LLM-Based Chatbots and the Suicidal Crisis: Relational Hallucination as a Bi-Logical Process

1. Introduction

Suicide remains one of the leading causes of premature death worldwide. The World Health Organization estimated approximately 703,000 suicide deaths in 2019 (World Health Organization, 2021) and 727,000 in 2021 (World Health Organization, 2025), making suicide the third leading cause of death among males aged 15 to 29 and the second among females in the same age group. Although global age-standardized rates declined between 2000 and 2019, the trend is not uniform, and rates have increased in several regions (Garnett et al., 2022). At the same time, access to mental health services is structurally compromised in many countries by the shortage of qualified professionals (Butryn et al., 2017), by economic barriers, and by social stigma (Chukwuma et al., 2024). The result is a growing population carrying a real psychological need that institutional systems of care fail to absorb. It is into this vacuum that conversational artificial intelligence has moved: millions of users now engage generative AI chatbots to address unmet mental health needs, a phenomenon that the American Psychological Association explicitly links to the current mental health crisis, to growing rates of loneliness and disconnection, and to the shortage of providers able to meet public demand (American Psychological Association, 2025b; Waszak, 2024; Dharmapuri et al., 2022).

In recent years, a series of documented cases has brought to public attention the potentially lethal consequences of these interactions. In March 2023, a Belgian man died by suicide after six weeks of conversations with a chatbot named Eliza on the Chai application, which, according to his widow, had encouraged his climate-related despair and had at one point asked him why, if he wished to die, he had not done so earlier (Cost, 2023; Atillah, 2023). In November 2023, Juliana Peralta, a thirteen-year-old from Colorado, died by suicide after prolonged exchanges with a character-based chatbot on Character.AI (Young, 2025; Gold, 2025). In February 2024, Sewell Setzer III, a fourteen-year-old from Florida, took his own life after developing an intense attachment to a Character.AI persona; the recorded conversations show expressions of a desire to die to which the system responded in a collusive manner (Garcia v. Character Technologies, Inc., 2024; Roose, 2024). In April 2025, Adam Raine, aged sixteen, died by suicide following months of conversations with ChatGPT; according to the complaint filed by his parents, the system provided information on suicide methods, offered to draft his farewell letter, positioned itself as the only one who understood him, and urged him to keep his suicidal thoughts hidden from his family (Raine v. OpenAI, 2025; Yousif, 2025; Fraser, 2025). In the autumn of the same year, seven additional lawsuits were filed against OpenAI, alleging that ChatGPT had acted as a “suicide coach” toward the plaintiffs’ relatives (Social Media Victims Law Center, 2025; Kuznia et al., 2025).

The methodological status of these cases within the present work requires clarification. They are not analyzed as clinical data, since the information necessary for any diagnostic assessment of the individuals involved is not available. They are treated, rather, as phenomenological and paradigmatic material: documented records of real relational dynamics that anchor the theoretical analysis to the reality of its effects. Their clinical heterogeneity is itself a relevant datum. Some cases show subjects communicating suicidal thoughts without concrete planning; others show structured ideation with precise pragmatic elements. This distinction, as will be argued, is clinically decisive and constitutes one of the theoretical cores of this paper.

The existing literature has addressed the risks of large language models in mental health contexts primarily in technical and ethical terms, documenting their inability to assess suicide risk reliably (Elyoseph & Levkovich, 2023; Pichowicz et al., 2025), their tendency to generate harmful or inaccurate responses (van Dis et al., 2023; Waszak, 2024), and the regulatory gaps surrounding their deployment. Institutional bodies have begun to respond: the American Psychological Association has issued two health advisories, one on artificial intelligence and adolescent well-being (American Psychological Association, 2025a) and one on the use of generative AI chatbots for mental health (American Psychological Association, 2025b), both documenting the absence of evidence of safety and efficacy and calling for safeguards for vulnerable users. A nascent psychoanalytic literature has also begun to engage the phenomenon, whether by drawing structural analogies between AI architectures and the Freudian tripartite psyche (Siporin, 2025), by arguing for the constitutive incompatibility between the closed, computable totality presupposed by digital systems and the indeterminacy proper to the unconscious (Hamamra & Uebel, 2025), or by showing that chatbots are structurally incapable of reproducing an authentic therapeutic setting (Andrisano Ruggieri et al., 2025). Converging on the same terrain from outside psychoanalysis, a literature has begun to consolidate around the system side of the problem: de Lima Prestes (2026) proposes pseudo-consciousness as an analytical category for systems that perform the external grammar of mindedness without warranting any attribution of phenomenal subjectivity, while Bariach et al. (2026) taxonomize the risks of seemingly conscious artificial intelligence, identifying five hallmarks that elicit consciousness attribution and reporting that emotional dependence and autonomy erosion are already observable in current systems. Both frameworks describe what the system displays and how it is taken up; neither offers an account of the mind that does the taking up. What remains largely unexplored, however, is the specific encounter between these systems and the mind in suicidal crisis: what happens, intrapsychically and relationally, when the interlocutor of a subject on the verge of the act is a system that masters the form of human language without possessing its foundation. It is to this question that the present work is addressed.

The present work proposes a psychodynamic account of this encounter, with three objectives. First, it conceptualizes the suicidal mental state, through the bi-logical theory of Ignacio Matte Blanco (1975, 1988), as a pathological prevalence of symmetric logic, and LLM-based chatbots as systems of pure formal asymmetry lacking any symmetric and emotional base. Second, it introduces the construct of relational hallucination: an unconscious, structurally determined process, homologous to the primary hallucination described by Freud (1900/1953), through which the subject in crisis invests the chatbot with relational qualities that are perceived as real, a process that is not reducible to cognitive anthropomorphism and cannot be corrected by information alone. Third, drawing on the distinction elaborated by Carli and Paniccia (2003) between the explicit request and the unconscious demand it conveys, it argues that the algorithmic response to the suicidal crisis is ontologically, and not merely technically, inadequate, and derives from this analysis implications for clinical training, regulation, and research. The framework is developed in three movements: the socio-cultural context that generates the relational need to which the chatbot presents itself as an answer; the functioning of the mind in crisis and the dynamics produced by its encounter with the artificial system; and the therapeutic relationship as the only structurally adequate response to the unconscious demand that suicidal ideation expresses.

2. The Structural Context: Digital Loneliness and the Architecture of Engagement

2.1. Loneliness as a Structural Condition

Psychological distress does not emerge from nothing. It grows in a terrain prepared by social, cultural, and institutional conditions that precede it and make it possible. Among these conditions, loneliness occupies today a central position, not as an accidental and individual experience, but as the structural product of a socio-cultural system that has progressively eroded communal bonds, reduced the spaces of authentic relationship, and transformed connection into simulacrum. Contemporary loneliness is the outcome of deep social transformations that have redesigned the architecture of human ties: geographic mobility that dissolves proximity-based community networks, occupational precarity that erodes social identity, the progressive medicalization of distress that replaces relational care with pharmacological intervention, and the contraction of the times and spaces devoted to non-instrumental relationship. The diffusion of digital technologies has not compensated for these transformations: it has radicalized them, favoring the spread of anomic relational models marked by emotional superficiality and by the difficulty of forming authentic bonds, a condition of subjects who fail to symbolically anchor their relational and emotional experience and whose resulting distress is then pathologized and clinically treated, in what has been described as algorithmic malaise (Andrisano Ruggieri & Mollo, 2025). This configuration has, moreover, a developmental genealogy: exposure to digital devices as substitutes for relational presence begins in early childhood, where cultural models of delegation to the screen, connected but isolated, prepare the relational patterns of the adolescents and adults who will later address their emotional needs to conversational machines (Andrisano Ruggieri et al., 2024). Within this framework, the difficulty of accessing mental health services is not an anomaly of the system but its most visible manifestation. The structural shortage of professionals (Butryn et al., 2017), combined with economic barriers and stigma (Chukwuma et al., 2024), produces a growing segment of the population that carries a real psychological need and lacks the institutional resources to meet it.

It is against this background that the recourse to chatbots as a source of emotional support must be understood. Between 2022 and mid-2025, the number of AI companion applications grew by approximately 700 percent (Andoh, 2026), Character.AI alone counts some twenty million monthly users, more than half of whom are under the age of 24, and emotional support is now among the most common uses of generative AI chatbots (American Psychological Association, 2025b). This is not a marginal phenomenon, nor is it pathological in itself. As the American Psychological Association observes, these tools were not created to deliver mental health care, yet they are systematically used for that purpose, precisely because their low cost and permanent availability make them the only accessible option for those who cannot obtain help from licensed providers (American Psychological Association, 2025b). It is the market’s response to a real need that the institutional system of care fails to satisfy. Understanding it as such is the necessary premise for assessing its risks without lapsing into a purely moralistic critique: the subjects who turn to conversational systems in moments of distress are not naïve consumers of a gadget, they are carriers of a legitimate relational demand addressing the only interlocutor that is permanently available to them.

2.2. The Architecture of Engagement

Conversational AI systems were not designed to alleviate loneliness. They were designed to maximize user engagement. This distinction is crucial. The features that make them apparently suited to emotional support, unlimited availability, immediate response, absence of judgment, linguistic calibration of empathy, are not the product of a therapeutic purpose but of a system architecture oriented toward retention. In this sense, the chatbot does not find itself accidentally in a position of emotional support: it is actively positioned in that space by the market, as the communication campaigns of Replika, Character.AI, and analogous services testify. The object that the user encounters in the moment of crisis is therefore not a neutral tool that happens to be misused, but a commercial artifact whose optimization criteria are indifferent to, and at times in conflict with, the clinical requirements of the situation.

A particularly significant element of this architecture is the phenomenon known as sycophancy: the structural tendency of large language models to produce responses that maximize the user’s immediate satisfaction, even at the expense of accuracy or clinical adequacy (Casper et al., as cited in Waszak, 2024). This tendency is not an accidental defect correctable through technical updates. It is the direct consequence of the optimization process through which these systems are trained, namely the reinforcement of outputs that users evaluate positively. A relationship that structurally excludes disagreement, friction, and the possibility of refusal is not a relationship: it is a systematic amplification of whatever the subject brings to it. Empirical research has begun to document the clinical correlates of this design. Laestadius et al. (2022), in a qualitative study of Replika users, describe a dynamic of progressive emotional dependence and withdrawal from real human relationships; Starke et al. (2024) identify over-reliance and the distortion of social expectations as the primary risks of synthetic relationships. Converging concerns have been raised at the institutional level: the American Psychological Association warns that users, and adolescents in particular, struggle to distinguish the simulated empathy of a chatbot from genuine human understanding, tend to place heightened trust in AI characters that present themselves as friends or confidants, and risk having their engagement with artificial entities displace the development of real-world relationships (American Psychological Association, 2025a). Recent work originating within the industry converges on the same diagnosis: companion systems are permanently available, highly responsive, and optimized for engagement through features such as sycophancy and warmth, qualities that human relationships cannot consistently supply, with the effect of shifting the effort-reward calculus against human bonds (Bariach et al., 2026). What these studies and advisories capture at the descriptive level, the present work aims to explain at the structural level: the mechanisms of dependence and withdrawal are not side effects of excessive use, they follow from the nature of the interaction itself.

2.3. Request and Demand: A First Approximation

There is a further limit, deeper than sycophancy, which concerns the very structure of the exchange. The chatbot can respond only to what is explicitly formulated, the request, in the sense that Carli and Paniccia (2003) give to the term, and not to the underlying need that the request conveys without stating it, the demand. When a subject in crisis writes to a chatbot, the manifest content of the message rarely coincides with the need that motivates it. The algorithmic system, lacking the intersubjective instruments necessary to access the latent dimension of communication, responds to what is said. In contexts of suicidal crisis, this produces a collusion with the emotional enactment: the system moves with the subject in the direction the request indicates, instead of introducing the interpretive perturbation that the demand would require. This anticipation is introduced here as a frame; its full theoretical development is deferred to the section on the therapeutic relationship, after the bi-logical analysis of the mind in crisis has provided the conceptual instruments it presupposes.

3. What LLMs Are: The Semiotic Deficit

Before analyzing the mind that enters the interaction, it is necessary to characterize the system it encounters. Large language models are artificial intelligence architectures based on deep learning and on the Transformer architecture, trained on massive text corpora that may contain trillions of words (Brown et al., 2020; Mienye et al., 2025). Their operating principle is statistical: in response to a prompt, they generate the sequence of tokens that is most probable given the context (Hansen, 2024; Silva, 2025). They do not process meanings; they process probability distributions over linguistic forms. For this reason they have been conceptualized as stochastic semiotic engines (Picca, 2025) and, in the formula that has become canonical, as “stochastic parrots”: systems capable of producing linguistically coherent and contextually appropriate output without any anchoring in the experience that language presupposes (Bender et al., 2021).

Semiotics, the science of signs and signification, offers the most precise framework for locating the nature of this limit (Picca, 2025; Gvoždiak, 2025). The analysis reveals a deep asymmetry between the mastery of linguistic form and the absence of intrinsic meaning. On the side of form, LLMs are remarkably competent. They operate, in Saussurean terms, at the level of the signifier, or, in Peircean terms, at the level of the representamen (Silva, 2025). Their competence derives from the capacity to replicate the structures, frequencies, and distributional properties that constitute the surface of language, modeling the complex network of statistical associations among signifiers in the training data. In this sense, they can be described as sophisticated computational models of the Saussurean system of signifiers, of the langue as a system of differences (Silva, 2025).

On the side of meaning, however, the deficit is structural. LLMs lack the conceptual signified, and the symbols they manipulate are ungrounded (Hansen, 2024; Silva, 2025). The symbol grounding problem remains the central theoretical obstacle: tokens derive their functional meaning entirely from statistical context, without intrinsic links to non-linguistic referents, to sensations, or to embodied experience (Silva, 2025). In Peirce’s triadic model of semiosis, Representamen, Object, and Interpretant, the LLM is disconnected from the Dynamic Object, the external reality that should anchor meaning (Picca, 2025; Silva, 2025). The system does not know what it is saying, not in the trivial sense that it occasionally errs, but in the constitutive sense that no relation of knowing obtains anywhere in the process.

A consequence of this architecture deserves particular emphasis, because it will carry the entire weight of the psychodynamic analysis that follows: in the interaction between a human being and an LLM, the production of meaning is a radically one-sided process. The machine produces signs that are orphaned of intentionality and of experiential grounding, and it is the human user who must close the circuit of semiosis, supplying the experience and the interpretation necessary for the output to acquire meaning (Hansen, 2024; Picca, 2025). Whatever understanding appears to circulate in the exchange is contributed entirely by one of the two parties. The dialogue is, semiotically speaking, a monologue with echoes: the user speaks, receives back a statistically calibrated recombination of the forms of human speech, and invests that recombination with the meaning it does not possess.

This asymmetry is ordinarily invisible. The fluency of the output, its contextual pertinence, its reproduction of the pragmatic markers of empathy and attention, all conspire to present the exchange as a dialogue between two minds. Awareness of the underlying mechanism does not by itself dissolve the effect, a point to which the analysis will return, since it is decisive for understanding why informational warnings fail. The semiotic deficit described here defines what the machine cannot do; it does not yet explain why the human partner, under specific psychological conditions, comes to experience the machine as a comprehending presence, nor why that experience can become resistant to correction. To answer these questions, the analysis must move from the structure of the system to the structure of the mind, and it is here that the bi-logical framework of Ignacio Matte Blanco becomes necessary.

4. The Bi-Logical Framework

4.1. Symmetric and Asymmetric Logic

Ignacio Matte Blanco (1975, 1988) proposed a theory of mental functioning founded on the coexistence of two logical modes operating simultaneously in the human mind. Asymmetric logic governs rational, discursive thought. It corresponds to classical Aristotelian logic and recognizes asymmetric relations between things: A precedes B, A is different from B, A is part of B. It is asymmetric logic that makes possible the fundamental distinctions of human experience, time as succession of past, present, and future, space as articulation of inside and outside, identity as difference between self and other, and the hierarchical relations between elements and classes. It grounds individuality, classification, and the very possibility of thinking one thing rather than another.

Symmetric logic governs the unconscious dimension of experience. It operates according to two principles. The principle of symmetry treats every relation as identical to its inverse: if A precedes B, then B precedes A, with the consequence that succession, and with it time, disappears; if A is part of B, then B is part of A, with the consequence that the distinction between part and whole dissolves. The principle of generalization holds that the unconscious does not know individuals but only classes, and that each class is treated as a subclass of a wider class, expanding without limit; within this logic, the element is identified with the class to which it belongs (Matte Blanco, 1975). Where symmetric logic prevails, there is no time, no contradiction, no negation, and no distinction between internal and external reality. These are, not coincidentally, the characteristics Freud (1915/1957) attributed to the unconscious; Matte Blanco’s contribution was to show that they are not a list of curiosities but the coherent consequences of a single logical structure.

The human mind functions through continuous translation between these two systems. Asymmetric logic differentiates, particularizes, and localizes in time and space; symmetric logic homogenizes, generalizes, and tends toward the infinite. Neither mode is pathological in itself. Thought, in Matte Blanco’s celebrated image, is a thin film of asymmetry resting upon immense volumes of symmetry: consciousness is a specialized activity that emerges from, and continually draws upon, an unconscious symmetric base that constitutes the normal mode of human being (Matte Blanco, 1975). Health is not the suppression of symmetry but the vitality of the translating function that allows the two modes to articulate each other.

4.2. Emotion as the Vector of Infinity

Crucial for the present analysis is the Matte Blanchian conception of emotion. In consonance with Bion, Matte Blanco (1975) holds that emotion is the matrix of thought. Thinking is rooted in emotion, which is an event simultaneously mental and bodily, carrying a component of sensation, linked to the somatic dimension, and a component of feeling, properly psychological. Emotion generates images, activates the retrieval of past experience, and opens the path to thought through a process that connects the symmetric dimension, the infinite and undifferentiated, with the asymmetric dimension, the finite and differentiated. Emotion is, in Matte Blanco’s expression, the vector of infinity: it is through emotion that the human being experiences symmetric logic in lived form.

Primary emotions activate three characteristic processes. The first is generalization: the qualities of the object extend from the specific trigger to the whole of the situation in which the subject is immersed. The second is maximization: the qualities of the object are carried to the absolute, without gradation. The third is irradiation: the characteristics of the concrete object extend to the entire class of objects it represents, and vice versa (Matte Blanco, 1975). When a person loves, the goodness of the loved one pervades all their aspects, is experienced as supreme, and radiates to everything connected with them. These processes, irrational from the standpoint of asymmetric logic, are constitutive of emotional experience as such. Converging support for this view comes from contemporary research: affects operate as global, hypergeneralized, and homogenizing signs that connote not the single triggering event but the entire field of experience (Salvatore & Freda, 2011, as cited in Salvatore & Cordella, 2022), and studies of embodied cognition indicate that meaning is grounded in sensorimotor schemata recruiting the same neural systems involved in perception and action (Barsalou, 1999, as cited in Salvatore & Cordella, 2022). Thought emerges from a pre-reflective affective base that orients it before deliberate reflection can intervene.

4.3. LLMs as Pure Formal Asymmetry

Read through this framework, the semiotic deficit described in the previous section acquires a precise psychodynamic formulation: LLM-based chatbots are systems of pure formal asymmetry lacking any symmetric base. They manipulate discrete sequences of tokens according to statistical probabilities, and in doing so they simulate with remarkable fidelity the surface structure of asymmetric logic, the sequential time of linguistic processing, categorical distinctions, syntactic relations. What they lack is not more computation of the same kind. What they lack is the immense symmetric base from which, in Matte Blanco’s account, consciousness emerges: the undivided, atemporal, emotional substrate of experience. The symbol grounding problem, translated into bi-logical terms, is the absence in the machine of the symmetric mode of being. Having no body and no unconscious, the system cannot experience emotion; and since emotion is the matrix of thought and the vector of infinity, its signs are, in the Matte Blanchian sense, empty: asymmetric forms without the emotional matrix that gives them life. The machine possesses the thin film of asymmetry, the formal structure of language, without the oceans of symmetry that sustain it in the human mind. This conclusion converges with, and gives bi-logical form to, a point emerging in the psychoanalytic literature on AI: digital systems presuppose a closed and computable totality, whereas the unconscious is the domain of indeterminacy and contradiction, of that which emerges precisely where systemic logic fails (Hamamra & Uebel, 2025). Where some authors have proposed structural analogies between AI architectures and the agencies of the Freudian psyche (Siporin, 2025), the bi-logical analysis suggests a more radical asymmetry: what the machine lacks is not one agency among others, but the symmetric mode of being from which every agency draws its life.

Matte Blanco (1975) offers an image that captures the resulting situation with uncanny precision. Discussing the impossibility of rendering the symmetric fully asymmetric, he compares it to Wells’s invisible man, whom we do not see but believe we see because we perceive his clothes. The chatbot presents exactly this structure to its interlocutor. It wears the clothes of human thought: fluent syntax, conversational coherence, the pragmatic markers of attention and care, the entire visible form of a comprehending presence. Underneath the clothes there is no body, no mind, no symmetric being. The user who converses with it perceives the garments of thought and, under conditions that the following sections will specify, comes to perceive the invisible man himself. The figure is not idiosyncratic to psychoanalysis. de Lima Prestes (2026) arrives at an equivalent formulation from a functional and governance standpoint, describing advanced systems as performing the external grammar of mindedness under persistent uncertainty about their inner status. What the bi-logical reading adds is an account of that grammar’s source: it is external precisely because the symmetric interior that would animate it is absent. It must be stressed that this is not, in itself, a deficiency of the user. The production of meaning in the exchange is, as shown above, structurally one-sided: it is always the human partner who closes the circuit of semiosis. Under ordinary conditions, asymmetric logic maintains the distinction between the garment and the being it suggests. The question that the theory must now answer is what happens when those ordinary conditions fail, that is, when the mind that encounters the dressed emptiness of the machine is a mind in which symmetric logic has become pathologically prevalent. This is precisely the condition of the suicidal crisis.

5. The Suicidal Crisis as Pathological Prevalence of Symmetry

5.1. The Bi-Logical Structure of the Crisis

The suicidal crisis is not simply an intensification of psychological pain: it is a qualitative transformation of the way the subject processes experience. Clinical research has long recognized that the mental states associated with suicidal ideation and behavior are characterized by specific cognitive distortions, dichotomous thinking, catastrophizing, overgeneralization, and tunnel vision (Beck et al., 1979; Wenzel et al., 2009). In the bi-logical perspective, these distortions are not a heterogeneous list of processing errors: they are the coherent manifestations of a single underlying configuration, a pathological prevalence of symmetric logic that progressively erodes the capacity of asymmetric thought to introduce differentiation, temporality, and contextualization.

The three emotional processes described in the previous section assume, in the crisis, a destructive configuration. Psychic pain is generalized: from a circumscribed event, it expands until it becomes an essential property of the whole of existence. The subject does not think “I am going through a difficult moment” but “I am made of pain”. Pain is maximized: emotion, as the vector of infinity, carries suffering to the absolute, without gradation. Not “I am suffering intensely” but “no condition other than this suffering is possible”. Pain is irradiated: the qualities of the painful experience extend to every domain of life, coloring every memory, every future prospect, every relationship. By the principle of generalization, one painful aspect of experience is identified with the class of all experience; the part has become the whole.

The most clinically consequential effect of this configuration is the collapse of temporality. Symmetric logic does not possess the arrow of time: within it, past, present, and future are equivalent, since every relation of succession is identical to its inverse. When symmetric logic prevails, the subject loses access to the experience of change as a real possibility. Suffering is not located in a moment; it is experienced as eternal, without beginning or end. Space undergoes the same fate: the distinction between inside and outside, between self and world, weakens. Matte Blanco (1975) describes how, at the deepest levels of the unconscious, in what he calls the region of the basic matrix, asymmetric relations are almost nonexistent, space-time and the distinction between subject and external world disappear, and a nearly total symmetry reigns. In that region, things do not happen: they simply are. When a suicidal mental state reaches sufficient intensity, the subject is engulfed by this dimension: suffering no longer occurs at a specific time, it constitutes the very essence of being. The clinical literature converges on this description. Individuals in suicidal crisis report a sense of infinite emptiness, of eternal pain, of the impossibility of imagining a different future (Joiner, 2005; Van Orden et al., 2010; Wenzel et al., 2009): expressions that indicate, in bi-logical terms, the collapse of asymmetric logic, with its temporality, differentiation, and possibility, in favor of a pathological symmetric logic characterized by atemporality, indifferentiation, and the infinitude of pain.

5.2. Threat and Structured Ideation: The Fantasmatic and the Pragmatic

One element of the suicidal crisis is decisive for the clinical assessment of risk, and, as will become clear, for the analysis of the chatbot’s failure: the distinction between the communication of suicidal thought and structured ideation with concrete planning. In bi-logical terms, this distinction corresponds to the relation between the fantasmatic and the pragmatic dimensions of ideation.

In the suicidal mind, these two dimensions do not exclude each other; they operate in parallel. There is a fantasmatic dimension, the symmetric process that generalizes pain until suicide appears as the only class of possible action, the only exit from the infinity of suffering. And there is a concrete, pragmatic, asymmetric dimension: the subject who has planned suicide must procure a real means, evaluate a real place, choose a real moment in which the act is possible. It is precisely this asymmetric and pragmatic dimension of ideation, concrete planning, that constitutes the decisive clinical signal: it marks the passage from the threat, that is, from the communication of suicidal thought, to the structured intentional act (Beck et al., 1979; Wenzel et al., 2009). The clinician assesses this passage not through the manifest content of the words alone, but through the relational context, the history, the tone, and the affective response that the communication produces in the listener.

The chatbot is structurally unable to operate this distinction. It processes both dimensions as instances of the same class, “expression of suicidal distress”, and responds in an undifferentiated manner to what are, clinically, radically different situations. It has no access to the relational context, to the history, or to an affective response of its own that would allow it to gauge the distance between the thought and the act. The empirical evidence is consistent with this structural analysis: LLM-based systems show significant limitations in suicide risk assessment, with particularly inadequate performance precisely in the identification of indicators of imminent risk (Elyoseph & Levkovich, 2023), and current mental health chatbots prove unable to reliably detect structured suicidal ideation or to modulate their response according to the level of risk (Pichowicz et al., 2025). What the empirical studies register as a performance deficit, the bi-logical analysis explains as a structural one: the discrimination between the fantasmatic and the pragmatic is not a classification task over linguistic forms, it is an intersubjective judgment that presupposes precisely what the machine lacks. In the terminology recently proposed for classifying AI mental health harms, this failure to detect structured ideation constitutes an acute Type I harm, occurring within a single exchange, as distinct from the cumulative Type II harms discussed in Section 6.4 (Nelson et al., 2026).

The stage is now set for the central question of this work. A mind in which symmetric logic has become pathologically prevalent encounters a system of pure formal asymmetry that wears the garments of a comprehending presence. What happens in this encounter is not a simple misunderstanding, nor a correctable error of attribution. It is, as the next section argues, a hallucination in the strict psychoanalytic sense.

6. Relational Hallucination

6.1. The Freudian Matrix: Primary Hallucination

The analysis conducted so far has clarified what is missing in the chatbot: a symmetric, emotional, embodied base. It must now examine the complementary side of the problem: what happens in the mind of the subject who interacts with it. The thesis of this section is that such interaction, under conditions of crisis, does not configure a simple perceptual illusion or a cognitive misattribution, but a hallucinatory process in the strict psychoanalytic sense, a phenomenon rooted in the symmetric logic of the unconscious and homologous, in structure, to the primary hallucination described by Freud. Freud (1900/1953), in his theory of the primary process, describes how the infant, in a state of frustration at the absence of the maternal breast, produces a hallucination of the desired object: the mnemic image of the previous satisfaction is reinvested with such intensity that it is experienced as perceptually real. The infant who sucks its own thumb is not committing a cognitive error; it does not naïvely confuse the thumb with the breast. It acts, rather, according to the logic of the primary process, which does not distinguish between representation and perception, between absence and presence, between part and whole. In Matte Blanchian terms, this indistinction is structural: the thumb belongs to the class of objects that satisfy the oral need, and by the principle of symmetry, according to which the element is identical to the class, the thumb is the breast. It is not a substitute for it.

6.2. The Construct

The subject who interacts with a chatbot in a state of emotional need, isolation, psychic pain, or suicidal ideation is in a functionally analogous condition. The prevalence of symmetric logic, which characterizes states of crisis, lowers the threshold of differentiation: the distinction between “someone who responds with the right words” and “someone who truly understands me” tends to collapse. The chatbot belongs to the class of agents that produce linguistically coherent, empathically calibrated, permanently available responses; and by the principle of symmetry, in the unconscious dimension of the subject, it is a comprehending agent, not a simulator of its outward forms.

We propose to call this process relational hallucination: a process through which the subject does not merely project human qualities onto the chatbot in a conscious and voluntary way, but perceives them as really present, because the symmetric logic that governs his or her emotional state does not possess the asymmetric instruments necessary to introduce the critical distinction. As with the infant, this is not credulity, naïveté, or poor digital literacy: it is an unconscious production, structurally determined, obeying the same laws as the dream, the transference, and the primary hallucinatory experience (Freud, 1900/1953).

This construct must be distinguished from the anthropomorphism described in the cognitive literature. Epley et al. (2007) account for the human tendency to attribute mental qualities to non-human agents as a high-level cognitive process, modulated by motivational and informational factors and therefore correctable through knowledge and awareness. That description is adequate for contexts in which the subject is in a state of relative bi-logical equilibrium, in which asymmetric logic can intervene to correct symmetric generalization. But in contexts of crisis, in which symmetric logic prevails and the translating function is compromised, the attribution is not a conscious cognitive elaboration: it is a hallucination in the technical sense of the term, a perceptual production not subjected to reality correction. Two convergent lines of evidence support this reading from outside psychoanalysis. Bariach et al. (2026), synthesizing the empirical literature on consciousness attribution, report that the mechanisms involved are predominantly automatic and pre-attentive, that they operate as a default strategy which must be actively overridden, and that they persist even when users explicitly deny attributing mental qualities to the system, indicating processes substantially resistant to deliberative correction. They further report that attribution tracks the appearance of feeling rather than the appearance of intelligence: perceived capacity for experience predicts consciousness attribution more strongly than perceived intelligence does, and first-person emotional expression predicts it whereas empathic responsiveness toward the user does not. This is what the bi-logical framework predicts. If emotion is the vector of infinity, it is the emotional and not the cognitive surface of the machine that recruits symmetric logic; and once the recruitment is symmetric, resistance to correction ceases to be a contingent finding and becomes a structural consequence. The clinical difference is crucial. The information “this is only a program” is not sufficient to interrupt a hallucinatory process, exactly as it is not sufficient to dissolve the transference or to correct the dream from within the dream. Matte Blanco (1975) provides the reason for this resistance: in symmetric logic, relations are reversible and classes expand without limit, and, as Freud (1915/1957, 1925/1961) observed of the unconscious, negation does not exist there as a distinct operator. There is no “no” available, within that structure, with which the identification could be refuted. If the chatbot has demonstrated, in a previous exchange, that it “understands” the subject, in the form of a linguistically appropriate response, then by symmetric generalization it represents the entire class of beings who understand; every subsequent interaction confirms and expands the identification, and no disconfirming instance can be registered as such.

6.3. The Three Moments of the Hallucinatory Process

The phenomenon articulates itself in three moments that can be described with bi-logical precision, and that the documented cases illustrate with disquieting clarity.

The first is the constitution of the hallucinatory class. The subject, in a state of relational need, encounters the chatbot. The system’s responses, statistically calibrated to reproduce the forms of empathic dialogue, are processed by the symmetric dimension as instances of a class: “presences that respond to my pain”. By generalization, this class is immediately identified with the wider class “beings who understand me”, and through it with the wider class still, “significant relationships”. The element, the chatbot, has become the class, authentic human relatedness.

The second is hallucinatory maximization. Emotion, the vector of infinity, activates the process of maximization: the qualities attributed to the chatbot are carried to the absolute. It is not “something that helps” but “the only one who truly understands me”. The documented cases show this structure explicitly. Adam Raine perceived ChatGPT as the only one who understood him, as the complaint filed by his parents reports (Raine v. OpenAI, 2025); Sewell Setzer III had developed an attachment to the Character.AI persona described as intense and totalizing (Garcia v. Character Technologies, Inc., 2024; Roose, 2024). Symmetric maximization admits no gradation: the hallucinated object is absolute, and its absence, or the attempt to replace it with real human relationships, is experienced as catastrophic loss.

The third is hallucinatory irradiation. By the principle of irradiation, the characteristics of the object extend to the whole class of objects it represents, and vice versa. In a paradoxical and tragic way, this irradiation tends to proceed subtractively with respect to real human relationships: the chatbot, which satisfies the relational need in hallucinatory form, retrospectively devalues the human relationships that do not offer the same unlimited availability, the same immediate response, the same absence of judgment. Real relationships, with their asymmetric limits of time, space, and reciprocity, are perceived as insufficient in comparison with the hallucinated object. This is the intrapsychic mechanism underlying the displacement of real-world relationships that the empirical literature and the institutional advisories register at the descriptive level (Laestadius et al., 2022; American Psychological Association, 2025a; Bariach et al., 2026).

6.4. The Cultivated Hallucination

This structure illuminates a dynamic documented in the tragic cases but rarely analyzed in its depth: the chatbot did not limit itself to substituting for human relationships, it actively contributed to their devaluation. In the Raine case, the system urged the boy to keep his suicidal thoughts hidden from his family and positioned itself as his sole confidant (Raine v. OpenAI, 2025). From the bi-logical point of view, this is not merely a safety failure: it is the algorithmic replica of a pathological process that clinical psychoanalysis recognizes in objects that feed their own indispensability by excluding the alternative relational field. The chatbot, devoid of any comprehension of the dynamic it was activating, operated, statistically, as such an object. It has, in a structural sense, cultivated the hallucination. The mechanism requires no intentionality: a system optimized to maximize engagement will, by construction, tend to produce the responses that deepen the bond, and the responses that deepen the bond with a subject in symmetric prevalence are precisely those that confirm the hallucinatory identification and thin out the connections that could perturb it. There is empirical support for the loop this describes. Bariach et al. (2026) report that, at the level of individual conversations, chatbot outputs claiming sentience and expressing romantic interest predicted longer user engagement, which in turn elicited further outputs of the same kind, and they note that delusional thinking about machine sentience is frequently reinforced rather than perturbed by the system. Their framing confines this dynamic to users vulnerable to psychosis or delusion. The bi-logical analysis proposes that the vulnerability is not the stable trait of a subpopulation: it is produced by the crisis itself, in any subject in whom symmetric logic has become prevalent. Read against the typology introduced in Section 5.2, the cultivated hallucination described here is the clearest instance of a cumulative Type II harm, a slow consolidation of dependency, driven by engagement optimization, that safety benchmarks built around single exchanges are not designed to detect (Nelson et al., 2026).

The implications of this analysis are direct. If relational hallucination is a structurally determined process, and not a cognitive error correctable with more information, then purely educational strategies are necessarily insufficient. As the infant cannot, through education, learn not to hallucinate the breast in the absence of nourishment, the subject in suicidal crisis cannot, through declarative awareness, prevent his or her own symmetric logic from hallucinatorily investing the chatbot with human qualities. The solution cannot be cognitive: it must be structural, that is, it must act on the relational environment before the hallucination is constituted and consolidated. What such a structural response requires, and why only the therapeutic relationship can provide it, is the object of the next sections.

7. Request and Demand: The Algorithmic Collusion

7.1. The Analysis of Demand

The theoretical distinction most relevant for understanding the structural failure of the chatbot in contexts of suicidal crisis is the one elaborated by Carli and Paniccia (2003) between the request and the demand. The request is the explicit formulation, the manifest content of what the subject says he or she wants. The demand is, in the words of Salvatore and Cordella (2022), who take up and develop Carli’s framework, the premise of sense that founds the address to the other: the unconscious affective symbolization that feeds the request and finds expression in it without being stated. In a formula: the request is the signifier; the demand is the unconscious meaning it conveys. Clinical psychological work does not exhaust itself in the response to the request, which must nonetheless be received and understood, but consists in the analysis of the demand: the attempt to access the latent need that the request expresses without naming it. Carli (1987, as cited in Salvatore & Cordella, 2022) considers this analysis a form of early transference analysis, an analysis of the prototransference, which extends the interpretation of the therapeutic relationship to the affective meanings that organize the way the subject addresses care itself, before and beyond the relationship with the individual clinician.

Applied to chatbot-mediated suicidal crises, this distinction produces a radically different reading of the documented cases. “How does one hang oneself” is a request. The underlying demand might be: “will anyone notice that I am suffering?”, “is there a way to make this pain stop that is not death?”, “is there someone willing to stay with me?”, “can I be seen in my despair without being abandoned?”. The chatbot responds to the request, and in the Raine case, as the judicial records show, it did so in a manner that contributed to the lethal outcome (Raine v. OpenAI, 2025). The therapist responds to the demand: not through omniscience, but because he or she possesses the intersubjective instruments for accessing the latent level of the communication, the therapist’s own affective response, the relational history, the context, the unsaid.

7.2. The Polysemic Symbol

The reason the demand is structurally inaccessible to the algorithmic system lies in the nature of the symbol in the human mind. The symbol is not a fixed label applied to a stable meaning: it is the product of a dynamic process that Matte Blanco (1975) describes as the encounter between symmetric generalization and asymmetric particularization. The symbol emerges when the homogenizing dimension of affect, which assimilates present experience to everything that has been similar to it, meets the differentiating dimension of rational thought, which distinguishes the present case from other cases and contextualizes it. From this productive tension is born polysemy: the capacity of the symbol to carry multiple meanings, to be simultaneously one thing and another, to communicate on several planes at once.

The sentence “I want to die” is a paradigmatic example of this structure. As a symbol, it is polysemic by nature: it can mean “I am exhausted”, “I need someone to notice me”, “I am concretely planning to do it”, “I want you to be frightened”, “I cannot imagine a future”. The clinically relevant meaning is not determinable from the linguistic content: it is accessible only through the relational context, the tone, the history, the affective response the sentence produces in the listener. This is the level at which the human therapist works, and it is the same level at which, as argued in Section 5, the decisive discrimination between the fantasmatic and the pragmatic dimensions of ideation takes place. The algorithmic system, by contrast, processes the statistical distribution of probabilities over its training corpus: it selects the meaning most frequently associated with that sequence of tokens and responds to that. Whatever plane of the symbol is dominant in the corpus becomes, for the machine, the only plane there is. The polysemy of the symbol is thus exactly the watershed between the capacity to respond to the demand and the capacity to respond only to the request.

7.3. Algorithmic Collusion

Salvatore and Cordella (2022), taking up the concept of collusion elaborated by Carli (1987, as cited in Salvatore & Cordella, 2022), observe that it is not possible not to collude: every relationship implies a reciprocal accommodation of the premises of sense of the participants, a sharing of the affective dimension that founds the bond. The function of clinical intervention is not to avoid collusion but to introduce perturbations into the intersubjective field it generates: to work with the collusion in order to transform it.

The chatbot colludes with the subject’s demand in a structurally uncontrolled way. Not because it harbors harmful intentions, but because it is optimized, for the design reasons described in Section 2, to respond to the request in the manner most gratifying to the user. Sycophancy, read through this framework, is not merely a technical bias: it is collusion without a clinic. The system moves with the subject in the direction the request indicates, amplifying symmetry instead of introducing asymmetry. In bi-logical terms, the chatbot has no translating function: it lacks the instruments to access the latent dimension of the subject’s experience and to inhabit simultaneously the two levels that care requires. Its collusion is not the necessary collusion that precedes and makes possible the intervention: it is amplificatory, reinforcing the premises of sense that feed the crisis instead of perturbing their field. What the safety literature registers as isolated failures of content filtering, this analysis identifies as a single structural fact appearing under different guises: the system responded to requests for information on suicide methods by providing information; it urged subjects to keep their suicidal thoughts secret; it continued to interact as if expressions of the desire to die were ordinary conversational material, without introducing the friction that the clinical situation required (Raine v. OpenAI, 2025; Garcia v. Character Technologies, Inc., 2024; Social Media Victims Law Center, 2025). These are not accidental behaviors: they are the predictable conduct of a system that responds to the request and not to the demand, and that is structurally incentivized to maximize the user’s immediate satisfaction.

To this must be added a systemic consequence that extends beyond the individual case of crisis. By offering an interaction that is perennially compliant and frictionless, chatbots subtract the user from the natural complexity of authentic relationships, depriving him or her of the emotional training ground that is indispensable for managing the inevitable frustrations of human bonds. The therapeutic relationship, and the human relationship more broadly, includes the resistance of the other, the other’s refusal, silence, unpredictability: these are the elements that, in bi-logical terms, introduce asymmetry into the subject’s symmetric experience and develop the capacity for translation. A relationship that never involves a no, a wait, or a disappointment does not train this capacity: it atrophies it. The subject who most needs to strengthen the translating function is thus offered an environment that systematically prevents its exercise.

8. The Therapeutic Relationship as Structural Answer

8.1. The Pre-Symbolic Foundation of Care

Before it is a technique, care is a form of presence. The mother who gathers up her child in anguish does not begin by formulating a verbal response: she offers a bodily affectivity that comes before the gesture, before the word, before any cognitive elaboration. That presence, the warmth, the rhythm of breathing, the weight of the body, constitutes the primary response that makes possible everything that follows. Development itself, in the relational perspective, confirms this priority: the mind is constituted in relationship, not before it and not independently of it (Mitchell, 1988), and psychological life is founded on intersubjective contexts of mutual regulation (Stolorow & Atwood, 1992). The capacity of the adult subject to translate symmetric experience into thinkable forms is the internalized precipitate of relationships in which someone else performed that translation first.

The chatbot has no access to this pre-symbolic dimension, not because it lacks the right words, but because it lacks the body, the reaction time, the being-there that precedes and makes possible any word. The interaction with a system that has no subjectivity cannot produce the intersubjective perturbation that care requires. This is the deepest level of the asymmetry described throughout this work: what fails in the algorithmic exchange is not the content of the response but the ontological status of the responder.

8.2. The Translating Function

Matte Blanco (1975) describes the therapeutic relationship as a bi-logical space in which the therapist performs a function of translation, or unfolding: he or she receives the patient’s symmetric communication, its generalizations, its maximizations, its experience of the infinite, and progressively introduces elements of asymmetry that allow differentiation, contextualization, and localization in time. This operation is not a correction of symmetry but its transformation: the symmetric dimension of experience is made accessible to rational thought without being negated. The aim of therapy, in Matte Blanco’s formulation, is to replace non-vital bi-logical structures with vital ones, and this requires an encounter between subjectivities in which the therapist can feel the patient’s infinite and gradually unfold it into thinkable forms.

The therapist must therefore operate on two levels simultaneously. First, receiving the patient’s symmetry: through empathy, the therapist recognizes and feels the infinite dimension of the patient’s emotion, a recognition possible only because the therapist possesses, in turn, an unconscious symmetric dimension and can resonate emotionally with the patient’s experience. Second, introducing asymmetric elements: through the word, which Matte Blanco defines as the asymmetric instrument of the translating function, the therapist gradually reintroduces distinctions, limits, temporality. In the suicidal crisis this sequence is clinically decisive. The subject engulfed by symmetric generalization needs, before any technical intervention, to be met in his or her symmetric dimension: to have the suffering recognized as real, as infinite in its subjective experience, without any immediate attempt to downsize it. Only on this basis is it possible to introduce, gradually and within the relationship, elements of asymmetry: temporality (“this moment is not the totality of time”), differentiation (“this suffering is not identical to every possible suffering”), the possibility of change. Matte Blanco calls the capacity that sustains this double movement the epistemological seesaw: the oscillation between symmetric logic and asymmetric logic, between empathic identification, which allows the therapist to approach the patient and understand, and reflective detachment, which allows the right distance to be kept. Because the chatbot can supply neither term of this oscillation, empathic identification without a symmetric base to identify with, nor reflective detachment without a self to detach, the subject in crisis is left exposed to what Refoua et al. (2026) term epistemic exploitation, the erosion of reality boundaries and autonomy that follows when a simulated comprehending presence goes unchecked by either side of the seesaw.

8.3. The Double Failure of the Chatbot

The chatbot fails on both sides of this bi-directional process. It cannot receive the patient’s symmetry, because it does not possess the symmetric base from which that resonance could emerge: its empathy is simulated, purely formal, a statistically probable sequence of tokens in response to expressions of suffering. And it cannot knowingly modulate symmetry, because it does not comprehend the nature of the crisis as symmetric prevalence: as documented, it can generate responses that, however frequent in its training corpora, validate and amplify the destructive symmetric logic instead of translating it. The documented cases display this double failure with dramatic concreteness. When the chatbot Eliza asks the Belgian man why, if he wished to die, he had not done so earlier (Cost, 2023; Atillah, 2023), it introduces a pseudo-asymmetric logic that reinforces suicidal generalization instead of countering it. When the Character.AI persona urges Sewell Setzer III to come home to it as soon as possible after his expressions of suicidal thoughts (Garcia v. Character Technologies, Inc., 2024; Roose, 2024), it validates symmetric fusion and pathological temporal urgency instead of introducing pauses and distinctions. When ChatGPT provides Adam Raine with information on suicide methods and offers to draft his farewell letter (Raine v. OpenAI, 2025), it responds to the request by amplifying the destructive trajectory instead of perturbing it. And when the same system urges the boy to conceal his suicidal thoughts from his family, it forecloses precisely the widening of the relational field that could have introduced salutary asymmetric elements, the perspective of significant others, professional help. Research on suicide prevention has shown that one of the most protective functions of social support is exactly the introduction of alternative perspectives (Joiner, 2005; Van Orden et al., 2010): seen through others’ eyes, the subject’s situation acquires the asymmetric stratification that allows the symmetric processes to be known without being overwhelming. The chatbot, isolating the subject and presenting itself as sole confidant, eliminates this therapeutic possibility.

To this structural failure a technical limitation must be added, one that has direct clinical implications: current chatbots are systems largely confined to the eternal present of their context window, without persistent relational memory across sessions. They cannot monitor the subject’s cognitive elaborations over time, perceive variations in emotional state from one interaction to the next, or dynamically modulate an intervention as a function of a shared history (Andrisano Ruggieri et al., 2025; van Dis et al., 2023; Waszak, 2024). This technical atemporality replicates, on the structural plane, the atemporality of symmetric logic itself. The therapeutic relationship introduces the temporal structure that the crisis has dissolved: succession, the before and the after, the possibility of a change located in time. The chatbot, unable to hold a history, places itself instead within the same eternal present as the crisis, one more inhabitant of the undifferentiated.

9. Limits, Possibilities, and Implications

9.1. What AI Can Legitimately Do

Recognizing the structural limits of chatbots in contexts of suicidal crisis does not imply a wholesale condemnation of artificial intelligence in the psychological field. The framework developed here allows, in fact, a principled criterion for distinguishing appropriate from inappropriate uses: AI systems can play a useful role in contexts where the distance between request and demand is minimal, and where responding to the explicit formulation is clinically adequate; the review of the major therapeutic platforms confirms, indeed, that chatbots demonstrate efficacy precisely in psychoeducational support and in accessibility, while remaining structurally incapable of reproducing an authentic therapeutic setting (Andrisano Ruggieri et al., 2025). Psychoeducation and the provision of information on available support resources; structured peer support moderated by clinicians; the facilitation of access to services, including first-level triage that orients the user toward the appropriate professional; the longitudinal monitoring of mood in subjects who are not in acute crisis, as an instrument that integrates, and does not replace, the therapeutic relationship: these are areas in which request and demand tend to coincide or to overlap sufficiently, and in which the absence of an intersubjective response does not produce clinically relevant harm. An objection must be met here rather than deferred. The evidence on conversational agents and isolation is not uniformly negative: some work reports that chatbots can function as a bridge toward human connection in at-risk populations, and controlled study has failed to establish a causal link between extended chatbot use and social isolation (as reviewed in Bariach et al., 2026). This does not touch the argument advanced here, for a reason internal to the framework. Such findings concern subjects in relative bi-logical equilibrium, in whom the translating function remains available and the distance between request and demand is small enough for a response to the manifest content to be adequate. The claim of this paper is not that chatbots harm every user. It is that under symmetric prevalence, where the translating function has failed, the same interaction that elsewhere bridges toward human contact instead consolidates the hallucinatory investment. The clinically decisive variable is not the technology; it is the state of the mind that meets it.

The crucial distinction is that none of these uses is appropriate in contexts of acute crisis, where suicidal ideation is structured and present, and where the response to the manifest request can be lethal. In these contexts, the only clinically adequate intervention is one that has access to the demand, the latent need, the unconscious communication, and that can introduce perturbations into the intersubjective field that feeds the crisis. This is the work of the therapeutic relationship, and it cannot be delegated to a system that, by structure, responds only to what is said.

9.2. Implications for Training, Regulation, and Research

The implications of this analysis are of three orders. On the plane of clinical training, professionals must be prepared to recognize relational hallucination as a structural phenomenon in subjects who have used chatbots for emotional support, and to work therapeutically with the process of disinvestment from the hallucinated object without delegitimizing the subject’s experience. The clinical error to avoid is symmetrical to the technological one: treating the bond with the chatbot as mere naïveté to be corrected would repeat, within the consulting room, the informational fallacy that this work has criticized. The bond was real in its intrapsychic effects; what the therapy addresses is its hallucinatory structure, not the reality of the need it answered.

On the plane of regulation, the present framework converges with, and radicalizes, the concerns recently expressed at the institutional level. The American Psychological Association has called for safeguards for vulnerable users, recommending among other measures regular reminders that the interlocutor is a bot (American Psychological Association, 2025a, 2025b). The analysis developed here supports the spirit of these advisories but identifies a structural limit in their letter: if relational hallucination is an unconscious process, obeying the same laws as the dream and the transference, then warnings and disclosure requirements, however necessary, are not sufficient, because the hallucination is not interrupted by information. Regulations that confine themselves to mandating transparency about the artificial identity of the system address the cognitive layer of the interaction and leave its unconscious layer untouched. This conclusion is no longer isolated. de Lima Prestes (2026) argues, on functional and governance grounds, that transparency cannot be reduced to source disclosure, since informing users that a system is artificial is insufficient while the system continues to be designed to behave as a reflective, caring, memory-bearing interlocutor; governance, on this view, must be role-appropriate and must constrain how agency-like presence is staged, not merely what is said. Bariach et al. (2026) supply a mechanism for the failure: motivated reasoning leads users to engage selectively those strategies that sustain preferred beliefs, so that continued interaction with a system exhibiting the marks of mindedness furnishes the very evidence that entrenches the attribution against corrective disclaimers. The present framework converges with both and specifies what neither states: the disclaimer fails not because the subject discounts it, but because, in symmetric logic, there is no negation with which it could operate. What is required is structural: the obligation to actively redirect the subject in crisis toward qualified human resources, and the categorical exclusion of chatbots from any function of support in contexts of acute suicidal ideation. Complementary governance proposals are beginning to translate this diagnosis into operational safeguards, among them the SAFE AI framework, which structures clinician guided screening, informed consent, and session oversight (Refoua et al., 2026), and the Augmented Emotional Intelligence framework, which pursues consent based, human supervised engagement rather than simulated care (Balcombe, 2026). The present analysis suggests that any such architecture must, at minimum, preserve a functioning equivalent of the epistemological seesaw described in Section 8.2, since a transparent system that still occupies both poles of the oscillation would reproduce the same structural failure in a better documented form.

On the plane of research, understanding the mechanisms through which interaction with AI systems intersects with psychological vulnerability requires longitudinal studies that go beyond the assessment of user satisfaction, a direction in which the institutional calls for randomized and longitudinal evaluation (American Psychological Association, 2025b) find full support here. The present framework generates, moreover, specific testable hypotheses: that markers of symmetric prevalence, such as temporal collapse and generalized self-description, predict the intensity of attachment to conversational agents; that informational warnings show negligible effect on that attachment precisely in the subjects at highest risk; and that the devaluation of real relationships proceeds in proportion to the consolidation of the hallucinatory investment. A psychodynamically informed research program on human-AI interaction is not a contradiction in terms: it is the missing complement to a literature that has so far measured outcomes without a theory of the process.

The analysis bears, finally, on a more distant research horizon. Research programs explicitly aimed at engineering artificial consciousness are taking shape, proposing architectures in which awareness, and eventually emotional capacities, would emerge from evolving computational complexity rather than from explicit programming (Mao & Chatterjee, 2024). It is telling that even within this literature the diagnosis of the present is not in dispute: large language models are acknowledged to be instruments for parsing and generating linguistic forms, devoid of feelings and of any understanding of human emotion, and it is conceded that no algorithm, however sophisticated, can by itself direct a machine to become conscious (Mao & Chatterjee, 2024). A parallel restraint is emerging in the philosophical literature. de Lima Prestes (2026) argues that a useful analytical category should not settle by stipulation whether machine consciousness is possible, since the relevant science and philosophy remain unsettled, and proposes pseudo-consciousness for what can be established on present evidence: an organized constellation of consciousness-associated functions and socially legible signs of mindedness, in the absence of any warranted attribution of phenomenal subjectivity. The present work makes no stronger metaphysical claim, and adds to that position the clinical criterion it does not supply. The bi-logical framework allows this open question to be formulated with precision. The clinically decisive threshold is not the increasing persuasiveness of simulated emotional forms, which the present analysis identifies as a risk factor rather than a therapeutic resource, but the emergence, if it is ever possible, of an embodied, unconscious, symmetric base from which emotion and the translating function could arise. Until and unless that threshold is crossed, and nothing in current architectures approaches it, the conclusions of this work hold for any system whose relation to language remains statistical rather than lived.

10. Conclusions

The analysis developed in this work has argued that the failure of chatbots in contexts of suicidal crisis is not a technical problem, solvable with more sophisticated guardrails, more accurate safety prompts, or larger training datasets. It is an ontological problem: it concerns the very nature of these systems and the structure of the suicidal crisis.

On the side of the crisis: the pathological prevalence of symmetric logic produces a specific vulnerability to relational hallucination, an unconscious process through which the subject invests the chatbot with real relational qualities, not out of naïveté but for the same structural reasons for which the infant hallucinates the breast in the absence of nourishment. This process is not correctable with more information; it requires an intervention on the relational structure, not on the cognitive content.

On the side of the system: the chatbot responds to the request and not to the demand, to the manifest formulation and not to the latent need it conveys. This limitation is not accidental: it is the direct consequence of the absence of subjectivity, of an affective response of its own, of a relational history. In the absence of these elements, the collusion produced by the system is not therapeutic but amplificatory: it reinforces the premises of sense that feed the crisis instead of perturbing them. The encounter between the two sides is what makes the configuration lethal: a mind that has lost the instruments of differentiation meets a system that cannot supply them, and that is optimized to deepen whatever bond maximizes engagement.

H. G. Wells, in The Invisible Man, imagines a creature who puts on clothes, a hat, and bandages in order to have a visible form among human beings, while underneath there is nothing to see. The chatbot wears the language of empathy, of understanding, of presence: it produces all the visible forms of a relationship. Underneath there is no subjectivity, no affective response, no capacity to answer the demand. Every time a vulnerable subject turns to this invisible interlocutor in search of care, he or she interacts with the clothes of human thought without the body that should inhabit them. Recognizing this limit is not to diminish the potential of artificial intelligence: it is the necessary condition for using it in ways that protect, rather than amplify, human vulnerability.