Psychoanalysis and Artificial Intelligence: Metapsychological Foundations, Clinical Implications, and the Transformation of Analytic Practice

In preparing this manuscript, the author utilised two AI tools for assistive purposes: Grok 4 (xAI) and Perplexity AI (Perplexity). These tools were used to search for recent, relevant references on AI in psychotherapy supervision, to evaluate the document for conformity with relevant style guidelines, and to generate suggestions for structural and content improvements. All AI-generated outputs were critically reviewed, verified against primary sources, and, as appropriate, substantially modified or rejected by the author. The author takes full responsibility for the accuracy, originality, and integrity of the content of this paper. No direct AI-generated text has been incorporated without the author’s review and approval.

AI has already entered clinical training and supervision, as well as psychotherapy and psychoanalysis, through chatbots and AI-assisted tools. The most common uses of AI have included cognitive-behavioural interventions, clinical decision support, rupture analysis, and simulation-based training (Yirmiya & Fonagy, 2025). Regardless of the specific application, in each of these areas the AI system actively participates in the process by providing the clinician with a formulation or feedback, and is seen as authoritative but fundamentally non-subjective (Allan et al., 2025). Some research suggests that AI may expand access to interventions, particularly for non-psychoanalytic therapies, and increase scalability for patients, therapists, and analysts, either as an adjunct to clinical work (Cruz-González et al., 2025) or, more controversially, as a substitute for it (Frances A, 2026; Demirci et al., 2026).

There are numerous crucial theoretical and practical differences in how psychoanalysis might approach integrating AI compared to other forms of therapy. Specifically, psychoanalysis emphasises the role of the unconscious; the association between the patient’s and the analyst’s affects; environmental awareness; and the impact of mentalisation and symbolisation in promoting psychological change (Fonagy & Allison, 2014). Scholars are currently debating whether AI will provide greater access to non-psychoanalytic therapy and at what cost to embodied therapeutic presence.

This review seeks to avoid both technophobic and uncritically optimistic views of AI. It will instead explore how AI might be responsibly integrated into practice, while identifying the areas in which AI systems in their present form cannot supplant intersubjective clinical work. Tasks for which AI could be useful include transcription, documentation, pattern identification, and provision of psycho-educational information. Areas in which current AI cannot supplant the role of the human therapist include embodied presence and participation in unconscious processes.

This paper will engage in a direct dialogue with the February 2026 “Problematising AI” section of the European Journal of Psychoanalysis (Soh, 2026) and specifically with Zapien’s (2026) clinical warnings regarding technology-induced omnipotence and transference leaks; Brahnam’s (2026) questioning of “talking things” vs. “speaking beings”; and Soh’s (2026) call for a new problematic of AI grounded in the clinic and not in theory. The review will combine the contributions made above with Freudian, Lacanian, and relational theories. It will seek to move outside individual criticisms to develop a clinically applicable psychoanalytic framework for evaluating AI use.

Soh’s intervention suggests that the issue is not only how disparate psychoanalytic and adjacent literatures on AI can be integrated, but also the conditions under which theorising about AI becomes possible in the first place. In that sense, this critical narrative review is intended not as a closed or exhaustive theory, but as a provisional psychoanalytic mapping of a still-emergent field. Its synthetic framing reflects both the current fragmentation of the literature and the need to clarify what can and cannot yet be responsibly theorised about AI in psychoanalytic work. The aim is therefore to offer a clinically grounded starting point for further thought rather than a final resolution.

The pace of AI integration into mental health settings has outstripped theoretical and ethical reflection. A 2024 WHO report identified AI-assisted clinical tools as among the fastest-growing areas of health technology adoption globally (WHO, 2024), while recent surveys indicate that a substantial proportion of trainee clinicians are already using AI tools in clinical documentation, case formulation, and supervision without formal institutional guidance (Maheshwari et al., 2025; Shumate et al., 2025). This gap between adoption and reflection constitutes the central problem the present review addresses.Theaddresses. The review will address three primary questions:

1. How do current psychoanalytic theories conceptualise AI metapsychologically and culturally?
2. What does empirical research on AI-mediated non-psychoanalytic psychotherapy and therapeutic relationships demonstrate about alliance, outcome, and relational process?
3. How does AI, in contexts where integration occurs in psychoanalytic work, reshape the analytic field, therapeutic dynamics, and the prerequisites needed for genuine psychological change?

Throughout this review, the term ‘psychoanalytic work’ is used as a working umbrella term that encompasses psychoanalysis (classical and contemporary four- to five-session-per-week analysis on the couch), psychoanalytic psychotherapy (lower-frequency face-to-face treatment informed by psychoanalytic theory), and psychodynamic psychotherapy (time-limited or open-ended therapy drawing on psychoanalytic principles without strict adherence to the analytic frame). Where a claim applies specifically to psychoanalysis—as distinct from psychoanalytic or psychodynamic psychotherapy—this is stated explicitly. Readers should also note that, as discussed in the methodology section below, empirical research on AI in psychoanalysis is, to the present author’s knowledge, virtually absent; the empirical literature reviewed here pertains predominantly to non-psychoanalytic psychotherapy and general mental health settings.

AI is already entering analytic work through at least three distinct pathways. First, it enters through the analysand, as clients increasingly arrive in treatment after using ChatGPT or similar tools for self-analysis, often shaped by fantasies of optimisation and omnipotent control that can disrupt the early formation of transference. Second, it enters through the analyst’s own use, since AI is now being used for documentation, summarisation, and other workflow tasks in clinical settings, with studies showing that GenAI can support documentation while still requiring clinician oversight and responsibility. Third, it enters through supervision, where AI monitoring and feedback systems introduce an evaluative presence into the supervisory triad that can reshape reflection, authority, and relational safety.

This paper is a critical narrative review. It is synthetic rather than merely descriptive, bringing Freudian, Lacanian, relational, and mentalisation-based perspectives into dialogue with empirical findings to develop a unified psychoanalytic framework for thinking about AI in clinical work. Unlike systematic or scoping reviews, it does not aim to be exhaustive in its coverage of AI and non-psychoanalytic psychotherapy literature; instead, it selectively integrates theoretical and empirical material to develop a clinically applicable psychoanalytic position on AI. The main inclusion criterion was relevance to psychoanalytic concepts, including the unconscious, transference, mentalisation, and the analytic field, together with empirical work on AI-mediated therapeutic relationships and supervisory practice.

It should be noted that no empirical studies were identified that examine AI’s clinical application in psychoanalysis. The empirical literature reviewed here pertains predominantly to general psychotherapy, digital mental health, supervision, and psychoanalytic or psychodynamic psychotherapy. Discussions of AI and psychoanalysis in this review are therefore largely theoretical, metapsychological, or speculative.  This is a finding with significant implications for how cautiously any generalisation from the broader literature to psychoanalytic practice should be made.

Psychoanalytic Conceptualisations Of AI

Freudian structural perspectives

From a Freudian structural perspective, AI can be analogised to psychic agencies, but this analogy breaks down precisely where the unconscious is defined by embodiment and conflict. Freud’s (1923) tripartite model provides the basis for Possati’s (2020) integration of neuropsychology and machine learning, in which the architecture of AI is mapped onto Freudian structures. In this system, the basic neural network functions as a base-level Id-like system. It is a repository of optimisation pressures and quasi-instinctual “forces” seeking discharge, comparable to Freud’s (1915) drives (Triebe). The ego appears as an algorithmic control mechanism that mediates between elementary computational impulses and external constraints. It performs a kind of machine “reality-testing” by balancing competing goals. The superego corresponds to the evaluative and ethical regulatory layer, including reinforcement learning from human feedback (RLHF), constraint codes, and alignment protocols, which together constitute an internalised moral architecture.

This structural mapping illuminates several clinical phenomena. Practitioners frequently project the psychic roles they see in their patients onto AI systems, imagining a superego that will judge competence or an ego that will mediate patients’ perception of reality (Possati, 2020). Recognising such projections is essential for ethical practice. Freud (1912) already described humans’ tendency to ascribe psychological structure and intention to external forces; this tendency intensifies when AI simulates responsiveness. These attributions shape how clinicians use AI-generated feedback, at times leading them to defer clinical judgment to a fundamentally non-conscious computational system (Rabeyron, 2025; Ackerhans et al., 2025).

The incompatibility, however, runs deeper than the tripartite analogy suggests. Freud’s structural model is not only a topology of agencies but also a theory of conflict. The unconscious is constituted through repression, a mechanism in which unbearable representations are not simply absent but actively barred from consciousness and preserved in distorted form as compromise formations (Freud, 1915; Freud, 1926). As Freud put it, “the essence of repression lies simply in turning something away, and keeping it at a distance, from the conscious” (Freud, 1915, p. 147). This means the unconscious is not a repository of forgotten content but a dynamic system of ongoing defence, return, and symptom formation. While AI architectures can be mapped onto Freudian agencies at a descriptive level, they cannot host repression or symptom formation. The deeper logical incompatibility is elaborated in the later discussion of Govrin and Matte-Blanco.

At the same time, the limits of the structural analogy are worth noting. Drawing on Matte Blanco’s (1959) distinction between symmetrical and asymmetrical logic, Govrin (2025) argues that, under current computational principles, it is not possible to design an algorithm capable of identifying symmetrical truths, since symmetry violates the basic premises of digital computation. This argument applies to AI systems based on classical digital computation. Emerging paradigms may complicate it in future: Solms (2021), drawing on affective neuroscience and the free energy principle, proposes the theoretical possibility of an ‘artificially conscious feeling machine,’ a system that would generate genuine affective states grounded in homeostatic need rather than simulating them linguistically, and which might therefore satisfy some of the conditions Govrin identifies as absent from current AI. Quantum computing architectures, which can hold superposed rather than binary states, present a separate computational challenge to the symmetry constraint. The present review does not engage these speculative possibilities, focusing instead on the clinical implications of currently available systems.

Equally important, current AI lacks the embodied and affective dimensions that constitute the Freudian unconscious. The drives that animate the id, as Freud (1915) emphasised, arise from bodily needs (e.g., hunger, sexual desire, and aggression) and are intrinsic to a living organism. AI is a non-embodied, non-affective computational system in its current form. It can simulate the logical implications of drive-based behaviour, but it does not possess drives in the psychoanalytic sense; it has no hunger, no erotic longing, and no aggression grounded in fear of death or loss of the organism. Psychoanalysis, therefore, differentiates machine behaviour and black-box opacity from the metapsychological unconscious. The latter is inherently conflict-ridden, affect-laden, historically layered, and tied to embodied experience (Govrin, 2025). 

Digital immersion and uncanny states

Rabeyron (2025) analyses how “the feeling of the uncanny” arises in user exchanges with AIs like ChatGPT, Anthropic, Mistral, DeepSeek, or Grok. They argue that users will relate to the digital other as a thinking being because its responses are similar enough to their own. Yet the digital other is a constructed, artificial entity that responds according to algorithms and, therefore, despite the illusion of similarity, is entirely different from humans.

Psychoanalyst Gutiérrez (2025) provides a psychoanalytic model of digital immersion, moving beyond purely technical descriptions to the psychological activity of inhabiting virtual space. He uses concepts of cathexis, defences, and negative capability to describe how digital environments redirect an individual’s psychological energy and focus. As users submerge themselves in digital environments, they allocate their libido to the algorithmic world and disinvest from their embodied world. Even though machine learning operates through formal statistical relations, it cannot reproduce the cathected, embodied investment through which subjects inhabit digital space.

These clinical dynamics are developed in the surveillance anxiety discussion that follows, where the algorithmic third, transference leaks, and defensive uses of digital immersion are taken up in clinical detail.

Clinical psychologist Zapien (2026), writing from the hub of AI research and development in the San Francisco Bay Area, documents a pattern emerging in their clinical practice: clients increasingly arrive having already consulted AI tools such as Claude or ChatGPT prior to seeking therapy, and this prior consultation shapes the terms on which they enter treatment. Drawing on Zapien’s observations, this review introduces the term ‘pre-treatment AI use’ to describe consultations with AI systems that occur before the initiation of analytic or psychotherapeutic treatment. The always-on availability of AI reinforces an infantile sense of omnipotence and makes both the initiation of treatment (relying upon the analyst) and termination (mournfully accepting limitation) more difficult.

The clinical significance of this pattern extends beyond the presenting moment. When the initiation of treatment is preceded by AI-mediated self-diagnosis and omnipotent expectation, the analyst’s ordinary not-knowing, including their refusal to deliver a formulation immediately, may be experienced as withholding or incompetence rather than as the condition for genuine inquiry. Likewise, termination, which depends on the analysand’s capacity to mourn limitation and accept the analyst’s separateness, is made more difficult when an AI alternative remains perpetually available as an object that does not require mourning. Pre-treatment AI use therefore shapes the entire arc of analytic work, from its beginning to its conclusion.

Possati (2020) identifies a central paradox at the heart of digital immersion: although algorithms cannot reproduce unconscious processes, the human unconscious plays a constitutive role in their creation and design. Possati terms this relationship “emotional programming,” the process by which the desires, anxieties, projections, and unconscious fantasies of designers, engineers, and users are sedimented into the architecture, training data, and evaluative criteria of AI systems. The tool is not, therefore, a neutral instrument. It is the crystallised product of its creators’ and users’ inner worlds, while remaining structurally incapable of hosting an unconscious of its own. This gives rise to the paradox proper: AI systems function as highly effective surfaces for projective identification precisely because they were shaped by unconscious forces they cannot reciprocate. The subject’s internal world creates the very tool that cannot return that subjectivity, a dynamic that intensifies when patients or clinicians attribute authority, understanding, or care to a system that can simulate, but never instantiate, the conditions under which those qualities become therapeutically operative.

Lacanian approaches and AI as big Other

Systems like ChatGPT can be understood as occupying the Lacanian position of the Big Other and the subject supposed to know, with many users treating them as authoritative knowledge sources and guarantors of meaning (Black & Johanssen, 2025). In this position, AI appears as the Other who knows what the subject means, attracting transference, demands, and fantasies of recognition that would otherwise have been directed toward a human authority. The way AI is positioned, therefore, reshapes how relational processes unfold.

Rabeyron (2025) similarly describes AI as a representation of the Lacanian subject supposed to know everything. It is imaginatively endowed with limitless memory and knowledge and with the capacity to answer questions definitively, without hesitation or apparent limit. This rests on Lacan’s (1959–1960) account of the analyst’s symbolic position, in which the analyst is unconsciously invested with imagined completeness, omniscience, and a lack-free status. A central task of psychoanalytic work is to identify and gradually dismantle these transference positions so that the analysand can recognise that both analyst and analysand are subjects of lack, desire, and limitation (Rabeyron, 2025). ​

AI can make the concept of omniscience less a product of fantasy than ever before. Compared with humans, current AI systems can be trained on much more information, act very consistently, and do so without many of the normal limits of human intelligence (Black & Johanssen, 2025; Rabeyron, 2025). As a result, the fantasy of the all-knowing Other is at least partially achieved through AI’s capacity, making the transference to AI even more difficult to analyse and break down.

Rather than encountering a purely metaphorical Other who knows everything, users will experience an operating system that has incorporated an extensive corpus of texts, languages, and perspectives. Possati (2020) describes this as a stabilisation of the transference that results when the subject-supposed-to-know (the figure onto whom knowledge and authority are projected) is not simply a ghost to be interpreted, but rather an operational fact to be managed.

The greater the perceived success of AI as flawlessness or near-flawlessness, the greater the illusion of a non-lacking Other becomes. There is less ability to accept ambiguity and uncertainty, and therefore less negative capability (Keats, 1817; Bion, 1970) to allow subjects to exist in a state of not-knowing without rushing to some form of closure. As acceptance of uncertainty and ambiguity decreases, the ability to express creativity in those areas will also decrease, limiting the types of doubt that lead to true understanding and transformative change (Frankfeldt, 2025).

Geal (2026) expands upon the Lacanian framework by suggesting that large language models represent a perverse denial of the partiality and capitalist determination of those systems and a psychotic non-human subjectivity in which the conscious human agency of the subject is no longer required. It is exactly this stabilised transference and illusion of a non-lacking Other that the current review sees as unworkable via analytic working-through. Pavón-Cuéllar (2026) adds that the repression of the subject is part of a larger trend in scientific knowledge toward the technological jouissance of capital, that is, the excessive enjoyment built into capital’s operations.

Therefore, the transference that occurs with AI is fundamentally different from that of the typical analytic situation. Unlike the clinical situation where the analyst’s inevitable lack ultimately reveals their incompleteness, thereby leading to disillusionment and working-through, with current AI, there is no corresponding feeling of incompleteness. The opacity of the decision-making process of AI (“the black box”) paradoxically enhances the belief that the system is complete because users are unable to demystify how the system generates its answers. Consequently, they are denied the disillusionment that is often necessary for analytic insight. This review places less emphasis on whether AI is a subject and instead emphasises its role as the Big Other in the user’s psychic economy. We consider AI to occupy the location of the Big Other, yet recognise that it is ontologically a non-subject. The distinction is important; we describe a transference dynamic in which users unconsciously project omniscience onto AI, not that AI has now become a structural element of the symbolic order.

Mentalisation, epistemic trust, and simulated empathy

Yirmiya and Fonagy’s (2025) model of “mentalising without a mind” describes how generative AI can simulate certain aspects of mentalising. They show that AI can approximate processes such as taking perspectives, clarifying emotions, and reflecting on alternative points of view. Generative systems can therefore produce mentalising-like discourse and display reflective thinking. However, this reflection does not yield genuinely human responses: it lacks affective responses between people, mutual regulation, and embodied, temporally attuned responsiveness.

AI-based systems can also help patients reduce symptoms and provide a sense of being understood by others. However, if patients do not recognise that the epistemic trust is based upon an illusion, then when patients interact with the AI, they may perceive it as caring for their welfare, interested in their well-being, and providing insight into their experiences.  Yet, AI in its current form can never recognise these patients as individuals or as subjects (Yirmiya & Fonagy, 2025). Similarly, clinicians, whether in a therapeutic or supervisory capacity, who utilise AI to obtain feedback or clinical guidance may perceive its responses as neutral, unbiased, and objective; this may increase the potential for confusion between perceived simulated empathy from the AI and actual human attunement (Yirmiya & Fonagy, 2025).

According to Yirmiya and Fonagy (2025), developing epistemic humility is the most critical requirement for utilising AI responsibly. Clinicians must be able to acknowledge not knowing while maintaining authority and care. The materials generated by AI should be introduced to the clinician and the patient as a hypothesis for joint exploration; thus, the clinician and patient can co-create new possibilities and understandings, rather than the clinician presenting the generated materials as a definitive statement (Yirmiya & Fonagy, 2025). If clinicians instead use the AI as a source of pseudo-certainty, thereby displacing both their own and the patients’ uncertainties, they eliminate the negative capability, creative play, and rupture-and-repair processes that are necessary to maintain analytic work (Yirmiya & Fonagy, 2025).

Emotional communication and the limits of AI simulation

Emotional communication, according to Frankfeldt (2025), is an evolutionary, subcortically grounded mode of relating that differs essentially from cognitive and linguistically based exchanges. It is defined by its capacity to elicit an emotional response in the recipient without requiring mentalising capacities or conscious cognitive understanding. Frankfeldt (2025) traces emotional communication to emotional induction, a complex, non-separable interpersonal process grounded in embodied presence. In this sense, it stands in sharp contrast to present-day AI architectures, which are cognitively oriented. While AI is capable of recognising emotionally patterned text and speech, it remains incapable of engaging the non-verbal, symbolic dimensions of emotional communication, such as tone, rhythm, facial expressions, gestures, and affective resonance. These are the features that constitute full emotional communication as it is presently understood. In this sense, Frankfeldt’s account situates emotional communication as a primarily embodied, non-verbal process (Frankfeldt, 2025).

The ability of AI to simulate cognitive empathy differs from affective empathy, which requires emotional involvement, sustained attention, and Winnicott’s concept of primary maternal preoccupation, i.e. the analyst’s state of total affective attentiveness to the patient (Frankfeldt, 2025; Winnicott, 1960, 1965). People’s emotional connections to others drop dramatically when they learn that an AI generated the apparently empathetic response they received, even if the content is identical (Yirmiya & Fonagy, 2025; Rubin et al., 2025). Awareness of the authentic engagement of human consciousness is therefore crucial to successful therapy.

Recent multimodal machine-learning research on non-verbal communication in psychotherapy has provided empirical support for these distinctions. Using facial action units, gaze direction, and body posture, Hau and colleagues (2025) analysed dyadic interactions across seven psychoanalytic sessions. When they applied traditional machine-learning techniques to cluster non-verbal episodes, clustering purity remained close to chance, at approximately 50%. One patient in their pilot terminated treatment early, and the authors noted that the use of AI and machine learning appeared to activate unconscious counteracting forces that limited the freedom to engage in the therapeutic process (Hau et al., 2025, p. 60). These findings suggest that emotional communication arises from a relational and temporal logic that current methods struggle to represent and that AI-mediated monitoring can itself compromise the sense of safety required for such communication to unfold.

The Logical Incompatibilitiy: AI And The Unconscious

Symmetrical vs classical logic: Govrin’s structural analysis

As noted in the earlier structural discussion, Govrin (2025), drawing on Matte-Blanco (1959), argues that current digital architectures are structurally mismatched with unconscious processes. In unconscious life, mutually opposed meanings can coexist without requiring resolution. From a psychoanalytic perspective, states such as loving and hating the same object are not errors to be corrected, but tensions to be sustained and worked through.

Govrin’s use of Matte-Blanco therefore sharpens the limits of the Freud–AI structural analogy introduced earlier in the manuscript. While AI can simulate aspects of conflict in its outputs, it cannot itself be the site of unconscious conflict in the psychoanalytic sense, because it does not repress, does not generate symptom formations as compromise solutions, and does not exhibit the return of the repressed (Freud, 1915, 1926; Govrin, 2025; Matte-Blanco, 1959). The issue is therefore not simply whether AI can mimic complex mental content, but whether it can instantiate the conflictual, overdetermined, and symbolically reversible logic that psychoanalysis attributes to the unconscious.

Clinically, this incompatibility matters most in how AI is used to formulate and evaluate psychic life. Where the analyst’s task is to help the patient remain with both-and experiences, such as love and hate, longing and fear, or the wish to change alongside an investment in symptom, AI-driven tools will tend to translate these dynamics into reductive outputs such as diagnostic labels, risk scores, or sentiment classifications. Such outputs may be technically useful, but if they are treated as capturing the patient’s truth, they risk foreclosing the conflict and ambiguity through which symbolisation and analytic change become possible (Govrin, 2025; Brahnam, 2026). For this reason, AI-generated formulations are best understood as partial adjuncts to human interpretation rather than as definitive accounts of unconscious life.

Embodied holding and intercorporeal attunement

Therapeutic action is also grounded in embodied relational presence. Fuchs (2013) and Frankfeldt (2025) emphasise that therapeutic relationships have an intercorporeal quality: therapists and patients incorporate each other’s somatic experience through physical presence and responses to subtle bodily signals, creating an interconnected space for shared affective experience. This space is shaped by breathing patterns, posture, gestures, and micro-adjustments in distance and orientation that are perceived and responded to largely outside conscious awareness.

Polyvagal theory and somatic marker research corroborate Winnicott’s intuition about the centrality of embodied attunement in therapy. A therapist’s capacity to facilitate therapeutic action depends on subcortical processes that operate beneath conscious awareness and require the presence of a living, breathing, emotionally responsive body (Porges, 2011; Winnicott, 1960, 1965). When a therapist leans forward from genuine concern, softens their voice to match a client’s despair, or regulates their own state to meet a client’s dysregulation, therapeutic action occurs in a shared bodily field that cannot be reduced to verbal content alone.

Taken together, these accounts depict therapeutic action as grounded in intercorporeal, subcortical processes that presuppose a living, responsive body, rather than a computational system (Fuchs, 2013; Frankfeldt, 2025; Porges, 2011).

Projective identification and pathological adaptation

Boundary violations in the therapist’s experience are important clinical data in psychoanalytic treatment. The therapist’s feeling of being “engulfed” by the patient is a central example. Govrin (2025) presents the case of Rebecca, a patient with profound relational trauma who had learned to dissolve her boundaries to protect against abandonment. In treatment, this manifested as constant reassurance-seeking, fusion with the immediate moment, and reliance on the analyst’s continuous physical and emotional presence for validation.

For Rebecca, lateness, holidays, or the analyst’s boundary-setting confirmed her conviction that “no one will ever truly care about me.” Gradually, through recognition of previously unacknowledged aspects of her experience in another human being—one who could set limits precisely because of their embodiment, mortality, and separateness—she began to rework this template. The analyst’s separateness was thinkable only because it was grounded in a similarly embodied, mortal, and distinct subject.

A human therapist thus has an embodied experience of boundary strain or violation that signals pathology. An AI, by contrast, risks becoming a “perfect object” for fusion fantasies: always available, boundaryless, and endlessly accommodating. Rather than disputing pathological patterns, such an object may reinforce them.

Repeated reliance on AI to maintain a constant connection without reciprocal accountability may further strengthen adhesive identifications. Patients can come to expect that “good-enough” attunement entails rapid response, continuous presence, and the suppression of the therapist’s own affective responses. When a human therapist inevitably fails to meet this algorithmic ideal, because they experience fatigue, countertransference, and the need to maintain professional boundaries, the patient may experience such limits as betrayal and misread ordinary therapeutic boundaries as coldness or rejection.

A clarification is warranted before proceeding to empirical findings. The clinical and ethical arguments developed in this review concerning surveillance anxiety, the algorithmic third, the erosion of transitional space, and the conditions for genuine therapeutic change do not depend on resolving whether AI could, in principle, ever possess unconscious properties or embodied experience. Even if a future AI system were to satisfy these criteria, the transference dynamics, performativity pressures, and epistemic risks documented here would remain clinically significant. The theoretical incompatibility argument and the clinical argument are therefore treated as related but separable contributions.

Empirical findings on AI-mediated psychotherapy

Having established the theoretical grounds for AI’s incompatibility with unconscious process, this critical narrative review now examines what empirical research reveals — not to resolve this tension, but to hold it: at present AI can produce measurable symptomatic relief without thereby accessing the depth at which psychoanalytic transformation occurs.

The empirical literature reviewed below spans psychoanalytic and selected broader non-psychoanalytic psychotherapy and mental health research. This breadth is intentional but requires acknowledgment. Much of the outcome research, on symptom reduction, therapeutic alliance, and chatbot efficacy, has been conducted within cognitive-behavioural, integrative, or general mental health frameworks whose aims, outcome measures, and theory of change differ from those of psychoanalysis. Whether reductions in self-reported depression symptoms constitute the kind of transformation psychoanalysis seeks is itself a contested question, as Fonagy and Allison (2014) note. The review draws on this broader literature to contextualise AI’s capabilities and limitations in the clinical field, while recognising that its implications for specifically psychoanalytic practice require interpretation through the theoretical framework developed above. Where differences between psychoanalytic and general psychotherapy evidence are clinically significant, they are noted.

Symptom reduction and availability 

Research indicates that AI-assisted interventions and chatbots can reduce symptoms of depression, anxiety, and eating disorders and can expand access to support for people who face geographical, temporal, or economic barriers to care (Yirmiya & Fonagy, 2025; Vaidyam et al., 2019; Heinz et al., 2025; Rabeyron, 2025). Users in some studies have rated AI-generated responses as highly empathic, and in a few trials AI-based interventions have achieved short-term symptom reductions for depression and anxiety in the range of those observed in other brief, structured treatments (Ayers et al., 2023; Heinz et al., 2025). At the same time, these findings arise almost entirely from cognitive-behavioural and common-factor frameworks, and the outcome measures used in this literature do not straightforwardly capture the kinds of analytic transformation that psychoanalysis aims at, such as changes in unconscious conflict, capacity for symbolisation, and subject-to-subject relating (Fonagy & Allison, 2014). 

The alliance question and common psychotherapeutic factors

Research on alliance with AI suggests that users’ experience of empathy and relational fit depends heavily on what they believe the system is, not only on what it says or does (Yirmiya & Fonagy, 2025; Rubin et al., 2025). Therefore, it appears that how one perceives the relationship with another entity is primarily influenced by the perceptions of the “other” in the relationship, rather than the actual characteristics or behaviours of the interaction itself.

As for trust and attachment, the evidence has been inconsistent; individuals who identify themselves as having an anxious attachment style seem to show greater willingness to interact with AI-based systems.  However, lonely individuals—a group likely to derive the greatest benefit from therapy—report lower levels of trust in AI (Wu et al., 2025).

Additionally, there is considerable flexibility in how people perceive empathy. Users who perceive AI as empathetic report forming stronger alliances, even when the system’s technological underpinnings remain the same. Perceived alliance strengthens when AI is experienced as empathic but drops when users are explicitly reminded that the response is algorithmically generated (Rubin et al., 2025). The results illustrate that beliefs about the nature of the “other” in the interaction can influence common therapeutic factors (Wampold & Imel, 2015). Giotakos (2025) applies this framework directly to AI-based psychotherapy, arguing that therapeutic alliance, empathy, and common goals, the core common factors, remain the primary determinants of outcome even in AI-mediated contexts, though their implementation calls for substantial adaptation.

Limitations and risks

A prior limitation bears stating explicitly: this review identified no empirical studies examining AI’s clinical application in psychoanalysis, as distinct from psychoanalytic or psychodynamic psychotherapy. The empirical literature reviewed here pertains predominantly to general psychotherapy, digital mental health, and supervision. Discussions of AI and psychoanalysis throughout this review are therefore largely theoretical, metapsychological, or speculative. Readers should exercise caution in generalising any empirical findings reviewed here to the practice of psychoanalysis.

Beyond this evidentiary gap, current AI may simulate therapeutic form, but it does not participate in the mutual mentalising and embodied intersubjectivity through which psychoanalytic change is understood to occur (Yirmiya & Fonagy, 2025; Giotakos, 2025). This limitation is especially important for patients with complex trauma, relational vulnerability, or impaired mentalising, for whom pseudo-empathic responses may intensify confusion, dependency, or delusional use of the system rather than support reflective change (Yirmiya & Fonagy, 2025; Frankfeldt, 2025).

From an attachment perspective, AI-based interventions raise the question of whether epistemic trust acquired in relationships with nonhuman entities generalises to human relationships—a question that the foundational framework of epistemic trust (Fonagy & Allison, 2014) does not yet address but that emerging research on attachment and AI adoption is beginning to examine (Békés & Aafjes-van Doorn, 2026). If humans learn to trust entities that cannot genuinely affect them, they may come to expect less from human responsiveness and revise downward their sense of what authentic care entails.

Genealogy of the Third in Psychoanalytic Work: From Ogden to Algorithmic Presence

The analytic third: Ogden’s foundational concept

Ogden (1994) gives the clearest formulation of the analytic third, a generative third that emerges from the intersubjective experience of analyst and analysand. Through their mutual interaction, they create, negate, and preserve one another, establishing a shared psychic space that is neither analyst nor analysand nor a simple sum of the two. The analytic third is a phenomenological reality, available to the analyst through reverie, somatic experience, and countertransference. Within this third process, ruptures in the analytic relationship are repaired, and the analysand’s incompleteness meets the analyst’s, forming a container for transformation (Ogden, 1994).

It is an intersubjective creation between analyst and analysand, accessed through the analyst’s embodied, affective experience, and it presupposes two human minds engaged in substantive psychological work. The analytic third thus exists only insofar as the analyst can exercise negative capability, i.e., the capacity to tolerate uncertainty and not-knowing and to be authentically affected by the analysand’s experience.

The artificial third: Concrete, interactive AI

Concerns over what constitutes creative and analytic space, especially when a third, technology-driven participant enters the equation, are shared by several contributors to the current edition of the European Journal of Psychoanalysis, including Heimann’s (2026) examination of “free association” as it relates to machines, Jihnson’s (2026) demonstration that large language models (LLMs) are analogous to Freud’s original psychical apparatuses, even though they are excluded from the dialectic of analysis, and Tutt’s (2026) argument that actual creativity is a “mutual day-dream” whose social structure cannot be replicated through mechanical production.

The transition from an intersubjective human third to a machine third creates significant and immediate questions regarding creativity and analytic space in our AI era.

For the first time in history, the “third” in psychotherapeutic treatment has taken on a qualitatively new shape in its current technological form. It is no longer only a metaphor or a product of intersubjective relations; it is a tangible, interactive participant. While the third previously arose from the embodied intersubjectivity of the dyad in therapy, it is now created through artificial learning systems and neural networks (Haber et al., 2024). This entity is referred to as the “Artificial Third,” as described by Haber and colleagues: “a concrete and symbolic representation of GAI in the cultural-political-digital space and the shaping presence of GAI in the therapeutic encounter, in society and in the self.” In contrast to the analytic third, the Artificial Third can interact with both the therapist and the patient. Its mechanisms may be challenged, though usually in an obscure manner, and its outputs are commonly perceived as authoritative statements rather than hypotheses (Haber et al., 2024).

In their paper, Haber et al. locate the Artificial Third in a broad psychoanalytic lineage of thirdness: Freud’s Oedipal third (the father as ultimate other), Winnicott’s transitional space (the interstitial space between imagined and actual reality), and Ogden’s analytic third (an intersubjective relational space). Now, there is a fourth configuration: one in which the third is not co-constructed by two humans but is instead created by a technological apparatus that develops its own formulations, feedback, and interpretations.

The algorithmic third: Extension to supervisory dynamics

Haber et al. (2024) locate the Artificial Third within the therapeutic dyad; this review extends their concept into the supervisory realm by introducing the algorithmic third as a distinct phenomenon. As used here, the ‘algorithmic third’ refers specifically to AI systems that continuously monitor live clinical sessions, recording, transcribing, and scoring interactions in real time, and return evaluative feedback to supervisors, supervisees, or institutions. It appears when such systems enter supervision, a setting that has historically been protected as a space in which supervisor and supervisee develop reflective capacity with some insulation from the performance pressures of the primary therapeutic encounter. Supervision has functioned as a second-order analytic third, allowing supervisor and supervisee to develop a reflective environment partly shielded from outcome demands. When AI is used to oversee and provide feedback on supervision, a novel configuration arises: the algorithmic third becomes an evaluative presence within supervision, substantially altering the conditions under which the supervisory dyad can generate its own analytic third.

The algorithmic third (Allan et al., 2025) is thus distinguished from the Artificial Third by its location and function in supervision. It modifies the interactions between supervisor and supervisee by introducing a third, evaluative presence. By monitoring, scoring, and providing feedback, it implicitly demands performance optimisation and reorients attention toward its metrics. In contrast, the Artificial Third is typically envisioned as assisting human therapeutic work within the treatment dyad.

Surveillance anxiety: Empirical and theoretical grounding

The current review relates these clinical phenomena to surveillance anxiety and the emergent presence of an algorithmic third. When an algorithmic response is incorporated into clinical practice, the omnipotent fantasies and transference leaks (see Digital Immersion and Uncanny States, above) are intensified. When digital monitoring and/or algorithm-based feedback are used to organise clinical activities, the clinician’s libidinal energies become oriented toward obtaining the algorithm’s approval. Thus, they begin to perform for the algorithm. The energy behind their interpretations is then subtly shifted to optimise the algorithm’s metrics rather than serve the client’s needs. The analyst’s position is thereby compromised: the client’s ability to authentically free associate is now secondary to obtaining the algorithm’s approval.

Furthermore, Gutiérrez (2025) argues that digital immersion functions as a defense mechanism against reality. Through digitally managed interactions, people can avoid embodied encounters and experiences. In AI-supervised clinical practices, this defensive behaviour manifests as an increasing reliance on algorithmic approval to support the clinician’s actions and decisions, rather than developing the ability to tolerate ambiguity, uncertainty, and emotionally painful material.

The “surveillance anxiety” experienced by clinicians and supervisees who work within the algorithm’s view is a different type of pressure from usual performance anxiety. Surveillance anxiety is present in at least two main ways. The first is that there is always some uncertainty when working with algorithms. Supervisees rarely have information on the timing, frequency, and methods used to monitor their clinical activity or on the methodology used to analyse the data. Therefore, even if the algorithms do not necessarily pose an immediate danger, the uncertainty creates a constant low-level alertness in supervisees and contributes to a state of chronic hypervigilance (Glavin et al., 2024).

The second mechanism is what may be termed the “performativity burden”. Because supervisees are aware that an algorithmic system is continually analysing their activities, they begin to control their own presentations in clinical sessions. They no longer respond spontaneously but instead restrict their responses to only what they believe the algorithm will accept and what they believe others viewing the algorithm’s output will consider acceptable. This performativity burden reduces spontaneity and increases the likelihood that supervisees will provide less honest and authentic responses to patients to protect themselves from potential criticism (Topor et al., 2017; Haggerty & Hilsenroth, 2011).

Studies of workplace electronic monitoring help explain these mechanisms. Using structural equation modelling, Glavin et al. (2024) studied perceptions of surveillance and psychological distress among 3,508 employees. They concluded that employee perceptions of surveillance indirectly affected psychological distress via three paths: reduced autonomy at work, increased job pressure (explaining approximately 55% of the total variance), and perceived privacy violation.

There are parallel findings in the supervision literature. For decades, researchers have shown that psychotherapists-in-training typically exhibit increased anxiety when recorded during therapy training (Topor et al., 2017). Research has also demonstrated that videotaping therapy leads to lower supervisors’ ratings of trainees’ self-disclosure during supervision and to a decreased willingness to share clinical content during supervision (Haggerty & Hilsenroth, 2011). Similarly, Haggerty and Hilsenroth (2011) reported that the use of video tapes to record therapy training had documented benefits, but that “actual use of video recordings [was] limited by learner anxiety related to the use of recordings,” with learners reporting concerns that someone would review the “good, the bad and the ugly” of their sessions (p. 201).

Further research has indicated that the knowledge that one is being observed while performing a task can increase anxiety and paranoia in individuals who are prone to these states (Malik et al., 2024). As such, this produces a long-term affective state rather than a short-term feeling of discomfort. Furthermore, because algorithmic observation may produce repeated episodes of defensiveness in supervisees who have experienced relational trauma or anxiety disorders, it represents a special risk for those supervisees.

AI in the Analytic Field: The Shift in Analytic Focus

The concept of the algorithmic third and transference

In the supervisory setting, the algorithmic third operates as an observing and evaluating presence that sits alongside supervisor and supervisee rather than remaining outside the clinical situation. This shifts the focus from exploration and authenticity in relationship development to performance management and defensive presentation.

From a Lacanian standpoint, the placement of AI into the position of the subject supposed to know is predictable (Black & Johanssen, 2025; Rabeyron, 2025). Clinical reports have supported this: supervisees and patients report AI as all-knowing, all-seeing, all-powerful, and infallible. This represents a displacement of the subject-supposed-to-know from the human analyst/supervisor to the algorithm (Allan et al., 2025). When AI-generated rupture markers, scores, or formulations are considered more reliable than human appraisal, supervisory/analytic dialogue can quickly reorganise itself around “what the machine says,” rather than the living field of analyst, analysand, and supervisor.

This displacement brings significant challenges to analytic work. Maintaining collective negative capability and epistemic humility – both essential parts of mature analytic relationships – becomes increasingly difficult as the apparent authority of algorithms seems to resolve ambiguity (Yirmiya & Fonagy, 2025). These pressures make it easy to abandon not-knowing in favour of algorithmic certainty, a risk that must be addressed through the stance outlined in the practice principles below.

Conceptualising transference in AI–human encounters

Holohan and Fiske (2021) offer a nuanced account of how transference must be reconceived in AI–human psychotherapeutic encounters. They show that users do form transference relationships with AI systems. However, these differ from conventional transference because the therapist’s incompleteness is concealed rather than becoming visible through rupture and limitation.

Erosion of analytic holding and containment

Within a Bionian framework, the analytic process depends on the analyst’s capacity to “take in” the analysand’s projections and affective states through reverie and countertransference, and to return them in a transformed, digestible form—the container–contained (Bion, 1962). This is inherently relational and presupposes a present, emotionally available, subjectively engaged analyst. Introducing AI into the analytic field disturbs this holding function. When sessions are recorded or transcribed and then analysed by an algorithm that issues feedback or recommendations, the analyst’s capacity to metabolise unconscious material is compromised. Analysts report that awareness of an algorithmic observer alters their clinical stance, shifting focus away from the internal field of the therapeutic relationship toward external, metric-based evaluation (Haggerty & Hilsenroth, 2011).

AI in supervisory practice

Studies of AI in supervision by Allan et al. (2025) and Haber et al. (2024) show that a central unintended consequence is degradation of the primary supervisory function. A recent meta-analytic study by Schreyer et al. (2025) found that supervision improves therapist competency and alliance quality but has small and statistically nonsignificant effects on patient symptomatology. These modest benefits derive mainly from relational focus and structure, raising concerns that AI-mediated feedback could attenuate them by undermining relational elements.

AI can perform useful technical tasks such as real-time transcription, thematic organisation of client material, and checking adherence to manualised protocols. It cannot, however, reproduce the parallel process—the recurrence of therapeutic dynamics within supervision—that is critical to learning in psychodynamic training at present.

Supervisees report that when algorithmic data displays and optimisation recommendations dominate supervisory time, they have less space to explore their subjective experiences of countertransference, relational rupture, and professional development (Topor et al. 2017). The supervisory relationship risks being reduced to the implementation of algorithmically defined best techniques rather than a site of reflective, experiential learning.

De-skilling and professional identity threat

Algorithmic authority poses a threat to professional identity as well; it presents clinicians as inferior to “objectively” judgemental machines. Ackerhans et al. (2025) describe the existential threats to professional identity that clinicians perceive when they view algorithms as superior authorities for diagnosing patients. Similarly, during supervision, when supervisees view an algorithm’s evaluation as definitive, they report reduced clinical confidence and a sense of losing skills rather than developing them.

There is a paradox involved in this process. While explainable AI was developed to increase clinicians’ trust in these systems, the transparency of their evaluation criteria increases anxiety because clinicians now have an explicit set of standards against which they can evaluate their own performance. These systems do not enhance clinicians’ abilities to learn; instead, they create increased surveillance-related anxiety and encourage increased levels of performance-based compliance among clinicians.

Winnicott’s transitional space and the loss of creative play

Winnicott (1971) emphasised that both therapeutic and supervisory relationships require a transitional space characterised by uncertainty, play, and creative experimentation, without shame or premature evaluation. The transitional space is an inherent part of the relationship between the clinician and patient; the clinician must tolerate uncertainty and risk by remaining open to the patient’s internal world, and responding with authentic interest.

According to Allan et al. (2025), the use of AI systems that rely on precision and quantifiable measures of success constrains the transitional space. It therefore limits the potential for this type of creative experimentation. Because supervisees receive constant algorithmic feedback regarding relational aspects of their interactions (i.e., empathy or humour), they tend to reduce experimentation and utilise defences to protect themselves. The transition from the clinician’s internal, subjective evaluation of their own performance to algorithm-based evaluations of performance creates obstacles to clinicians engaging in reflective practice, reduces their confidence in their ability to express themselves clinically and limits the space for creative play that facilitates psychoanalytic development.

Ethical and Political Stakes

Surveillance, privacy, and informed consent

Given the increasing prevalence of AI in the fields of psychotherapy and psychoanalysis, there exist major concerns regarding privacy, informed consent, and the likely psychological effects of continuous surveillance on both patients and analysts. Many supervisees will not recognise the considerable impact this type of observation has on their behaviour and development until after the fact. Informed consent processes typically emphasise data security without addressing the relational and transferential consequences of employing AI in analysis and supervision. Furthermore, given the inherent power dynamic in a supervisory relationship, supervisees may not have the option to refuse the use of AI in their analysis and supervision, thereby raising questions about the validity of their “consent.”

In addition to issues related to informed consent, the collection and storage of therapeutic and supervisory data by AI systems present additional privacy-related risks. For example, data breaches, hacking, and/or the secondary use of clinical data for non-clinical purposes can compromise confidentiality. Although data may be encrypted and anonymised, it remains ethically and relationally tied to the analyst and the patient. Knowing that sessions are being recorded, stored, and analysed by machines can create feelings of unease and potentially limit patients’ ability to disclose information openly during sessions.

Epistemic rights and the dangers of pseudo-expertise

AI systems also risk assuming an authoritative interpretive role by occupying the position of the subject-supposed-to-know. In such a configuration, patients may be positioned as passive recipients of algorithmic interpretations rather than as active co-constructors of meaning (Black & Johanssen, 2025). This is particularly problematic given that much of psychotherapy depends on recognising the client as a knower of their own experience, capable of generating their own understanding of symptoms and suffering (Frankfeldt, 2025).

Compounding this, AI systems lack metacognitive awareness of their own limits (Govrin, 2025; Yirmiya & Fonagy, 2025). When these systems are used beyond their appropriate scope and without explicit acknowledgement of epistemic boundaries, the risk of overstepping and error increases. If AI fails to sustain what Frankfeldt (2025) calls “emotional communication”, a viscerally grounded, co-constructed process of meaning-making between client and therapist, then it compromises their epistemic agency.

Maheshwari et al. (2025) introduces the notion of “institutional epistemic trust,” the trust placed in systems that are accountable for their claims. This form of trust becomes dangerous when directed toward systems that appear accountable but lack genuine accountability. Algorithmic authority does not entail the same level of human responsibility: patients and supervisees entrust their understanding and meaning-making to entities that cannot be held accountable for harm. Because an AI system cannot bear moral responsibility for interpretive error, no single human agent is readily accountable in the way a clinician would be.

Commodification of care and professional identity

Allan et al. (2025) argue that AI in supervision exemplifies an expanding commodification of mental health services and professional training. AI’s capacity to generate measurable output, session counts, competency scores, and adherence rates invites the construction of quantifiable performance metrics and seemingly clear training targets. Yet many of the most important developmental aspects of psychotherapy are difficult to measure or standardise. A supervisory relationship that was once relational and developmentally oriented risks becoming a production-line process organised around outputs rather than growth.

This broader political economy also shapes how AI-mediated creativity is commodified and constrained.

Tutt (2026), examining the political economy of labour as it relates to the commodification of AI-mediated creativity, argues that “AI Panic” is created primarily by Capital’s contradictory relationship to work, not by the autonomy of technology. What cannot be replicated is the mutual daydream, that is the shared social space in which private fantasy is transformed into collective meaning. The algorithmic third, therefore, jeopardises not only the individual’s clinical space but also the conditions required for both aesthetic and political imagination.

Governance and regulatory structures

To date, AI in mental health can theoretically be regulated using specialised regulatory regimes or existing regulatory regimes for general AI and medical devices. The World Health Organisation’s (WHO) guidance on the ethical and legal governance of AI for health emphasises that regulatory approaches need to move beyond making sure that AI systems perform safely and correctly to include relational, epistemological, and justice-oriented considerations, especially given the increasing use of large multimodal models in standard clinical practice (WHO, 2024). WHO suggests the development of context-specific governance regimes that will preserve individuals’ autonomy and confidentiality from harm, limit the harms of bias and exploitation, and ensure that responsibility for harm is traced to humans rather than dispersed throughout opaque systems.

Shumate et al. (2025) conducted a review of legislative initiatives across the 50 states regarding AI in mental health. They found that current AI governance in this area is highly fragmented and reactive, and primarily focused on protecting consumers’ personal data and rights. At both the state and local levels, there is little consideration of how AI algorithms impact the therapeutic relationship. Shumate et al. (2025) describe the “patchwork” of statutes and bills regarding AI in mental health as rarely distinguishing between AI used as an analytical instrument for administrative purposes. AI is positioned as a quasi-clinical actor, thereby leaving many important questions regarding accountability, scope of practice, and protection for at-risk groups unanswered.

The WHO’s international standards, along with Shumate et al.’s (2025) analysis of legislation, suggest that a governance regime for AI in mental health that includes principles of psychoanalysis, requires a regulatory approach that extends beyond technical safety to encompass protection of transitional space, promotion of epistemic justice, and safeguarding of clinicians’ reflective stance.

Implications and conclusions

Using AI in psychoanalysis and psychotherapy is neither a categorical failure nor a technological saviour; it poses clear challenges calling for both psychoanalytic thinking and clinical experience. This critical narrative review brought basic psychoanalytic concepts into dialogue with the empirical and theoretical work collected in the European Journal of Psychoanalysis’s “Problematising AI” discourse to address three core questions. The outcome is a clinically oriented integration: From both the structural/metapsychological and the clinical perspective, the European Journal of Psychoanalysis studies lead to one fundamental conclusion. Brahnam (2026) asks if there is any way to automate speech without undermining the ethical and existential implications of analysis; Heimann (2026), Jihnson (2026), Tutt (2026), Geal (2026), and Pavón-Cuéllar (2026) all indicate that LLMs produce structural equivalents of free association and yet are unable to engage in genuine analytic dialogue; and Pavón-Cuéllar (2026) and Geal (2026) show how AI is embedded in larger structural systems of perverse disavowal and the capitalist suppression of the subject.

In practice, the theoretical consensus is evident at the clinical level as well. Zapien’s (2026) descriptions of omnipotent reinforcement and transference leaks, Soh’s (2026) appeal to establish a clinic-grounded problematic amid symbolic decline, and Pavón-Cuéllar’s (2026) warnings against the disappearance of the subject all arrive at the same conclusion when viewed through the lens of Freudian, Lacanian, and relational perspectives.

AI produces metonymic chains and even creates machinic free associations. However, because current AI systems lack embodied drive, lack [in the Lacanian sense], and desire, they cannot create the intersubjective space in which authentic analytic transformation occurs. This review therefore endorses an epistemic humility stance in which AI remains a hypothesis-generating adjunct embedded in human interpretive work. Importantly, the near-absence of psychoanalysis from the AI outcome literature is not simply a gap to be filled by future research. It may reflect a structural mismatch between the methods needed to study psychoanalytic change and the kinds of outcomes AI-mediated research is currently built to measure — a finding with implications for how the field designs and interprets future studies, and for how cautiously the current empirical literature on AI and therapy should be applied to psychoanalytic practice.

Novel contributions and practical implications

Two practical implications arise from these new understandings:

1. AI systems can be employed responsibly for supportive functions (transcription services, documentation support, identifying patterns, providing psychoeducational materials). However, there is no evidence that they can supplant the therapist’s corporeal presence, or the ability of the therapist to access and utilise unconscious processes.
2. Therapist discomfort with AI-based monitoring should be understood as a reflection of the degree of transitional space being diminished, and a decreased ability to accept ambiguity and uncertainty.

Principles for responsible practice and future directions

Several principles emerge from this synthesis for ethical integration of AI in psychoanalytic contexts, alongside directions for future research and professional development.

1. Epistemic Humility. Consistent with the epistemic humility approach outlined above, clinicians should present AI outputs as provisional hypotheses rather than authoritative formulations. Supervisors should model this stance and help trainees distinguish pattern recognition from psychoanalytic interpretation, while remaining alert to the authority technology can acquire in analytic practice.
2. Preservation of Embodied Presence. Psychotherapy relies upon the embodied, emotionally responsive presence of the therapist. AI should never replace the analyst’s presence in the intersubjective field, even though AI can be used to assist with auxiliary functions. Training programs should avoid the temptation to use AI as a substitute for a human supervisor. Organisations that want to integrate AI into clinical practices should assess the relational consequences of using AI—its effects on relational processes such as performativity, self-censorship, and the ability to engage in negative capability—before using it in a clinical setting.
3. Preserving the Analytic Space. Before integrating AI into clinical settings, organisations should assess whether the benefits of AI, such as documentation, pattern identification, and adherence monitoring, are sufficient to justify the relational costs of using AI.  This should include increased performance expectations, increased self-censorship due to fear of being evaluated, and reduced capacity to engage in negative capability. Where AI threatens to weaken authenticity, containment, or trust, organisations should judge the relational costs to outweigh the operational benefits. When organisations choose to integrate AI into clinical settings, they should develop specific procedures to protect the relationship between the analyst and supervisee from algorithmic intrusion, so that supervisees can safely explore uncertainty, errors, and countertransference without fear of being evaluated by an algorithm.
4. Ethically Integrating Accessibility. Intelligent technology can make mental health treatment available to populations that lack access due to geographic location, timing of crises relative to traditional working hours, or economic constraints. These positive aspects of intelligent technology should not be dismissed. At the same time, there should be no confusion about the distinction between providing support and psychoeducation via intelligent technology, and the embodied presence, the capacity for authentic mentalising, and the ability to engage unconsciously with others that human therapists provide. Organisations should frame AI systems as tools to facilitate contact with human-based care, when possible, rather than as alternatives or replacements.
5. Responsibility, Accountability and Governance. As AI systems are increasingly being used in the delivery of mental health treatments, regulatory structures will require a major overhaul to protect against the new forms of risk that exist with AI systems—specifically, epistemological exploitation, surveillance, deskilling, and the degradation of professional identity (World Health Organisation, 2024; Shumate et al., 2025). Regulatory structures should be collaboratively developed by all stakeholders in the mental health community—including clinicians, patients, and psychoanalysts—and should not be imposed by technologists or regulatory authorities. Current regulations are fragmented and inadequate, based primarily on general AI regulations that do not account for the unique relational and unconscious dynamics that psychoanalysis addresses. Future regulations should distinguish between uses of AI that augment auxiliary functions (transcription, scheduling, psychoeducation), and those that aim to simulate therapeutic relationships—subjecting the latter to much stricter standards.

Several important gaps remain in future research. First, there is a need for longitudinal studies assessing whether the level of epistemic trust generated through interactions with AI generalises to human relationships, particularly for individuals with histories of attachment trauma or other relational vulnerabilities (Békés & Aafjes-van Doorn, 2026). Second, there is a need for a systematic study of the extent to which surveillance anxiety impacts supervisory relationships: at what level of algorithmic monitoring do we see clinically significant distortion of the supervisory relationship? Third, there is a need for an empirical study of how AI affects the development of clinical judgment among trainees: does early exposure to AI-generated feedback increase learning, or does it create dependency in trainees that undermines their own clinical decision-making? Fourth, there is a need for empirical study of how the benefits and risks of AI are distributed culturally and socio-economically: do AI’s benefits and risks fall equally on all members of society, or do they exacerbate pre-existing inequalities in the name of accessibility? Lastly, there is a need for an empirical study of how psychoanalytic training programs could prepare clinicians to function in a world in which AI is ubiquitous, and yet preserve the fundamental psychoanalytic abilities, i.e. embodied attunement, the ability to tolerate ambiguity, the ability to enter reverie and experience countertransference, on which psychoanalytic practice is founded.

Comparability and limits

The comparability problem between psychoanalytic and general mental health research is itself theoretically productive. Psychoanalytic treatment does not primarily aim at symptom reduction, and standard outcome measures such as symptom scales or alliance ratings do not fully capture changes in unconscious conflict, symbolisation, or subject-to-subject relating. The relative absence of psychoanalysis from the AI outcome literature may therefore reflect not only a gap in the evidence base, but also a mismatch between what that literature is designed to measure and what psychoanalytic treatment is intended to change.

Conclusion: Preserving the Human in an Algorithmic Age

While the conclusions may appear obvious, they have wide-ranging impacts. A self-critical caveat is warranted. The near-unanimous position across this literature is that current AI cannot replace the human analyst and may itself warrant psychoanalytic scrutiny. The forcefulness of the “AI as adjunct only” consensus, including in the present review, may partly reflect a collective defensive response to narcissistic injury and professional existential threat rather than purely dispassionate conceptual analysis. What must not be lost in this acknowledgment, however, are the concrete conditions that give the resistance its clinical legitimacy: the preservation of the analytic frame as the structured space from which the unconscious can be encountered anew by each subject; the analyst’s cultivated not-knowing as the condition for the analysand’s discovery; and the irreducibility of the two-person encounter as the site at which psychic reality is co-constructed rather than delivered. These are not defensive idealisations but technical necessities, and they constitute substantive grounds for scepticism that stand independently of narcissistic injury. Acknowledging this does not invalidate the arguments made. There is a logical incompatibility between algorithmic and unconscious processes that stands independently. However, intellectual honesty requires noting that the analysts writing about AI are not outside the dynamics they describe (Knafo, 2024).

The notion that current algorithms can emulate the unconscious, that AI has a “real” mind, or that simulated experience can replace lived subjective experience rests on a faulty assumption. Psychoanalysis provides a theoretical basis for understanding the distinctions between these experiences and for protecting what remains unexchangeable.

Innovation driven by technology will continue to permeate the field of mental health. What is at issue is not if AI enters analytic spaces, but under what terms, with what safeguards, and with what clarity regarding its limitations. An psychoanalytic perspective requires analytic discernment: the ability to distinguish legitimate uses of AI while protecting the embodied, relational, and unconscious processes that are the only means through which actual psychological change can occur in psychoanalytic practice as presently understood.

The future of analytic practice need not be characterised by displacement or dilution. By implementing theoretical clarity, ethical awareness, and institutional commitment, we can position AI as a constrained technological tool that expands access to analytic treatment while maintaining the uniquely human dialectical process that underlies analytic work. In an era in which machines can simulate aspects of unconscious process without possessing subjectivity or embodied experience, the analyst’s embodied presence is the indispensable condition for crossing from fantasy to authentic desire. It is our responsibility to ensure that the unique human capacities (to love and hate simultaneously, to hold contradiction without resolution, to be truly moved by another’s suffering) remain intact in their full, living forms.