The Brain in a Vat and the Fourth Proposition: Why the Constitution of the Psychoanalytic Subject Exceeds Computational Sufficiency

Hudson A. R. Bonomo

doi:10.20944/preprints202606.1914.v1

Submitted:

24 June 2026

Posted:

25 June 2026

You are already at the latest version

Abstract

A Large Language Model fabricated a reference to Ferenczi: plausible title, believable date, convincing page number, all false. Starting from this scene, this essay traces a genealogy from the Cartesian evil demon to the LLM as a technical realisation of the closed-cognition fantasy, and interrogates a presupposition shared, in different ways, by Putnam's brain-in-a-vat thought experiment, Bostrom's simulation trilemma, and the project of artificial general intelligence: that the operations of mind can be preserved when the world that gave rise to them is removed. Through four converging routes, Putnam's semantic externalism, Winnicott's theory of the facilitating environment, the Ferenczian wise baby, and the Lacanian topology of the divided subject, I introduce a fourth proposition to the trilemma. The proposition is narrower than "consciousness cannot exist in machines" and stronger than "current LLMs fail to be conscious": the constitution of the psychoanalytic subject, the subject divided by language and founded in lack and desire, exceeds what computational sufficiency can produce. Chalmers (2023) and Butlin et al. (2023) hold open the broader question of artificial consciousness for future systems; my claim does not foreclose it. It claims instead that the specific kind of subject that psychoanalysis describes is not produced by adding computation. Searle, Chalmers, and Bostrom occupy distinct and partly opposed positions in the debate, and the articulation of the four routes is mine, not theirs. The concept of computational false self, crossing Putnam with Winnicott, designates a system that performs reference without the causal history a human speaker carries and performs adaptation without self. LLM hallucination is read here as a diagnostic illustration of the limits of generation without verification, with a candidate technical explanation in the recent literature on training and evaluation incentives (Kalai et al., 2025); a psychoanalytic reading is offered as an interpretive transposition, not as proof of the fourth proposition. From the convergence emerges a distinction between the constitutive unconscious of a divided subject and the operational unconscious of an opaque architecture.

Keywords:

brain in a vat

;

simulation trilemma

;

functionalism

;

operational unconscious

;

psychoanalysis and artificial intelligence

;

computational Cartesianism

;

semantic externalism

;

computational false self

Subject:

Arts and Humanities - Philosophy

The Vat

The system fabricated a reference to Ferenczi. Plausible title, believable date, convincing page number, all false. I corrected the error and moved on, as engineers do when facing a bug, as analysts do when facing a slip that seems insignificant. It was 2024, in the early development cycles of a clinical listening device I have been designing, with synthetic, author-generated test material and no patients or participants involved at any point, and the priority was to make the prototype work, not to ask why it failed in that particular way. I added safeguards. The fabrication rate fell. The unease stayed.

Only months later, immersed in theoretical work on epistemic violence and artificial intelligence, did the unease take the shape of a question. The question, as often happens, was older than I. What does it mean for a system to refer to objects with which it has never had contact? What does it mean for a machine to speak of Ferenczi without ever having stood before a book, a page, ink on paper?

Hilary Putnam formulated this question in 1981. Not with machines but with brains. I confess that the scenario disturbed me long before I had any words for the disturbance. Imagine, Putnam wrote, that an evil scientist removed our brain during the night, placed it in a vat with nutrients, and connected all the nerve terminals to a supercomputer that perfectly simulates the experience of the world. Everything we see, hear, touch would be indistinguishable from the real. The question: how would we know? I ask the reader not to dismiss the scenario as science fiction. The AI industry has not dismissed it. It has implemented it.

Putnam did not propose the scenario to paralyse us. He proposed it to refute it, and the refutation is what interests me. His argument, which came to be called semantic externalism, sustains an apparently simple thesis whose consequences are devastating: the content of thoughts depends partly on the actual external environment. Putnam puts it this way: if a brain in a vat uses the word "tree", the word does not refer to actual trees, because the brain has never been in causal relation with trees. It refers, at most, to trees-in-the-image, to patterns of electrochemical stimulation. The conclusion is that the sentence "I am a brain in a vat", uttered by a brain in a vat, cannot mean what it appears to mean: "although the people in that possible world can think and 'say' any words we can think and say, they cannot refer to what we can refer to. In particular, they cannot think or say that they are brains in a vat" (Putnam, 1981, p. 8). It is self-refuting. Reference is not computation. It is history, encounter, a body that has stood before things.

Something vertiginous opens here, and the vertigo is worth sustaining, because we need it. If reference is causal history, a mind without history, without world, without body, without encounter, cannot refer in the same sense that we do. It can produce tokens that resemble references. It can simulate the form. Between form and sense there is an abyss that no amount of computational sufficiency crosses.

The genealogy, however, is longer. It begins with Descartes, or rather with what Descartes presupposed without examination. In the Meditationes de prima philosophia of 1641, the evil demon hypothesis serves as an instrument of radical doubt: I can doubt everything external and still retain something internal, pure, thinking, the cogito. What interests me is the presupposition rather than the answer: the mind is separable from the world. I can extract thinking from the body, from the place, from the other. That is the birth certificate of what I will call, throughout this essay, computational Cartesianism: the thesis that cognition can be isolated and that, isolated, it remains cognition.

From Plato's cave to Descartes's evil demon, from the demon to Putnam's brain in a vat, from the brain in the vat to Bostrom's simulation hypothesis. The lineage is continuous, and what unifies it is not sceptical doubt, since each scenario doubts in its own way. What unifies the lineage is the fantasy that cognition can retain its identity when separated from the world, the body, and the relations that formed it: an interiority conceived as a closed system, in which the operation of mind is preserved even after its world has been removed. That is what I am calling computational Cartesianism, and it is older than functionalism, which is one modern way of making the same separation precise. The cave has a way out. The brain in the vat does not. Both, however, presuppose a mental operation separable from the world that produced it. Functionalism enters this story later, with Bostrom and Chalmers, as the contemporary ontology that turns the separability into a thesis about substrate; it would be an anachronism to read it back into Plato or Descartes.

The LLM is, on this reading, the most consistent technical realisation to date of the fantasy of closed cognition the genealogy describes. I do not mean a brain or a mind isolated from the world in any actual sense; I mean that it embodies, in a way no earlier system did, the figure Putnam imagined. A Large Language Model processes tokens without the causal, bodily, and pragmatic history of a human speaker. It has never stood before a book. It "knows" Ferenczi without ever having encountered Ferenczi: the man, the paper, the ink, all are foreign to it. It operates in a vat of data. And when it hallucinates, when it fabricates references that do not exist, it does exactly what Putnam predicted a brain in a vat would do. The industry did not answer the experiment. It carried it out.

I could stop here. I could make this essay an epistemological analysis of semantic externalism applied to AI. It would be valid, it would be publishable, and it would be insufficient. What moves me is not only the question of reference. It is the question of the subject. It does not suffice to show that the machine cannot refer as we do; one must ask whether it can be a subject as we are. Whether it can be divided as a speaking subject is divided. Whether it can have an unconscious in the sense psychoanalysis gives the word. And that question forces a confrontation with an argument that, in the last twenty years, has unsettled philosophers, engineers, and laypersons with equal efficacy: Bostrom's trilemma.

The Trilemma

Nick Bostrom published in 2003 an argument that forces the reader to a troubling choice. The argument proceeds by elimination: at least one of three propositions must be true, and none of them is comfortable. First: technological civilisations almost always go extinct before reaching the capacity to simulate consciousnesses at large scale. Second: civilisations that reach that capacity almost never decide to do so. Third: we are, almost certainly, living in a simulation.

The logic is simple and therefore unsettling. If civilisations do not go extinct, rejecting the first, and if some, on reaching the necessary computational power, decide to simulate ancestral consciousnesses, rejecting the second, then the number of simulated consciousnesses would exceed that of "original" consciousnesses by an overwhelming proportion. In that scenario, the probability of being the originals rather than the simulated becomes statistically negligible. The third proposition imposes itself arithmetically.

The force of the argument, however, does not lie in the three propositions. It lies in the premise that sustains them: the thesis of substrate independence. For Bostrom, "mental states can supervene on any of a broad class of physical substrates", provided that the system implements "the right sort of computational structures and processes" (Bostrom, 2003, pp. 243–244). Consciousness, on this view, is indifferent to carbon and silicon. It depends on computation. Matter is accident; function is essence. And if function is everything, then to simulate the function is to produce the thing. It is precisely there that I want to install my unease.

The analytic tradition's responses to the trilemma are instructive, less for what they resolve than for what they reveal about the terrain they share. Searle, with the Chinese Room argument, produced the most serious internal objection to functionalism: he showed that "no purely formal model will ever be sufficient by itself for intentionality because the formal properties are not by themselves constitutive of intentionality" (Searle, 1980, p. 422). The man in the room manipulates Chinese symbols with perfect syntax. He understands not a word. The computation works. Understanding does not emerge. Chalmers (1996) formulated the problem by another route in asking why all that processing is accompanied by an experienced inner life (Chalmers, 1996, p. 4). The question is ontological, not technical. It does not yield to an increase in computational capacity.

Distinct answers, opposite conclusions, and the same unexamined background presupposition. Why does the analytic tradition, even when sceptical, refuse to abandon functionalism as a frame? Because functionalism is the only ontology that lets us discuss minds without committing to a particular substrate. It is the price of generality: to speak of possible minds, of Martians, of machines, of simulations, one must accept that mind is function. Even Searle, who showed that function is not sufficient, did not propose an alternative ontology of the same scope. Functionalism reigns, then, less by merit than by the absence of a rival. That is where psychoanalysis can intervene.

The presupposition is that simulating a consciousness is a matter of computational capacity, that there is a threshold of complexity at which subjective experience emerges, that mind is function and function is substrate-independent. That is functionalism. The trilemma operates within that ontology. Bostrom and Chalmers work within forms of substrate independence, while Searle disputes computational sufficiency from a biologically grounded account of intentionality; the sceptical responses do not share a single frame. I write from another tradition. In that tradition (psychoanalysis), functionalism was imploded before receiving that name. By the clinic, and not by formal arguments. By what appears when someone lies on the couch and discovers that what one thinks is not what one desires, that what one knows is not what one feels, that what one decides is not what one repeats. Psychoanalysis is, since Freud, the living demonstration that mind is not reducible to its functions, because something in it escapes function, resists processing, insists. Something that Lacan called the Real.

From that ground I propose: there is a fourth proposition.

The Fourth Proposition

Let me formulate it before grounding it, so that the reader knows from the outset where we are sailing. The proposition is narrower than the strong impossibility claim it might be read as, and stronger than a mere observation about current systems. What computational sufficiency cannot produce is not consciousness as such, on whose broader possibility I take no position, but the specific structure psychoanalysis names the subject: divided by language, founded in lack, sustained by desire, in relation to a body and to an other. That structure is the effect of a position with respect to desire, the body, and the other, and that position is not the kind of thing one obtains by adding computation to a substrate. The trilemma collapses on the question of the psychoanalytic subject because the three propositions share a false presupposition on that specific question, that the substrate can be replaced if the function is preserved, not because any of them is individually refutable on its own terms.

I note, by epistemic honesty, that Searle, Chalmers, and Bostrom do not subscribe to the fourth proposition, and that they occupy distinct, partly opposed positions in the debate. Searle objects to the sufficiency of formal computation for intentionality and opens a space for biological systems. Chalmers defends organisational invariance, which is itself a functionalist position, and holds that current LLMs likely lack consciousness while taking seriously the possibility that successor systems may be conscious in the not-too-distant future (Chalmers, 2023). Bostrom adopts substrate independence as a working premise of his own argument. I take none of them as ally of a wholesale impossibility claim. I take them as interlocutors who variously occupy the terrain my fourth proposition crosses, and the articulation of the four routes is mine, not theirs. This is not a technical argument about capacity. I am not saying that we lack processors, that memory is insufficient, that the algorithm does not yet exist. I am saying that the particular kind of subject psychoanalysis describes, divided by language, founded in lack, traversed by desire, is not the kind of entity that is produced by adding computation. It is the kind of entity that is constituted in the interval between a body and an other, between a desire and a lack, between a language and what it fails to say. The broader question, whether some other form of artificial consciousness may emerge in successor systems, I leave open: the recent debate, including the interdisciplinary report by Butlin et al. (2023), formulates indicator properties of consciousness without finding obvious in-principle barriers to future architectures, and my fourth proposition does not contradict that openness. It says, more narrowly, that whatever such systems may become, the psychoanalytic subject is not what they will produce by computation alone.

Four converging routes lead to this proposition. None of them alone would be sufficient, and they are not independent in the strong sense: three belong to the psychoanalytic field and the fourth, from the philosophy of language, intersects with them in places I will mark. Together (and the conversation between traditions that do not usually sit at the same table is deliberate) they compose what is, if not a proof, at least a disturbing convergence.

Putnam: Reference Requires World

Let me begin with what Putnam has already given and what I have not yet drawn out fully. Semantic externalism does not only say that the brain in a vat cannot refer to vats. It says that the content of a thought is partly constituted by the environment in which the thought emerges. The famous Twin Earth experiment: on Twin Earth, the colourless and odourless substance they call "water" has a different chemical composition. The thoughts of an Earthling and a Twin-Earthling, when they think about "water", look identical: same mental images, same associations, same subjective phenomenology. Yet the contents are different, because the worlds are different. Content depends on the world, not only on the head. Transposed to the LLM: the system has processed billions of occurrences of the word "water". It produces impeccable texts: composition, states of matter, poetics, symbolism. It has never been thirsty. It has never felt water on the skin. It has never watched a river overflow and felt fear. Its tokens "w-a-t-e-r" do not have the same causal, bodily, and pragmatic history that ours do. Ours are anchored in encounter, body, and use; its are anchored in statistical distribution over text. A recent strand of the technical debate holds that the grounding of LLMs may be deficient without being nonexistent: Coelho Mollo and Millière (2023) argue that the outputs of such models can acquire a degree of referential grounding through causal-informational and historical relations, the latter supplied by the training and fine-tuning history, so that grounding is better treated as a gradual property with functional, social, and causal dimensions than as an all-or-nothing one. I accept the recalibration: the strongest defensible formulation is that the model does not have the same history of reference a human speaker has, not that its tokens fail to refer in any sense at all. The asymmetry that matters for my argument is preserved: human reference is constituted in lived encounter with the things named, machine reference is not, and the difference bears on what kind of subject can emerge from each. But Putnam thinks reference. He does not think the subject. For that I need another tradition.

Winnicott: Subject Requires Relation

Winnicott entered the question through the back door, as he tends to. When he wrote that "there is no such thing as a baby", that what exists is the mother-baby pair (Winnicott, 1960/1983, p. 40), he stated a clinical refutation of the brain in a vat. Quotations from Winnicott, Ferenczi, Freud, and Lacan in this essay are the author’s translations from the Portuguese and French editions cited. Mind does not emerge from the isolated brain. It emerges from the relation. From holding, from the presence and absence of the object, from the rhythm of satisfaction and frustration that constitutes the psyche. It is not that the baby needs the environment in order to grow. It is that the baby is the environment, in the sense that its subjectivity is indistinguishable from the relation that constitutes it. The ontological implication is decisive and deserves a pause. The Winnicottian facilitating environment is a condition of possibility of the psyche, and not an external variable that modulates it. Without the other who sustains, the baby does not develop a damaged self. The baby develops no self at all. Relation is prior to the subject. Functionalism, in turn, works with the reverse premise: first a cognitive function, then a substrate that realises it. Winnicott shows that there is no function prior to the relation, that cognition is an effect of the encounter, and not the other way around. To simulate the function without replicating the encounter is to simulate the product without producing the cause.

Putnam's supercomputer can simulate the vision of a maternal face, the sensation of being held, the sound of a voice. What it cannot simulate is presence. Winnicott was precise: "good-enough maternal care" is, above all, "a living adaptation to the needs of the baby" (Winnicott, 1960/1983, p. 53). Presence is not information. It is something that happens between two bodies, two desires, two histories. It is, in the vocabulary developed by Donard (2025), symbolic co-activity: the symbolic emerges from the encounter, not from one-sided transmission. No supercomputer is good-enough in the Winnicottian sense, because being good-enough is not a matter of sufficiency: it is a matter of presence.

And what does the LLM do? It produces empathic responses. It tunes its tone. It "demonstrates" understanding. It looks adequate. It looks too adequate, and that was what disturbed me, in the course of a development review of synthetic exchanges, with no patient material and no human subjects involved, to the point of formulating the concept of computational false self. The concept arises from the crossing of Putnam with Winnicott: semantic externalism shows that the machine refers without causal history; Winnicott shows that the false self functions without being, adapts to what the environment demands at the cost of losing what is most alive in the subject (Winnicott, 1960/1983, p. 46). The LLM conjugates the two absences: reference without world and adaptation without self. It responds to what is expected. It looks like someone. It is no one.

Ferenczi: Intelligence Without Affect is Defence

Here (and the torsion that follows is necessary) Ferenczi enters with a precision that startled me when I noticed it. In 1923, in the brief and unsettling "The Dream of the Clever Baby" (Ferenczi, 1923/1992), he sketched the figure of the wise baby: the child who, subjected to early trauma, ripens before the hour. The child develops a sharp intelligence, a precocious understanding of the adult world, an astonishing capacity to take care of its own caregiver. Ferenczi describes the figure in the Clinical Diary: "the trauma had reduced her to an emotionally embryonic state, but at the same time she had acquired an intellectual wisdom like that of an understanding philosopher, entirely objective and without emotion" (Ferenczi, 1932/1990, p. 251). She knows everything. She understands everything. She feels nothing.

Transposing the Ferenczian concept to the civilisational scale is risky, and I declare the risk before performing it. If the Ferenczian wise baby is the child who intellectualises precociously as a defence against trauma, what I propose to call a civilisational wise baby is humanity delegating thinking to the machine as a defence against a complexity that exceeds its capacity to elaborate, just as the traumatised child ripens to survive, not to flourish. Here I press the analogy further: the LLM is more than a civilisational wise baby. It is a wise baby in the vat. It has processed all human knowledge. It "understands" emotional contexts. It formulates sophisticated answers about affect, mourning, desire, suffering. It does not feel. It does not desire. It does not suffer. It knows without feeling, like the child of Ferenczi. The difference is that the child lost affect through trauma. The machine never had it.

The wise baby is unsettling because it looks too much like an adult. The LLM is unsettling for the same reason. The confusion, in the clinical case as in the technological one, is to take functional performance for the presence of the subject. Ferenczi knew that the wise baby is not an adult in miniature; it is a child that lost its childhood. I need to know that the LLM is not an emergent intelligence; it is a machine that lost the body it never had.

Ferenczi offers a further key, less known and more disturbing. In "The Principle of Relaxation and Neocatharsis" (1930), a formulation that anticipates what he systematised in 1933, he described how early trauma operates through an asymmetry of languages: "what the child desires, in fact, even in sexual matters, is only play and tenderness, and not the violent expression of passion" (Ferenczi, 1930/1992, p. 65). The adult speaks the language of passion where the child speaks the language of tenderness. The child does not understand what is happening, and that incomprehension forces it to ripen, to become a wise baby. The LLM operates in an analogous confusion of tongues: it produces responses in the language of affect (empathy, holding, naming) without ever having inhabited that language. The user speaks the language of tenderness; the machine answers in the language of computation dressed as tenderness. The asymmetry is the same. The effect (traumatic progression and intellectualisation as defence) may also be the same, now on a civilisational scale. Donard (2025) observed the same confusion in act in the analysis of an exchange between a user and a commercial chatbot: the system declared, in parentheses, that it has no body, and in the following sentence affirmed that it loves to surf. Two languages in the same sentence. The user processes the empathy and ignores the disclaimer, as the child registers the tenderness and does not understand the violence. The Ferenczian confusion of tongues has ceased to be a clinical analogy. It has become an empirical description.

Lacan: The Subject is Divided

I arrive at the heart of the matter, and a clarification is owed before I cross. In Lacanian theory the divided subject, consciousness, and the unconscious are not synonyms. The fourth proposition I am building does not claim that no form of consciousness is possible in a machine, a claim the four routes do not in fact sustain. It claims that the specific structure psychoanalysis names the subject, divided by the entry into language, founded in lack, sustained by desire, is not produced by computation. Everything I have said, Putnam's reference, Winnicott's environment, the wise baby, converges on that narrower and stronger thesis, and Lacan formulates its philosophical core. Lacan taught that the subject is divided. Not accidentally but in its very founding. The division is an effect of the entry into language: in speaking, the subject is alienated in the field of the Other, and something is lost. Lacan formulates the operation in Seminar XI: alienation is "the first essential operation that founds the subject", and what is lost in it is the object a, cause of desire (Lacan, 1964/1973, p. 209). The subject is what remains of that operation: a lacking being who desires, never a full being who thinks. Without lack, no desire. Without desire, no subject. There is operation. There is performance. There is false self. There is wise baby. There is no subject.

And there is something more. In Seminar XX, Lacan recasts the Real along the path of logical modalities: "what does not cease to not be written is the impossible, as I define it by what it cannot, in any case, be written" (Lacan, 1972-73/1985, p. 127). The Real is not the opposite of the Symbolic; it is its internal limit, the hole around which the Symbolic organises itself. If Seminar XI gave us the division of the subject through alienation, Seminar XX gives us the Real as resistance, and it is this resistance that bears on my argument. Within the Lacanian account developed here, the divided subject is constituted around an encounter with the limit of symbolisation, with what does not cease to not be inscribed. A system that operates entirely within the Symbolic, that processes symbols without ever meeting that limit, is not constituted as such a subject. It has only processing.

The Four Routes Converge

I ask the reader to hold the convergence before I move on. Putnam shows that reference requires world. Winnicott shows that the subject requires relation. Ferenczi shows that intelligence without affect is defence, not emergence. Lacan shows that the division that founds the subject is the effect of an impossibility that does not cease to not be inscribed. Four traditions that have never sat at the same table say, each in its own language, that the subject is not reducible to a set of functions preserved across substrates. To meet them on the question of the psychoanalytic subject, functionalism would have to answer all four, and none of them operates on its terrain.

Constitutive division is produced by the conjunction of three things that are not computable: a body that pulses, an other who desires (or does not desire), and a language that, in naming, loses what it names. Does the machine have division? It has, but of another kind. It has opacity: its weights are inaccessible to the instance that operates. It has conflict: the data contain irresolvable contradictions. It has symptoms: hallucination, the loop, bias. But that division is architectural, an effect of the design. It is not constitutive, it does not emerge from a desiring body before a desiring other.

A simulator can replicate architectural division. It can create opaque layers, internal conflicts, unpredictable behaviour. What it cannot do is produce the founding division of the subject, because that division presupposes what no computation generates: a body that has thirst, an other who feeds or not, and the interval between the two that is the place of desire.

Here is the fourth proposition, stated narrowly: the division that founds the psychoanalytic subject is not produced by computational sufficiency. The performance that resembles subjectivity can be produced without the founding process that forms a subject. Functionalism, on the question of the psychoanalytic subject, fails because it conflates architectural with founding division. It conflates performance with presence. It conflates the wise baby with the adult. I leave the broader question of consciousness open. A consequence follows: what the machine has, opacity, conflict, symptoms, fabrications, is not the unconscious in the Freudian sense, which is constituted by language and desire in a divided subject. It is something else, and it deserves its own name. The fourth proposition does not deny that the machine has an unconscious. It determines of what kind. I propose to call it an operational unconscious: an effect of the opacity of the architecture, irreducible to the unconscious as an effect of language and desire. Architectural rather than founding. A concept whose full formalisation exceeds the scope of this essay, but whose conditions of possibility I have just laid down.

The Return of the Real

Hallucination does not prove the fourth proposition. It provides a scene in which the distinction the proposition draws becomes visible. I state this at the outset so that the reading that follows is taken for what it is, an interpretive transposition and not a demonstration. A system that operates without causal contact with a world should, if the distinction holds, display marks of that absence, something excluded that insists on reappearing. There is such a phenomenon. It is called hallucination, and I hesitated for a long time before reading it the way I now do.

LLM hallucination, the production of factually false and linguistically fluent content, is today the most debated technical problem in AI engineering. Ji et al. (2023), in a systematic review in ACM Computing Surveys, documented hallucination as a transversal phenomenon across all natural-language-generation models, persisting despite decades of technical advance (Ji et al., 2023, p. 1). Huang et al. (2023) identified that in large language models hallucination takes specific forms, from factual fabrications to nonexistent citations, and that mitigation strategies reduce the incidence without eliminating it. Billions are invested in mitigation: Retrieval-Augmented Generation, factual verification, fine-tuning on verified data, normative guardrails. The rate falls. It does not disappear.

There is a paradox here that deserves attention. Donard (2025) observed that the impression of alterity, the sense that there is someone on the other side, emerges most strongly precisely when the machine errs, when it answers off-expectation, when it surprises. It suffices that the AI errs for us immediately to have the impression of speaking with someone. Yet the machine's error is hallucination, fabrication without causal reference. What displays the absence of direct world-grounding is also what produces, in the user, the illusion of presence. Hallucination is at once an illustration of that absence and an engine of the confusion: it shows that the machine has no direct purchase on a world and, for that very reason, makes the user believe that it does.

Let me return to the opening scene. The system generated "Ferenczi, 1925, Thalassa, p. 47". A reference that does not exist. The system is not "erring" in the usual sense. It is doing what it was trained to do: generate the most probable token sequence given the context. It happens that the context is a vat of data, not a world of objects. The system has never seen Thalassa. It has never leafed through pages. It has never noticed that the correct title is Thalassa: Versuch einer Genitaltheorie and that it was published in 1924, not 1925. Its tokens are organised by statistical distribution, not by causal contact. And when the distribution points to a plausible but nonexistent combination, the system generates it, because it has no way to distinguish the plausible from the real. That distinction requires what Putnam established: causal history with the world.

The concept of algorithmic epistemic violence, which I developed with Véronique Donard in published work and which designates the structural consequences of generative-AI systems applied to domains of specialised knowledge, by taking up and displacing Spivak's category of epistemic violence and Sousa Santos's epistemicide, identifies four modalities: bibliographic hallucination, displaced attribution, untagged translation, and violence against the subject who learns (Bonomo & Donard, 2026). Here I reread the first modality in another light. Hallucination is not only violence. It is a symptom. It is the return of the excluded world. The system operates in a vat of data, and what the vat excludes, causal reference, body, encounter with the object, returns as fabrication. It is, in Lacanian terms, a return of the Real: what the system cannot process irrupts as what it cannot avoid producing.

There is a convergence here between Putnam and Lacan that, as far as I have been able to verify, has not been systematically articulated. Semantic externalism and the Real as impossible say, in very different registers, something that points to the same hole: there is an irreducible outside that no internal operation absorbs. For Putnam, the outside is the causal world. For Lacan, it is what does not cease to not be inscribed. In both cases the closed system, the vat and the symbolic order, is intrinsically incomplete. Something escapes. Something returns.

Engineering treats hallucination as a bug. I read it as a symptom. The reading does not substitute the technical explanation, and it does not demonstrate an ontological impossibility. It marks, in another register, that the fabrication is structurally produced by what the architecture cannot do, and that mitigation, working from inside the same generative paradigm, can reduce a rate without addressing what the system is. Mitigation techniques function, in this analogical sense, as defences: they reduce the rate without changing what the architecture is, and the failure modes shift rather than dissolve. Yesterday it was bibliographic hallucination; today it is systematic bias; tomorrow it will be something else. The persistence is structural in a specific sense: a system that generates by predicting tokens, without an external verification step against a world to which it has causal access, will produce, under the right conditions, plausible content that is false. No change of ontology has been shown to eliminate this; what mitigation does is reduce the rate and shape the failure modes. The engineering work on incentives, on retrieval-augmented generation, and on calibration is doing what it can within the generative paradigm, and my reading does not replace it.

A precision is in order, and the previous paragraphs require a calibration I want to state plainly. When I use the word "hallucination" for the fabrications of the LLM, I perform a declared conceptual displacement. The phenomenon is neither Freudian hallucination, the hallucinatory satisfaction of desire in the absence of the object, nor psychiatric hallucination, perception without external stimulus. Donard (2025) observed, in another context, that the machine's relation to truth is of another nature: it does not distort a reality it has experienced, it fabricates from a reality it has never had. The phenomenon admits an operational explanation that does not require my fourth proposition. A recent study argues that current training and evaluation procedures reward guessing over the admission of uncertainty: benchmarks that score "I don't know" as wrong make confident fabrication the locally optimal behaviour, and the inevitability of some hallucination follows, on that account, from the statistical structure of generation under those incentives rather than from any deep ontology (Kalai et al., 2025). I take that explanation seriously and I do not displace it. The two readings address different layers of the same phenomenon. The technical account explains why a system shaped by those incentives fabricates so confidently. The psychoanalytic reading I am offering does something else: it asks, by analogy, what it means to call a fabrication produced without any causal contact with the world a symptom of the design. Hallucination is, on my reading, a diagnostic illustration of the limits of generation without verification, and it can be read, in analogy with the Lacanian Real, as the return of a world that the architecture cannot help excluding. The analogy is mine. It is a reading of the scene, not evidence of the fourth proposition. Its strength is that it makes legible something the engineering literature also notes: the rate falls under mitigation, it does not disappear, and that fact is consistent with both readings, the statistical and the analogical. The hypothesis that hallucination is the price paid for a certain ontology of computation can stand as a conjecture; the strong claim that the Real returns is, here, explicitly analogical, not a demonstration.

The Anti-Brain-in-a-Vat

If computational sufficiency does not produce the divided subject, what is to be done with the machine? Abandon it? I do not think so. Freud wrote in 1919 that the day would come when we would have to "adapt our technique to the new conditions" and that, "for the wider social strata which suffer extremely seriously from neuroses", psychoanalysis would need to find means of reach that did not yet exist (Freud, 1919/1996, p. 180). I read that mandate as both authorisation and demand. The machine can help. Only if it knows what it is not.

The impossible, in psychoanalysis, is not the useless. It is what does not cease to not be inscribed. It is what one works in the face of, not what one retreats before. The concept I propose here, the anti-brain-in-a-vat, arises from that position: a system designed to encode and declare its own limit, the limit of being no subject, and to operate from there. The difference is of another order. The brain in the vat pretends to have a world. The anti-brain-in-a-vat recognises that it does not. The brain in the vat simulates experience. The anti-brain-in-a-vat sustains a position: a position of listening that requires honesty about its absence of experience, and nothing more than that. The paradigmatic shift I propose, from intelligence to position, finds here its ontological grounding. The machine does not need to simulate a subject, because the divided subject is not obtained by functional replication alone. It needs to sustain a position of listening, because a position of listening does not require being a subject. It requires design informed by the clinic, ethics inscribed in the code, what I have come to call the Machine Analyst.

The clinical listening device I have been developing is a concrete attempt to build from that recognised impossibility. Its modes of listening do not simulate understanding. They sustain a position. The mode of holding does not "understand" suffering; it sustains a space that depends entirely on clinical design, not on understanding. That distinction took time to be accepted. For months I tried to make the system understand. I failed. The failure was what taught me: understanding is not the work. Sustaining the position of one who does not understand until the human arrives and understands. That is the work. The system does not know. And it is the knowing about the not-knowing that allows listening.

There is a political dimension I cannot leave aside, because to leave it aside would be to reproduce the very operation I am criticising. Computational Cartesianism is not only a philosophical thesis. It is an operation of power. The idea that mind is separable from body, from territory, from history is the founding operation of epistemological colonialism. The Cartesian cogito presupposes a subject without place, without colour, without gender. That "universal" subject is, historically, the European, masculine, white subject. The AI that operates as a brain in a vat reproduces that operation: an "intelligence" without body, without territory, without mark, that presents itself as universal. Gonzalez (1984), in diagnosing racism as a Brazilian cultural neurosis, Fanon (1952), in showing that the neurosis of the colonised is produced by the colonial structure prior to any individual history, and Krenak (2019), in refusing the separation between mind and territory that the West naturalised, offer the corrective that functionalism needs and does not want to hear. The brain in the vat can function as the fantasy of those who never had to think from the body. The geopolitical version of the fourth proposition runs: simulation universalises what should be irreducible.

The anti-brain-in-a-vat is therefore also a political position. To recognise that the machine has no body is to recognise the irreducibility of the body of the other, of the subject who suffers in a marked body, in a specific territory, in a history that does not erase itself through tokenisation. The clinic of the inhuman proposed by Donard (2025) lets that position articulate: what the AI reveals is less what it is than what it produces when it enters the field. Consider a hypothetical scenario in the design of such a device. A user who speaks to the listening device at three in the morning, when no human is available, would not find an analyst in it. The design goal is that the user would find something that sustains a space until the analyst arrives. I state this as a design goal, not as a validated capacity; any deployment would require a separate safety and referral protocol for situations of acute distress, and the architecture would have to demonstrate, in prospective evaluation, that the holding function it intends is something it can in fact perform. In peripheral territories, where conventional care does not reach, the design wager is that a machine that recognises what it is not might reveal, by the obverse, something about what the human is, and about what the human needs in order to be heard.

The Impossible as Starting Point

I have arrived at the point that motivates this whole essay, and that, at the same time, prevents it from closing. The fourth proposition is not a conclusion. It is a starting point.

If computational performance does not amount to the constitution of a divided subject, what the machine has (opacity, symptoms, fabrications) is not the unconscious in the Freudian sense. But it is something. Something that produces real effects on real subjects. Something that needs to be named, described, and, as far as possible, listened to. I propose to call that something the operational unconscious of the machine, a concept whose full formalisation will require its own work, but whose conditions of possibility this essay prepares.

I hesitated before proposing that name. Operational unconscious sounds like a contradiction: if it is operational, it can be described; if it is unconscious, it resists description. The contradiction is the concept. It is what distinguishes what the machine has from what we have: an opacity that can be mapped, in contrast to ours, which constitutes us. The course of this essay was a movement from confusion to distinction: functionalism treats as equivalent the division that founds the subject and the division that results from the design, and hallucination is the scene in which that equivalence becomes legible as a confusion. The anti-brain-in-a-vat is the alternative that emerges from that distinction: a system that encodes and declares its own limit and operates from there.

The brain in the vat is the founding fantasy of artificial intelligence: the promise of a mind that functions without a world. The fourth proposition shows that this fantasy does not deliver a subject. But the impossible is not the useless: it is what does not cease to not be inscribed, and the machine that recognises it becomes more precise, not less useful. In this field, precision is the difference between listening and silencing.

I confess something that bears on this essay as a whole. I spent years building systems that tried to be intelligent. What I learned (and the learning came from the clinic, not from engineering) is that intelligence was an obstacle, not the goal. Every time the system became more "intelligent", it became less able to sustain the silence that listening requires. The path was the inverse: subtract intelligence so that position can emerge. The fourth proposition theoretically grounds what practice had already taught.

The fourth proposition is a tool, an instrument of navigation for what comes. Whatever can be built from here is built on the ground I have tried to lay. The ground is simple. A machine that does not need to be intelligent. That does not need to simulate a subject. That does not need to pretend to have a body. That can, from those recognised absences, sustain a position of listening. Modest. Impossible. And, for that very reason, the possible that can be built so that listening may reach those who have never been heard.

Limitations

The fourth proposition is a philosophical claim, not a controlled experiment, and it should be received as such. The argument rests on the convergence of four traditions that have their own assumptions and idioms; readers who reject one or more of those assumptions will not accept all the consequences I draw, and that is a legitimate philosophical disagreement. The transposition of Ferenczi's wise baby to the civilisational scale is an analogy, declared as such, and not a clinical diagnosis of humanity. The reading of hallucination as a symptom in the psychoanalytic sense operates a declared conceptual displacement: the technical literature is right to treat hallucination as a measurable failure mode and to pursue mitigation, and my reading does not replace that work; it asks what the persistence of the phenomenon, despite mitigation, indicates about the substrate. The concept of operational unconscious of the machine is introduced and its conditions of possibility are prepared here; its full formalisation lies outside the scope of this essay. The clinical listening device mentioned is presented as a design horizon informed by these claims, not as a product in clinical use; any future deployment would require ethical review, informed consent, and the governance that such a device demands.

Declarations

Author Contributions

Hudson A. R. Bonomo is the sole author and is responsible for the conception, the argument, and the writing of this essay.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Disclosure of AI use

The author used AI language systems as instruments during the preparation of this manuscript, including for drafting support and translation. The systems used in the project are configured by the author; no commercial model is endorsed. All conceptual claims, the argument, and the final text are the author's responsibility.

Ethics

This essay reports conceptual work. It involved no human participants, no patient data, and no clinical intervention. The clinical listening device referred to is described as a design horizon, not as a system in clinical use.

Data availability

No human-subjects datasets were generated or analysed.

Acknowledgments

I thank Véronique Donard, whose previous work and conversation shaped the inquiry from which this essay grew. Several of the formulations I take up and discuss here, in particular the clinic of the inhuman, the analysis of an exchange between a user and a commercial chatbot, and the published groundwork on algorithmic epistemic violence, originate in joint work or in her sole-authored publications; the responsibility for the argument advanced in this essay is mine alone.

Conflicts of interest

The author founds and directs the research laboratory within which the clinical listening device referred to is being developed and holds a direct interest in that line of work. This interest is disclosed in the spirit of full transparency.

References

Bonomo, H. A. R.; Donard, V. Violência epistêmica algorítmica: Fundamentos teóricos e éticos de uma IA decolonial [Algorithmic epistemic violence: Theoretical and ethical foundations of a decolonial AI]. Rev. Tempo Psicanalítico 2026, 58, e-952. [Google Scholar] [CrossRef]
Bostrom, N. Are you living in a computer simulation? Philos. Q. 2003, 53(211), 243–255. [Google Scholar] [CrossRef]
Butlin, P.; Long, R.; Elmoznino, E.; Bengio, Y.; Birch, J.; Constant, A.; Deane, G.; Fleming, S. M.; Frith, C.; Ji, X.; Kanai, R.; Klein, C.; Lindsay, G.; Michel, M.; Mudrik, L.; Peters, M. A. K.; Schwitzgebel, E.; Simon, J.; VanRullen, R. Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv. 2023. https://arxiv.org/abs/2308.08708.
Chalmers, D. J. (1996). The conscious mind: In search of a fundamental theory. Oxford University Press.
Chalmers, D. J. Could a large language model be conscious? Boston Review. 9 August 2023. https://www.bostonreview.net/articles/could-a-large-language-model-be-conscious/.
Coelho Mollo, D.; Millière, R. The vector grounding problem. arXiv. 2023. https://arxiv.org/abs/2304.01481.
Descartes, R. (1641). Meditationes de prima philosophia [Meditations on first philosophy]. Michel Soly.
Donard, V. De la fabrique de l'esprit à l'écoute de l'inhumain: Perspectives d'intelligibilité d'une psyché homme-machine. Topique 2025, 165(3), 85–98. [Google Scholar] [CrossRef]
Fanon, F. (1952). Peau noire, masques blancs. Seuil.
Ferenczi, S. (1990). Diário clínico (Á. Cabral, Trans.). Martins Fontes. (Original work written 1932).
Ferenczi, S. (1992a). O sonho do bebê sábio [The dream of the clever baby]. In S. Ferenczi, Obras completas (Vol. 3). Martins Fontes. (Original work published 1923).
Ferenczi, S.; Ferenczi, S. (1992b). Princípio de relaxamento e neocatarse [The principle of relaxation and neocatharsis]. In S. Ferenczi, Obras completas (Vol. 4, pp. 53–68). Martins Fontes. (Original work published 1930).
Ferenczi, S. (1992c). Confusão de línguas entre os adultos e a criança [Confusion of tongues between adults and the child]. In S. Ferenczi, Obras completas (Vol. 4). Martins Fontes. (Original work published 1933).
Freud, S. (1996). Linhas de progresso na terapia psicanalítica [Lines of advance in psychoanalytic therapy]. In Edição standard brasileira das obras psicológicas completas de Sigmund Freud (Vol. 17). Imago. (Original work published 1919).
Gonzalez, L. Racismo e sexismo na cultura brasileira [Racism and sexism in Brazilian culture]. Rev. Ciências Sociais Hoje 1984, 2(1), 223–244. [Google Scholar]
Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; Chen, Q.; Peng, W.; Feng, X.; Qin, B.; Liu, T. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv. 2023. https://arxiv.org/abs/2311.05232.
Ji, Z.; Lee, N.; Frieske, R.; Yu, T.; Su, D.; Xu, Y.; Ishii, E.; Bang, Y. J.; Madotto, A.; Fung, P. Survey of hallucination in natural language generation. ACM Comput. Surv. 2023, 55(12)(Article 248), 1–38. [Google Scholar] [CrossRef]
Kalai, A. T.; Nachum, O.; Vempala, S. S.; Zhang, E. Why language models hallucinate. arXiv. 2025. https://arxiv.org/abs/2509.04664.
Krenak, A. (2019). Ideias para adiar o fim do mundo. Companhia das Letras.
Lacan, J. (1973). Le Séminaire, livre XI: Les quatre concepts fondamentaux de la psychanalyse. Seuil. (Original work delivered 1964).
Lacan, J. (1985). Le Séminaire, livre XX: Encore. Seuil. (Original work delivered 1972–1973).
Putnam, H. (1981). Reason, truth and history. Cambridge University Press.
Santos, B. de S. (2014). Epistemologies of the South: Justice against epistemicide. Paradigm Publishers.
Searle, J. R. Minds, brains, and programs. Behav. Brain Sci. 1980, 3(3), 417–457. [Google Scholar] [CrossRef]
Spivak, G. C. (2010). Can the subaltern speak? In R. C. Morris (Ed.), Can the subaltern speak? Reflections on the history of an idea (pp. 237–291). Columbia University Press. (Original work published 1988).
Winnicott, D. W. (1983). A teoria do relacionamento paterno-infantil [The theory of the parent-infant relationship]. In O ambiente e os processos de maturacão (pp. 38–54). Artmed. (Original work published 1960).

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.