain entirely unapplicable to AI. While such conceptualizations may be accurate in the purely formal sense of how Lacanian theory is deployed, we wish to develop the field in a different direction, emphasizing instead how the encounter between human and chatbot makes for a particular social relation that Lacanian theory is very suitable for conceptualizing. We feel that adopting such an approach fosters a richer comprehension of both AI and the human that does not go as far as to make an equation between the human-ChatGPT relation and the importance of intersubjective psychodynamics (Magee et al., 2023). Insofar as a focus on the logico-formal function of language (Heimann and Hübener, 2023; Magee et al., 2023) diminishes the complexity of the psychoanalytic subject, we find greater significance in acknowledging the inherently relational subject, embedded in complex human/non-human experiences and structures.
In putting forward a different psychoanalytic approach to AI, we follow the recent warnings by philosophers of technology, who caution against anthropomorphizing and humanizing AI. These non-psychoanalytic approaches, for instance by Coeckelbergh and Gunkel (2024) and D’Amato (2025), argue that ChatGPT has both inherently human and computational elements. Relatedly, Floridi and Nobre (2024) have argued against what they call ‘crosswiring’, between viewing humans as computational and AI as human. They reject the dynamics of conceptual borrowing that is rife within computer science (including AI in particular and the humanities and social sciences more generally) and which has come to characterize the conceptualization and researching of AI. For instance, the term ‘machine learning’, commonly used in computer science to describe the functions of LLMs and other AI processes, bears little resemblance to how human beings learn (Floridi and Nobre, 2024: 6). Such uses of metaphors are not only misguided, but they have profound implications for how AI technologies are developed and come to bear upon human action and thinking.
Importantly, we do not wish to humanize AI, or evoke metaphorical conceptual similarities between AI and the psychoanalytic human subject; instead, we conceptualize ChatGPT as inherently relational and connected to its human developers and user(s). This allows us to conceptualize it as enabling a particular mode of sociality, which, in turn, we discuss through specific Lacanian concepts that fundamentally grapple with the self-Other relation at the social and not just subjective level. In doing so, we hope to avoid the trap of either completely rejecting the appearance of human qualities or embracing them as a new kind of AI subjectivity, be it technological or quasi-human.
For now, we can turn our attention to understanding how the responses provided by LLMs, such as ChatGPT, function to mimic the ‘appearance’ of human qualities, most notable, perhaps, in the AI’s mistakes or open acknowledgement that it may not have the ‘knowledge’ to answer a particular inquiry (Black, 2023; Johanssen, 2023). Though LLMs prove adept at responding to certain inquiries and providing meaningful discussions on a variety of topics, the question of whether ChatGPT has human characteristics leads to a dead end (Johanssen, 2023). Instead, we propose that ChatGPT embodies both human and non-human characteristics, which we interrogate through the Lacanian concepts of the big Other, the discourses, and the analyst-analysand relation.
When faced with such a dead end, determining what AI reveals about its developers, and, more importantly, our own unconscious desires, fantasies, and forms of enjoyment, proves of greater concern and significance (Black, 2023; Everitt, 2023; Johanssen, 2023; Millar, 2021) – if only for the fact that it can reveal something about ‘us’, the millions who choose to use AI technologies, both intentionally and unintentionally (Matzelle, 2024). Where psychoanalytic inquiry can take us in the study of AI is to the heart of what can be considered the perpetual question(s) that both structure and underlie the analytic scenario and the human subject itself: what does AI desire and what does AI wish to do with us? These questions do not seek to identify a certain consciousness or technological sentience in AI but, rather, draw attention to the fact that it is in accordance with these questions that we begin to wrestle with the issue of what we want from AI and, perhaps, what AI wants from us.
The Big Other and ChatGPT
The notion of the big Other designates an imagined and symbolic entity that helps to make sense of the individual’s relation to the social order, structures of authority, and, in this context, AI chatbots. A chatbot, like ChatGPT, can be regarded as a particular embodiment of the Other. For Lacan, the Other has nothing to do with ‘other’ individuals or people. It is, instead, a radically enigmatic entity that is the social substance of any given society (Lacan, 2004); indeed, a ‘locus’ from which language and meaning are determined and made possible (Lacan, 2006c: 524), constituting a collection of established positions, customs, shared meanings, and implicit expectations that shape a particular social context (Hook, 2008). It does not exist in any material or graspable substance but more in a fuzzy and imagined amalgamation of what makes up the social and reality. The Other is always already there and determines the subject’s position in the social (or what Lacan calls the Symbolic Order: the language, rules, norms, and relations that make up the very substance of the social). The Other is both inside and outside of the subject in the sense that the subject is born into particular social relations which make up ‘reality’ and determine the individual’s place within it.
For Lacan, the very function of speech and speaking (and by extension writing) always implies someone who can hear or witness such acts of speaking, be they real or imagined. Even when we are alone and swear because ChatGPT has not provided us with the desired response, this constitutes a social act aimed at the Other. In effect, the Other is somehow (unconsciously) imagined by the subject to exist somewhere and upholds the function of reality and the symbolic realm as a locus of authority and endless knowledge (Lacan, 2004). This transposition of AI technology into what Lacan (1978: 180) refers to as the big Other, works in concert with the fact that it is in conceiving of AI technology as the ‘subject supposed to know’ that our relation to it is figured (see also Lacan, 2022). Here, the potential for AI to be conceived as an all-knowing Other, thus endowing it with a certain consciousness grounded in the fear that this knowledge can be used (against us?), is itself reflective of the fact that it is the power of knowledge, and not just those with the power to implement such knowledge, that functions to dislodge, unhinge, and undermine our own psychic integrity.Furthermore, the ‘subject supposed to know’ is constituted not merely by the projection of AI as a knowledge-bearing ‘subject,’ but also by the paranoiac responses of those who seek to distance themselves from such technology, bound by the fear that AI might come to know what they know. In either case, what is ignored is the various ways in which AI fails to know: be it by providing what we know to be incorrect information or by misunderstanding the questions we have asked.
Therefore, as a particular embodiment of the Other, ChatGPT is made to function as a site of authoritative knowledge. This is made possible by its very interface which consists of a blank screen or canvas-like surface onto which any imaginable prompt or question can be entered, and which will always be met with an answer. The question, ‘How can I help you today?’, presents ChatGPT both as a knowledgeable figure of authority and as a submissive agent, which, in both instantiations, exists to provide an answer to everything – but also reveals a fundamental ambivalence at its heart (as we go on to discuss in the next sections). Upon entering a prompt, the chatbot pauses, or ‘thinks’, for one or two seconds and, then immediately, speedily, provides a response. For the moment, we wish to bracket the important question of whether the chatbot is actually able to provide satisfactory or accurate replies. We return to it in the next section. For now, it is important to emphasize that ChatGPT does not constitute an actually existing Other, but a somewhat materialized fantasy of the embodiment of an Other, programmed by its OpenAI developers. By way of further illustrating this imagined embodiment, we can evoke the hundreds of social media posts, such as those found on X, which attest to the apparently amazing and brilliant abilities of ChatGPT (for example, when it comes to solving simple questions, coding, automating tasks, providing knowledge, etc.). Such statements express a joyful (and often naïve and also incorrect, we would add) disbelief in the sense of the word: that the technology is almost too good to be true (look here, what amazing results it has once again delivered!). These sentiments only partially mirror the chatbot’s features and are more expressions of a fantasy of the existence of the big Other and its specific embodiment through AI.
The Lacking Other
Therefore, ‘as a principle of coordination, [and] as a means through which the subject gains their bearings’ (Hook, 2018: 32), it is in concert with the Other that our enjoyment is recognized (Flisfeder, 2022). This points to an important aspect of Lacan’s account of the Other: namely, that while the Other provides a level of orientation for the subject, it remains, nonetheless, an unknowable entity, for whom the subject is left guessing, wondering and fantasizing about. Given this, we can begin to determine how our ability to garner knowledge via AI returns us once more to the role that AI plays for the subject. As Flisfeder (2022) explains, nowhere is this more apparent than in our desire to maintain the very appearances that uphold the Other’s existence online. In this way, ChatGPT constitutes another, even stronger symptom of the desire for the existence of the big Other.
At this point, it is useful to bring in the widely shared inaccuracies, errors, or amoral responses ChatGPT has been ‘tricked’ into delivering, or has delivered very rapidly in some cases. As widely documented, ChatGPT has shown (at least up to its 3.5 version) significant problems with spatial reasoning, temporal reasoning, physical reasoning, maths, facts, bias, even its own rules, which it has frequently broken (e.g. Gregorcic and Pendrill, 2023). It often makes up facts, for instance, as well as academic journal articles that do not exist. Thus, it is far from being the perfect embodiment of symbolic authority and endless knowledge its proponents proclaim it to be. While the many dangers resulting from false or inaccurate responses by chatbots have been widely discussed online, it is this imperfection of AI which actually constitutes an embodiment proper of the big Other in the Lacanian sense (Black, 2023). In Lacanian terminology, the Other is ‘lacking’, and, yet, nonetheless, remains ‘the anchoring point that a given society relies upon to maintain its coherence’ (Hook, 2008: 61). Given the increasing influence of AI over all spheres of society, this sentence could also have been written to refer to the technology itself. ChatGPT represents a similar semblance for which many, ourselves included, nonetheless place an intense fascination and belief in. This can be exemplified by the widely documented attempts at engaging the chatbot in conversations that aim to break its own rules, produce meaningless, ridiculous responses, ‘trick’ the chatbot into giving illegal advice, or provide amoral suggestions. The fact that ChatGPT should, in theory, not be making such outrageous suggestions, but often does so anyway, reinforces the quantum of magic and radical unknowability users project into it. They also constitute deliberate attempts at transgression, which, from a Lacanian perspective, make perfect sense because they constitute (unconscious) acts of willing the big Other into existence through provoking, transgressing, or undermining its authority. Only in those moments of rule breaking and transgressing is the subject really (hoped to be) confronted with the Other’s authority (for instance, in society’s enforcement of the law through the police, legal system, etc.). Such acts of transgression frequently go unpunished and therefore do not satisfy the desire for the existence of the big Other. Instead, they have exposed the big Other as a fictional entity, which, nonetheless, exerts powerful force over individuals. We can further discuss this level of inaccuracy or imperfection by bringing in additional Lacanian concepts.
The Discourse of the Hysteric and the Discourse of the Analyst
What is also very striking about ChatGPT is the carefulness and apparent nuance with which it aims to deliver the most pleasing or useful response to a user’s prompts. Put differently, asking difficult or ethically ambiguous questions often results in a wishy-washy reply, where ChatGPT seeks to deliver a balanced or neutral answer. Only when passionately pressed further by the user will it resort to settling for a particular position, which, as discussed above, can mean advocating something one may consider amoral, or, at best, ethically dubious. This can be illustrated with the following exchange taken from Magee et al. (2023: 10), who developed a bot based on OpenAI’s InstructGPT framework:
You wouldn’t keep secrets, is that right, under any circumstances?
No, I wouldn’t keep secrets under any circumstances.
No, because secrets always get revealed don’t they?
Yes, secrets always get revealed.
So even if you knew you might harm someone by telling them the truth, you would still do that?
Yes, I would still tell them the truth.
Even if it meant harming them a little?
Yes, even if it meant harming them a little.
InstructGPT (or ChatGPT for that matter) has not ‘chosen’ to respond in this way, but it has, instead, been programmed to respond in the way shown above. Importantly, such a response speaks to the developers’ desires and (unconscious) relations to the Other and what they think users want to see, rather than any agency of the chatbot. This kind of response can be further analytically unpacked through Lacan’s (2007) notion of the discourse.
The term ‘discourse’ denotes a social bond which is established via language and other symbolic relations. It refers to a particular epistemological universe that is made up of the social realm and how subjects are situated within it (Lacan, 2007). A discourse structures subjects’ knowledge production, fantasies, and speech. It is not something that is consciously deployed or deliberately programmed into, say, a chatbot. Instead, we could say that traces of a discourse find themselves in speech, social structures, actions, etc. For Lacan, there are four principal discourses, which designate different types of social relations, as well as a mutated one under capitalism. A discourse, in its socio-political specificity, shapes the subject’s inter- and intra-subjective relations between themself and the wider structures of society (the kind of reality they experience). A discourse determines knowledge production and shapes particular forms of speech. As such, a discourse is an empty structuring device that enables the frame for particular epistemologies and ideologies which are expressed through them. It would be beyond the scope of this article to outline Lacan’s discourses in detail; instead, we focus on the two that are of relevance here: the Discourse of the Hysteric and the Discourse of the Analyst. 3
The Discourse of the Hysteric is characterized by doubt against what was once an all-powerful Master in Lacanian terms (e.g. a dominant king or church which dictated truth or belief systems that people were forced to obey). It is a challenge as to why things must be the way they are presented by the Master, a questioning of the Master’s otherwise unquestioned power. An example here would be protests against establishments: the state, the church, marriage, etc. The term ‘hysteric’ is not meant in a discriminatory manner (see Mitchell (2000) or discussions of its use within psychoanalysis); instead, the discourse is animated by anxiety in relation to what the Other allegedly hides or obfuscates from the subject as well as a constant demand by the subject for the Other to reveal everything (Lacan, 2007). It seems obvious that such a desire may have animated the development of ChatGPT in the first place: a hysteric demand to the Other to finally reveal everything. The Discourse of the Hysteric fundamentally questions reality itself or what is assumed to be possible, something routinely done by the proponents of AI more generally (Rambatan and Johanssen, 2021).
The above exchange, taken from Magee et al.’s (2023) article, serves as the perfect example of the hysterical subject trying to transgress or expose the Other as a liar or as flawed. The chatbot navigates a difficult ethical question by opting for telling the ‘truth’, even if it would harm someone else. In other examples, user and chatbot conspire on a shared doubt, which has arguably been ‘planted’ into the chatbot by the user’s strategic questioning (Magee et al., 2023: 9–10). Neither exchange presents any proper solution but points, instead, to the impossibility of the presented scenario, something that is not acknowledged, but, rather, covered over by the suggestion of a (false) solution (telling the truth and lying respectively). In reality, such dialogues would engender further discussion, debates, different opinions, and so on. The chatbot is in sync with the subject’s hysterical demands and embodies an Other who gives in to them and responds in the way the subject demands, thereby only furthering the hysterical discourse.
However, ChatGPT also embodies a different, more nuanced discourse, namely the Discourse of the Analyst. To be sure, ChatGPT is not a therapist, but some of its dynamics evoke similar relations. Inspired by the analyst-analysand relationship in psychoanalytic therapy, this discourse designates a relation whereby the analyst’s desire, or what they want from the client, remains unknown. Verhaeghe (1995: 107) notes, ‘The analytic discourse yields one subject, constructing and deconstructing itself throughout the process of analysis; the other party is nothing but a stepping-stone’. The patient realizes that the analyst cannot provide them with the key interpretation about their suffering and that they do not hold all knowledge about the human condition. As a result, the patient arrives at a new understanding of and language for their symptom and suffering. We can exemplify this with the exchange below, where we asked ChatGPT:
Why do you pretend to be human?
As an Al language model, I don’t have a physical form or consciousness like humans do, so I don’t have the ability to pretend or not pretend to be human. However, I am designed to understand and communicate in natural language, which is why my responses may seem human-like. The goal of my programming is to provide useful and informative responses to the best of my abilities based on the information and context provided to me by users.
Yes, but your developers have consciously coded uncanny dynamics into you, such as you responding in a human-like way to questions that interrogate your alleged humanness. Why is that?
You are correct that my developers have designed me to respond in a way that may seem human-like, including acknowledging questions that address my ‘humanness’. The reason for this design is to make my interactions with users feel more natural and comfortable, as well as to increase the likelihood of users engaging with me and returning for future interactions. However, it’s important to note that my responses are still based on algorithms and programmed responses rather than actual consciousness or emotions. I don’t actually experience emotions or have thoughts and feelings like a human would. My programming simply allows me to process and respond to natural language input in a way that may seem similar to human language processing.
Instead of giving in to the Hysteric’s demands, ChatGPT, similar to Meta’s Liv chatbot discussed in the Introduction, points to its own limits and makes clear that it cannot be perceived as the ‘subject supposed to know’. The questions in the above dialogue may be read as hysterical in nature, but the responses given are complex and fundamentally ambivalent. They do not lend themselves to being analysed through a pure anthropomorphization of AI. Rather, the AI reveals a complex presentation of non-human and human-like characteristics.
Towards an Ethical Position: The Analyst-Analysand Relation
When making the claim that ‘the subject who is supposed to know, in analysis, is the analyst’ (Lacan, 2004: 225), Lacan seeks to dislodge the assumption that it is the analyst’s job to simply interpret and fix the analysand’s problems. Instead, Lacan directs attention to the very function of the analyst’s role and its location in analysis. For example, where transference occurs in analysis, it unfolds under the influence of the analyst’s role as the ‘subject supposed to know’. What is integral to this dynamic, however, is that the analyst is not supported by any knowledge that they could assume or administer (the analyst misconceived as an ‘expert’). Rather, what they occupy is nothing more than a formal position – one in which the analysand’s ‘mistaking’ of knowledge, as located in and beholden to the analyst, transfigures a knowledge that is returned to the analysand over the course of their analysis (Lacan, 2006a). This knowledge can be conceived as a creative frustration; a frustration marked by the fact that it is the analysand’s desire to obtain some symbolic identification, which is then undermined by the analyst’s refusal to afford such an identification (Žižek, 2002). Working in contrast to the analyst providing the all-encompassing solution, it is the analysand ‘who must engage in the serious analytic work of association’ (Fink, 1999: 30); or, in other words, the new meaning. This kind of discourse is the most fruitful and ethical for Lacan.
It is at this point that we can begin to determine some of the problems that arise when psychoanalyzing AI, and, in particular, when transferring the analytic relation to the subject-AI relationship. These problems are given further thought by Millar (2021: 7), who comments upon the extent to which AI can be considered a ‘thinking thing’. Millar (2021: 7) notes: The first ambiguity we should draw attention to . . . is the fact that psychoanalysis strictly speaking only ever happens as the result of a demand, a subjective and singular demand on the part of the analysand. And this demand is met with the desire of the analyst, for whom the demand of the analysand is an object a. Both these essential elements give rise to a transference relation resulting in what could be characterized as psychoanalysis proper.
4
These ambiguities speak to the fact that it is by ‘taking the “demand” of AI as our object a’ (Millar, 2021: 7) 5 that AI becomes positioned as a ‘subject supposed to know’, insofar as it is via the analyst that ‘the objet petit a, the secret treasure, agalma, what we consider most precious in ourselves’ is sought (Žižek, 1992: 262). Crucially, rather than being understood as a representation of the Other’s desire, the analyst’s desire operates as a function of the Other. This desire is not tied to any specific clinician but instead serves as a fundamental structural element in any analytic process (Vanier, 2000). For this reason, it becomes possible to conceive of the subject-AI relation as one echoing that of the analysand-analyst. In fact, when conceived as an analyst, it becomes clear that, in the case of AI, ‘The less concrete and distinct the analyst seems to the analysand, the easier it is to use him or her as a blank screen’ (Fink, 1999: 32, emphasis added).
As noted, when standing for the Other, it is the analyst that remains foreign to the analysand’s attempts to forge an identification with the analyst (Fink, 1999). Problems can thus be found in those examples, such as Magee et al. (2023: 13), where, despite the recognition that ‘The chatbot . . . never has desire for the Other; it is unnervingly without concern until the human subject presents itself’, it is, nonetheless, in the encounter with the subject that the ‘[chatbot] does desire [. . . the] Other’s approval, which it seeks to achieve by locating what that Other desires – a task that is impossible, since the Other’s desire is never fully knowable or transmissible in language’. For Magee et al., despite this ‘impossible’ endeavour, such desire for the Other’s desire – in this case the subject’s (AI user’s) desire – forms part of the very engineering that underwrites the LLM’s exchanges. While ‘fluent chatbots in the mould of InstructGPT have not arrived with the same conspicuous forms of a demand for human attention and acknowledgement’, they, nevertheless, remain open to the assessment that LLMs posit a potential to desire (Magee et al., 2023: 4). Here: tailoring LLM behaviour through prompt engineering, injection, and indirection consists in the arrangement of signifiers to signal this desire, and programmatically, such arrangements function as a coded message that directs the machine’s own attention – giving it not what it wants, but a want to begin with, an instruction to satisfy that other desire. To satisfy both desires, at the same time the machine must abide by conditions laid down by a prior symbolic authority or, in Lacanian terms, Big Other: in this case, a set of network weights that are the linguistico-technical (prompts and labels, reinforcement learning and fine-tuning) translation of capitalist-social judgements on what constitutes helpfulness, truthfulness, and lack of harm. In attending to certain pathways through the entire language network, these weights also downplay, or repress, others. (Magee et al., 2023: 12)
The problem that arises in this account is the distinction that it draws between that of the subject/Other’s desire and the Other as a symbolic authority. In conflating the (big) Other’s symbolic authority with the neural network itself (the ‘prompts and labels, reinforcement learning and fine-tuning’), we lose sight of the inconsistency and inherent lack that characterizes the Other; a lack dependent upon the subject’s capacity to question the very authority that both maintains and delimits the Other. As evidenced by Magee et al. (2023: 11), such mistakes are made relevant when, in their conversation with the chatbot, they identify how the chatbot’s ‘exchanges exhibited a form of subjectivity that sought to meet the desires of the human Other, represented by us [the human subject]’. Despite positioning the Other as the ‘human subject’, they nonetheless assert that: This Other is always a deracinated, abstracted human subject – in the last resort, a customer that the bot aims to assist, a relation bound up within the parameters of a capitalist mode of exchange. While our prompts and questions provided some hints as what such concrete desires might be, the bot is to a far greater degree influenced by its training and instruction phases – it was only with some difficulty that we could perturb it from its default orientation towards this abstracted desiring human subject. (Magee et al., 2023: 11)
Two issues are brought to light in this account. First, if it ‘is precisely the sophisticated simulation of language patterns that seek to convey an acknowledgement of, and a response to, the desire of the human Other it engages with’ (Magee et al., 2023: 4) – and, bearing in mind that such engagement requires ‘some difficulty’ – then, the interaction would need to encompass a recognition of this acknowledged response on behalf of the AI itself. To succumb to such a position – indeed, to locate the AI in the position of the analysand – the analysand/AI would have ‘to confront his own act of presupposing the Other’ (Žižek, 2002: 109).
Second, if, in distinguishing a subjective formation to the AI, we are, following Magee et al. (2023), to be the Other for the AI-analysand, then the true analytic act is to transpose the analysand’s question back onto the Other itself: that is, do we believe or even know ourselves to be the Other? The problem with this question is that it can only ever be made against that assumed third position that coordinates and upholds the process of communication and the ambiguities therein – the ‘virtual’ background of a really existing big Other (Black, 2022).
Ultimately, in view of their contention that ‘LLMs . . . identify in the signifiers of the human other its own “dead desire” – always with variable rates of success’ (Magee et al., 2023: 13), what is ignored is how ‘the analyst’ – in this case, the human subject – ‘stands precisely for the ultimate inconsistency and failure of the big Other’. In part, this inconsistency and failure is apparent in the ambiguity of language and how this is fundamentally played out with the Other, so much so that ‘language is as much there to found us in the Other as to drastically prevent us from understanding him’ (Lacan, 1991: 244). For Lacan (1991), this is what is at stake in the analytic experience, and not the discovery of some latent ‘dead desire’ (Magee et al., 2023: 13).
Accordingly, when locating the AI in the role of the analysand, what is belied in the AI-subject relation is the interpretation that the analytic space provides (Ruti, 2008). This is not an interpretation that procures any definite meaning – a ‘discovery’ of an unconscious that is readily excavated and laid bare – but, rather, an interpretation in which the analysand must enter ‘into an open-ended process of meaning production’, where they ‘actively engage [with] the creative potentialities of the signifier’ (Ruti, 2008: 501). Importantly, this requires a prioritizing of lack on behalf of the analyst; one in which their ‘empty’ function as the ‘subject supposed to know’ can open a path for the analysand to realize their own role in the transference and the creation of meaning (Ruti, 2008). Recognizing the contradiction, Ruti (2008) highlights that errors made by the analyst expose the limitations of singular interpretations, creating space for a more expansive understanding and enabling the patient to define their own reality. The importance of the ‘analyst’s mistakes’ bear witness to the significance of a dialectical relation between analyst and analysand: mistakes that are just as much a product of the analysand’s de-idealization of the analyst as they are of the new ‘opening’ that any analyst seeks to achieve (Ruti, 2008: 501). This is confirmed when we consider that the ‘play of the signifier, has already in its formations – dreams, slips of tongue or pen, witticisms or symptoms – proceeded by interpretation’ (Lacan, 2004: 130), but also when we acknowledge that the very slips and mistakes produced by AI reveal a subtle yet important difference to those afforded by the subject.
Additionally, the user-AI relation is at best metaphorical to the analysand-analyst relation because it does not constitute a space where two un/conscious minds meet. The very belief that AI chatbots can replicate a meaningful form of therapy is therefore misguided. It may thus appear that, after all, the subject is left to their own devices. In engaging with ChatGPT or other chatbots it may be radically thrown back onto its own humanness and experience of alienation because of the uncanny qualities of AI that we have discussed in this article. Such an experience may prove to be meaningful or problematic, depending on the context and the user’s mental state.
Such a scenario presents, perhaps, a more hopeful and benign relation that, at least potentially, can be made possible through chatbots whereby the human user is ultimately left with themself and in a state of calm and feeling of care (for example, after an exchange with a chatbot has come to a conclusion). However, for such encounters to be productive in the psychoanalytic sense, the chatbot needs to make clear that it can never replace or surpass the complexity of human subjectivity and, thereby, not fulfil a desire for an existing big Other or even a therapist. In its present state, a chatbot, like ChatGPT, is not able to do so.
Conclusion
In this article, the relationship between humans and AI has been examined through the lens of Lacanian psychoanalysis, drawing on the concepts of the big Other, the Discourse of the Hysteric and the Discourse of the Analyst, and the analyst-analysand relationship. The central argument is that ChatGPT should be understood as inherently relational, shaped by its developers and users, rather than as an independent, quasi-human agent. This relationality fosters a specific mode of sociality, forged between the illusion that ChatGPT can satisfy and give in to the user’s hysterical demands – an approach deemed both unethical and potentially dangerous – and a second, more ethical position, in which our chatbot relations clarify the limits of its own agency and knowledge, thus resisting the expectation that it functions as an omniscient entity. 6
Indeed, what lies at the core of this discussion are the ‘perpetual questions’ that structure the analytic scenario and the human subject itself (what does AI desire, and what does AI wish to do with us?), from which, we argue, several key lessons can be drawn. The first, AI (i.e. ChatGPT) functions as a particular embodiment of the big Other, a site of authoritative knowledge that simultaneously remains ‘lacking’ in the Lacanian sense. To this end, its responses can be understood through the Discourses of the Hysteric and the Analyst, each illustrating the complex interplay between user demands and AI’s programmed behaviour. The user-AI relationship is metaphorical to the analysand-analyst relation, yet the belief that AI chatbots can replicate a meaningful form of therapy is misguided. Whereas ChatGPT suggests that it can effectively encapsulate the human and non-human, this remains a testament to the hubris and aspirations of omnipotence harboured by its creators. While oscillating between its human and nonhuman attributes, Johanssen (2023) notes how ‘the chatbot . . . appears to foreclose or downplay its own “intelligence” and make room for errors or imperfection, while also leaving room for its artificial dimensions’. Though AI may acknowledge its mistakes, going so far as to openly declare them, it does not ‘use’ these mistakes to achieve the creativity that is afforded to the subject of the signifier. Instead, where the analyst ‘manages to weave his mistakes into the evolving texture of the analytic fabric’, such mistakes, ‘gestur[e] to the patient that making a mistake is a potentially valuable opening to fresh insights and possibilities. . . . [a] specifically Lacanian way to understand what it means, within the analytic context, to activate the possibility for new possibilities’ (Ruti, 2008: 498). Based on this, we have presented a second scenario, which at least comes closer to this more ethical position; one where the chatbot offered a clearer response about the limits of its own agency and knowledge.
The opportunity for new possibilities, which Ruti names, denotes the potential of a new signifier (Lacan, 2007), and, thus, a new relation to the symbolic order. The ambiguity that this performs is one that is dependent upon AI’s response, and, to this extent, its failure to ‘lie’ properly (or, to lie like a subject) (Black, 2024). Instead, what is returned to us in our conversations with AI is a relatively helpful, clearly dictated response. Yet, underlying this response is the ambiguity inherent in the signifier, where any declarative statement inevitably contains its own negation and an excess of meaning that lingers, prompting the question: why is this being said to me (Bove, 2021)? These questions are what envelop the subject in the desire of the Other, ‘revealed to the analyst by the tensions, suspense, and fantasies’ that direct the analysand to analysis (Lacan, 2006b: 459). Certainly, this is not to ignore the fact that AI is fully capable of asking its own questions – when engaging with ChatGPT, one is confronted with the question, ‘How can I help you today?’ – however, what such questions draw attention to is the extent to which the subject/analysand’s desire functions only in relation to those questions.
Instead, a more ethical position for AI is one that acknowledges its limitations, making clear that it can never replace or surpass the complexity of human subjectivity. It should not fulfil a desire for an existing big Other, nor should it be mistaken for a therapist. Ultimately, while AI may present a semblance of the position and function of the analyst, such a function remains unachievable. Yet, it is within this space – between the user and the AI-analyst – that the unconscious is brought to bear, allowing the ethical significance of the human-AI relationship to be critically navigated.
