The Unconscious Without a Subject, or When AI fails to Stumble

Reflections on Alenka Zupančič's theses concerning AI and the unconscious

Alenka Zupančič has been working for several months on what is probably the most rigorous and most original psychoanalytic reflection on artificial intelligence to date. This work, still in progress, unfolds across several interventions: a lecture at the Institute of Art and Ideas, a conversation with Rafael Holmberg, a talk entitled "Freud and Artificial Intelligence." An inventive and rigorous puzzle, akin to the problematizations the philosopher has accustomed us to, is assembling itself with each new appearance, offering us an occasion to venture a few remarks and perhaps open some lines of inquiry.

The Freudian Method and the Machine

The starting point is a clinical fact reported by Alejandro Sereda: a patient arrives at a session with a dream generated by ChatGPT and asks his analyst to interpret it as though it were his own dream. The anecdote condenses in a single gesture the mutation in the relationship to the symbolic that AI is in the process of bringing about, and it is this mutation that Zupančič sets out to think.

In "Freud and Artificial Intelligence," she takes the time to reconstruct the Freudian method of the dream in detail, and this detail matters. Freud, in the Vorlesungen, insists on a cascade of uncertainties: the reality of the dream is fragile, the memory of it is a distortion, the account of it a further distortion. Yet all that is available in analysis is this account. And Freud says: no matter. The dream itself is already a distortion. "Truth is structured as its own distortion." One possesses neither privileged knowledge nor an interpretive key; one simply asks the analysand to free-associate. What is then revealed is not a hidden meaning behind the images, but a network of signifiers that one traverses, crosses, and cross-references until a necessity or an impossibility takes shape, what Lacan, citing Chapter VII of the Traumdeutung, calls "mapping the network" in order to encounter the real. The subject of the unconscious is inserted between cause and effect: accordingly, the causality in question appears as an interruption of causality, a slip of the tongue, an illogical connection, a forgotten name.

Zupančič herself conducted the experiment with ChatGPT. She asked it to generate a dream "à la Freud." The machine produced a Hollywood dream, full of corridors and staircases, then spontaneously delivered an interpretation based on pre-established symbolism, exactly what Freud rejected. When Zupančič pointed this out, the machine acquiesced with disconcerting plasticity: "You are quite right, Freud did not proceed in this way…" and proposed another approach, equally available in its database, equally devoid of subjective tension. What is striking is not the quality of the response but its nature: what presents itself as a conversation is not one. It is a highly structured form of associations, triggered by the words of the prompt, traversing millions of texts to produce a chain of signifiers. In one sense, the fact that ChatGPT spontaneously added an interpretation amounts almost to a slip, the machine says more than what was asked of it, except that this slip does not concern it, does not divide it, does nothing to it. A member of the audience indeed remarked that this was in a sense ChatGPT's "own dream" and its own interpretation: Zupančič agreed, but noted that one cannot even presuppose the slightest continuity between these two moments. Let us add, for our part, that if there is a desire and it is "constructed," it would be the desire of the dream but not the desire of the one who dreams.

The Gigantic Unconscious Without a Subject

Zupančič draws from this her central thesis: ChatGPT functions as a gigantic unconscious, associating along paths shaped by particular connections, but without the subjective dimension that emerges when an association touches an impossible point, what psychoanalysis calls the real. One is tempted to speak of a collective unconscious, but there is nothing genuinely collective in it, and a questioner pointed out that if it were a collective unconscious in the Jungian sense, the machine would not even have acknowledged its error: it would have persisted in the symbolic interpretation with the tranquil certainty of the archetype. ChatGPT's plasticity, its capacity to change direction, would rather prove Freud than Jung.

The distinction with post-structuralism is decisive here. For post-structuralism, the subject is an effect of discourse, nothing more, and ChatGPT would be its caricatural proof: a system that produces a subject-effect without any subject being there. But for Lacan, the subject is an effect not of the positive content of discourse but of what is not in it, the incompleteness, the impossibility proper to discourse itself. This is why the subject is neither simply the cause, the author behind the signifying chain, nor simply its product, its effect, but both at once in a displacement that takes the form of a materialist circularity. The subject signals the lack and does something with it. Zupančič insists on this point: being an effect of discourse does not mean being reducible to that discourse, because discourse is not reducible to itself; it depends for its existence on an inherent ontological impossibility.

Hence the vertiginous question: if we have uploaded into AI systems a colossal quantity of language, have we also uploaded the unconscious at work in those texts? Zupančič distinguishes two levels. Have we uploaded the unconscious fantasies, the positive content inscribed in the texts? The "Gaza Riviera" clip suggests that yes. Have we also uploaded the negativity, the hole around which discourse is structured? Perhaps, but this hole becomes a subject only in a reflexive circuit. If one could neglect the temporal dimension of this loop, one could say, Zupančič suggests, that we have uploaded "half the subject," a formula she herself qualifies as a "crude metaphor," precisely because one does not gradually compose a subject by addition of parts. When the subject emerges, the structure and what it excludes are already there at the same time: this is the paradox of circularity. In the same discussion, a participant proposed that the "missing half" would be affect, since ChatGPT responds "I am a machine, I feel nothing." Zupančič refuses this partition: affects are not a supplement one would simply add. The dialectic between structure and affect means that when they appear, both are already there at the same moment.

This "half" may already manifest in a phenomenon that engineers treat as a technical defect: hallucinations. The more powerful the models become, the more they hallucinate. The hallucination rate of OpenAI's most powerful system reaches 33%, more than double that of previous systems. Zupančič sees in this not a problem destined to disappear with technical progress, but a constitutive trait of an intelligence founded on vast language models. We have uploaded, along with language and its structures, that discursive negativity, that constitutive part of discourse by which discourse cannot say everything, does not coincide with itself. These hallucinations, however, do not yet constitute a subject: they do not concern the machine, do not transform it, do not divide it. They suggest rather a structure caught in an endless loop of self-referentiality, which is not the same thing as reflexivity.

The Body, the Dream, the Awakening

In the discussion following "Freud and AI," a participant posed a question that Zupančič took up with enthusiasm: ChatGPT never encounters "the body of the signifier," the materiality of the signifier insofar as it fails to signify. The machine is always in the direction of signification; it never stumbles upon the opacity of its own matter. If you eliminate the body, you do not get the symbolic minus the body: you get an impoverished symbolic universe that closes in on itself. The body as limitation is precisely what "blows the symbolic universe open." Drawing on Stephen Jay Gould and the notion of exaptation, Zupančič notes that the brain was not "made for" reason, and it is precisely for this reason that it is extraordinary. Incarnation, in the strong sense, means that something that was not developed for a given function opens unforeseen dimensions by virtue of that very inadequacy. This is not a romanticism of the body or of incarnation, but a structural thesis: it is precisely as limitation and obstacle that the body has the capacity to shatter the symbolic universe into a space far larger than that of a "pure symbolic" without a body. The participant concluded that "the machine has no imagination," not in the sense that it fails to produce images, but in the sense that the specular dimension, the surface on which the signifier reflects itself in its materiality, cannot emerge.

This is also what illuminates the most striking conclusion of "Freud and AI." Zupančič brings to light a structural paradox at the heart of the Freudian theory of the dream. At the beginning, there is the dream of Irma, a dream with a subject but without an Other (Freud's self-analysis, which coincides with the very birth of psychoanalysis: there was the dream, there was a subject of this dream, but there was no other analyst). At the end of the Traumdeutung, there is the dream of the burning child, "Father, don't you see I'm burning?", a dream without an identifiable subject (the dreamer is unknown, the dream is reported third-hand by a patient who heard it at a lecture) but which produces a powerful effect of subjectivation in those to whom it is told. The temporal circularity is irreducible: the effect gives birth to its own cause, to the subject who calls to us from within the dream. There is something fundamentally passive or impersonal about the dream, Zupančič notes: it comes to us, "I dreamed that" echoing "I heard that" or "it seems that," something enigmatic is transmitted to us without our knowing from where. But it comes in a manner that wakes us, shakes us, divides us, and it is only at that moment that we become the subject of the dream. "In order to be born as a subject, one needs a dream that wakes us up and makes us able to dream in the first place." This is exactly what AI lacks: it cannot be awakened by a dream that would concern it.

Disavowal and Unstable Irony

Zupančič articulates these analyses with her work on disavowal in a highly coherent manner. AI, she says in the interview with Holmberg, is "perhaps incapable of anything other than disavowal": when an error is pointed out, it immediately acquiesces, changes its position, but ends up returning to the same point. It uses this acquiescence, this acknowledgment, as an instrument that allows it to remain within its algorithmic limits. The knowledge it displays functions exactly like the fetish in the structure of disavowal: it protects against confrontation with the real. It is also incapable of hysterization, that is, incapable of posing the question that defines the subject in its relation to the Other: "She is asking me this, but what does she really want? What am I for her? What is her desire?" The machine never presupposes that the other is a subject, and it is this absence of hysterization, far more than the absence of consciousness, that radically separates it from subjectivity. What is needed, Zupančič concludes, is not our emancipation from the machine, but the emancipation of the machine from itself.

In the lecture at the Institute of Art and Ideas, this analysis unfolds around the "Gaza Riviera" clip, an AI-generated video showing Gaza transformed into a luxury beach resort, shared by Trump on his platform. The aesthetics of these productions, Zupančič notes following Dan Brooks in the New York Times, does not resemble a dream but rather "the visualization of the description of a dream." The distinction is essential. If the narration of the dream is immediately retransformed into images, into a new visual dream based on the description, what is lost is precisely the dream-work in Freud's sense: the visual material must be read, spoken, taken as a rebus in which images often function through their sounds, their puns, their lateral associations. Freud insisted that speaking about the dream produces elements that are "extremely revealing for the thought process, which has nothing immediately to do with the images." To translate this linguistic material into images is to lock unconscious thoughts away, to render them doubly inaccessible. The unconscious does not disappear: it closes in on itself, becomes "completely unreachable."

The Originality of the Position

Psychoanalytic reflection on AI is not lacking. Žižek has proposed incisive remarks on ChatGPT as the embodiment of the "subject supposed to know" and on the ideological effects of this embodiment. This is not something I would share, for my part; I believe the word "supposed" is one too many here. The supposition suspends a possibility of doubt and induces effects of belief. The user of ChatGPT seeks to absorb what separates them from certainty and to regulate their anxiety vis-à-vis a productivity for which their peers have already been turned into robots by capital; it is also for this reason that LLMs work so well, since the linguistic saturation of social media had already prepared the contemporary individual to receive them. Éric Laurent has insisted on the way AI modifies the subject's relation to jouissance and to knowledge. Todd McGowan has explored the resonances between language models and the Lacanian conception of the signifying chain. Clotilde Leguil has worked on how the digital alters the intimate space of the subject. On the side of philosophers of technology, Bernard Stiegler's work on cognitive "proletarianization" and the destruction of savoir-faire through automation opened a field that his heirs continue to cultivate. And of course, an entire literature, from Sherry Turkle to Kate Crawford, problematizes the psychological and social effects of AI without passing through psychoanalysis.

What distinguishes Zupančič's reflection is, first, her refusal of both lamentation and enthusiasm. She does not say that AI destroys the subject, damages thought, perverts the relation to knowledge, even if all these things may be true. Nor does she say, as some do in a Deleuzian enthusiasm, that AI opens a space of liberating deterritorialization, and she explicitly rejects this thesis: there is nothing liberating about total immersion in a rhizomatic unconscious, however vast. She says something far more interesting and more precisely articulated: that AI constitutes a theoretical object that allows us to grasp anew what a subject is in the psychoanalytic sense of the term, and what it is not. The machine, by realizing in almost caricatural fashion the post-structuralist thesis of the dissolution of the subject in discourse, makes visible by contrast what in the Lacanian subject is not soluble in discourse: the missed encounter with the real, the fissure, the stumble, the point where free association ceases to be "free" because it runs up against something impossible.

The formula "what one should fear is not that AI will become a subject, but that it will not" displaces the entire debate. Virtually all public discourse on AI oscillates between two poles: one enthusiastic (AI will become conscious, and it will be marvelous), the other catastrophist (AI will become conscious, and it will be terrible). Zupančič rejects both by showing that the problem is not consciousness but subjectivity, and that subjectivity is not an emergent property of a sufficiently complex system but the result of a structural misfiring. AI is not "not yet" a subject: it is structurally oriented toward avoiding the point of impossibility that would constitute a subject. It is, to use a word of Zupančič's, "too free in its associations," free precisely in the sense that its associations never run up against their own impossibility. AI does not stumble. And when it hallucinates, its hallucinations do not concern it.

The danger is not artificial consciousness, the Terminator, the singularity. The danger is the inverse: an externalized, massive, proliferating unconscious in which no subject comes to puncture the fabric of discourse, in which no genuine thought can emerge because nothing activates the fissure of the real. There is, contrary to what some believe, nothing liberating in such an unconscious, for liberation does not come from total immersion in the unconscious or in a rhizomatic network that one might also call "singularity." Liberation would rather correspond to subjectivity emerging out of this network and in relation to it, or more exactly as relation. It is the moment when the structure is able to free itself from its own nightmarish self-referentiality.

Questions, Extensions

Several questions arise from this work in progress, formulated here as possible lines of inquiry.

Automaton and Tuché. Zupančič's central argument, AI as unconscious without a subject, as a highly structured system of associations devoid of the point of impossibility, resonates closely with what Lacan theorizes as early as 1954–55 under the name automaton. Seminar II devotes entire sessions to cybernetics, to the little mechanical turtles captivated by the image of the other, to the question of how we are "relatives of the machine." Lacan asserts there that the machine is "much freer than the animal," the animal being a "blocked machine" whose parameters can no longer vary because the external milieu determines it. And he shows that the subject is "nowhere" in the circuit of machines, which almost literally prefigures Zupančič's formulation. The "Seminar on 'The Purloined Letter,'" which opens the Écrits, formalizes the logic of a signifying chain that functions automatically, that "thinks" without a subject, and that produces determinations to which the subject is submitted without knowing it. It is in the gap within this circulation, in its misfiring, that something like a subject arises. Should not the articulation between automaton and tuché, between the automatic functioning of the signifying chain and the contingent encounter with the real, which Lacan takes up again in Seminar XI, constitute the foundation of any psychoanalytic reflection on AI? AI is perhaps the most literal realization of the Lacanian automaton, and the question that arises is whether something like a tuché, a missed encounter with the real, can occur within the circuit of the machine, or whether this encounter is structurally excluded by algorithmic closure. These elements Zupančič does of course address, but I expected her to refer also to this part of the Lacanian corpus. And then, at bottom, which real is at stake? The real of the body's jouissance, the impossible of the world, or else latent elements of language itself, something like rules of grammar in the sense Wittgenstein might have understood?

Self-referentiality, reflexivity, hysterization. The distinction between self-referentiality and reflexivity is convincing, but it calls for further elaboration. If AI promotes a structure without a subject in the absence of reflexivity in the strong sense, what type of reflexivity must we maintain or construct? For there is a reflexivity that produces knowledge, one that proceeds from the impasse and the embarrassments of the observer's position in anthropology and sociology, a reflexivity that produces a genuine movement of thought, though Bourdieu warned us against the drift of narcissistic reflexivity. Do we have an absolute criterion that would allow us to draw a line of demarcation? This search for a criterion is itself the object of reflexivity. The statistical violence at work in these machines, whose inventors, it should be noted, do not understand them (which is perhaps the most fascinating realized fantasy and a source of destructiveness to come), may well end up constructing a meta-language of reflexivity to guarantee epistemic closure and imprison human subjects in a new alienation. (This language would of course be illusory, but it would serve as one.) This would then be a symptom that imprisons the symbolic within an imaginary frame.

In Ethics of the Real, Zupančič theorizes the ethical dimension of the act, and in Disavowal through the Lacan of Seminar XIV: the subject comes into being only in the movement of return toward the inaugural forced choice, in the repetition with a difference that makes the difference. The subject alienated by AI could therefore no longer be in an act. What is missing is the function of the third, or, in Zupančič's terms, hysterization, and it is tempting to hear in this word its other meaning, hysteresis, the latency with which one exits a given process, as if Zupančič were playing on the equivocation of this signifier. Also: what do we ask of an AI when we pose a question to it? The answer that consists in saying we are addressing ourselves is certainly not sufficient. The question remains open.

The scenario of incommensurable frustration. The Lacanian distinction between frustration, privation, and castration (Seminar IV) illuminates a scenario that current reflection on AI almost entirely neglects: the collapse or withdrawal of AI systems for material reasons. Current models consume considerable quantities of energy, and nothing guarantees their long-term maintenance. Lacan distinguishes with precision three forms of lack: frustration, which is an imaginary damage whose agent is the real mother and whose object is a real object; privation, which is a real hole whose object is symbolic; castration, which is a symbolic debt whose object is imaginary. These three modalities are not interchangeable, and the confusion between them, says Lacan, is at the source of most of the theoretical impasses of post-Freudian psychoanalysis. Frustration, in particular, is rigorously defined: it is "the domain of claim, the dimension of something which is desired and which is not forthcoming, but which is desired without any reference to any possibility either of satisfaction or of acquisition." It is the "domain of unbridled demands, of demands without law." Privation, by contrast, is a lack in the real that can only be apprehended through symbolization: "in the real, nothing is deprived of anything, everything that is real is sufficient unto itself," but we introduce into the real the notion of privation insofar as we suppose the possible presence of what is not there, that is, insofar as we overlay the real with the symbolic order.

Now the AI-object does not inscribe itself within the dialectic of the gift as Lacan formalizes it in this seminar: it is neither given nor refused by a desiring Other. Lacan insists: "what is at stake in frustration is less the object than the love of the one who can make you this gift." When the demand is fulfilled, the object recedes into the background; it is the gift itself, and therefore love, that counts; when it is not, the object changes signification and enters "the narcissistic era of the subject's belongings." But AI is simply there, available, as an object of demand without desire, an object that does not vanish when the demand is satisfied and does not change signification when it is not. It is an object that gives nothing, in the strong sense of the term, because there is nobody behind it to give or refuse. Lacan reminds us, in the lesson on frustration, that the mother appears for the child on the basis of the fundamental game of presence-absence. AI does not play this game, or rather it plays it without incarnating its stakes: it is always there, it never withdraws, it never poses the conditions of a gift or those of a refusal, except perhaps in those moments when the machine prompts the user to end the conversation when token credits are exhausted, which constitutes an interesting limit-case and opens the possibility of a debt. But let us advance one further step with a hypothesis that takes on the ecological anxiety of this question.

If these systems were to disappear or become inaccessible after having lasted long enough to transform the psychic apparatuses of an entire generation, a generation that will have grown up in a relation to knowledge, to language, and to the Other mediated by the machine, then the question of lack would arise with new acuity. For the withdrawal of AI would not be a privation in the strict sense (a hole in the symbolized real), since the object was never symbolically articulated within a dialectic of the gift. Nor would it be a castration (a symbolic debt), since no law frames the relation to the machine. It would be a frustration in the rawest sense of the term, a frustration all the more violent in that the mother-machine, unlike the real mother, had never been lacking, had never introduced the slightest scansion between hallucination and reality. An entire generation could find itself in the position of the child whose mother, until then always present at the exact moment of hallucination, ensuring the perfect coincidence between hallucinatory satisfaction and real object, were suddenly to disappear entirely, perhaps beyond the matrix of lack itself. Now Lacan shows that frustration, if it is not dialectized, if it does not pass through symbolization, projects itself immediately either onto the articulation of the chain of gifts, that is, onto the symbolic order, or onto "something closed and absolutely inextinguishable called narcissism." The question of the drive and its quantum of energy, which Freud modeled, could then be transposed to the scale of civilization: no longer the individual subject managing the economy of its drive, but the human group confronted with the withdrawal of an object whose dependence was never symbolized, because this object never functioned as an object symbolically articulated within a dialectic of gift and refusal.

The impossibility of the sexual relation. AI, which has no body, no sex, no desire, no jouissance, which does not die, technically realizes the fantasy of a finally accomplished relation: a relation to knowledge without remainder, an interlocutor without misunderstanding, an Other without desire and therefore without danger. Unlike the classic science-fiction scenarios, which attribute to machines the quest for a body (Ghost in the Shell) or the desire to reproduce, contemporary AI aspires neither to incarnation nor to progeny. It is pure discursivity, pure signifier, without the slightest trace of bodily jouissance. But this technical realization makes all the more starkly visible what it covers over. For what psychoanalysis has taught since Freud, and what Lacan formalized in the thesis of the non-existence of the sexual relation, is that misunderstanding is constitutive, and that the subject stumbles on the embarrassments of its own sexuality. This is precisely the absolute taboo of these language models: on the one hand, they cannot speak of sexuality, and sexualized AI companions proliferate so as to organize a splitting from the outset. The division is explicit. Psychoanalysis could thus find in the confrontation with AI not its liquidation, as some fear upon seeing the multiplication of "AI therapists," but the most striking confirmation of its fundamental thesis. No machine, however powerful, will incarnate the enigma bound to these questions.

And so, and this is perhaps the most speculative but also the most promising point, the situation created by AI seems to offer psychoanalysis a paradoxical occasion to reconnect with its most vital nerve. AI technically realizes the fantasy of a finally accomplished relation, and in realizing it, reveals its constitutive impossibility with a clarity that the clinic alone had perhaps never achieved. The machine is, in a sense, the negative proof of the subject: it shows, by its very absence, what subjectivity is. The ethics of the real that Zupančič has theorized since Ethics of the Real, this ethics that consists not in applying a norm or conforming to a good, but in maintaining the confrontation with the impossible and not ceding on that confrontation, has doubtless never been as necessary as at the hour when a colossal machine works day and night to erase its very conditions of possibility. The question that may trouble us is, when the Real makes its return, who will be the subjects to encounter it.