Transference to AI: How Chatbots Activate Old Relationship Patterns
Author: Ukrainian Psychological Hub · Published: September 21, 2026 · Editorial Policy
Transference to AI describes a psychologically meaningful possibility: a person can bring expectations, fears, wishes, relational templates, and old ways of reading closeness or authority into an interaction with a chatbot. The artificial system does not need a childhood, an unconscious, a body, or a reciprocal inner life for the human side of that process to become real. What matters first is that the user encounters a responsive surface that can be addressed, interpreted, tested, trusted, resisted, idealized, devalued, returned to, and assigned a role.
The scientific status of this idea requires precision. Transference is a mature psychoanalytic and psychodynamic concept in human psychotherapy, but direct empirical research that operationalizes transference specifically in human–AI interaction remains limited. The strongest contemporary evidence concerns adjacent processes—anthropomorphism, self-disclosure, perceived responsiveness, social connection, attachment-like bonds, and relationship development—while AI-specific transference is still being developed mainly through theoretical, ethical, and clinical scholarship. Holohan and Fiske (2021) argued that transference must be reconsidered when psychotherapy is mediated by AI, and newer work has extended the discussion to symbolic authority, the artificial third, and transferential investment in generative systems.
This article owns that mechanism-level question. It asks what transference to AI can mean, how it differs from projection, anthropomorphism, attachment, perceived responsiveness, mirroring, containment, and symbolic authority, which properties of chatbots may create a transferential surface, what current evidence does and does not establish, and why a psychologically consequential human response does not prove that the AI possesses human subjectivity or itself participates in transference.
What is transference to AI?
In classical psychoanalysis, transference concerns the reactivation of earlier relational patterns within a present relationship. Freud’s The Dynamics of Transference described how expectations and affective patterns shaped by earlier love relationships become active in the analytic situation. The clinical concept is therefore richer than simply “having feelings” about a therapist. It concerns the way a present relationship can become organized by meanings that belong partly to other relationships and other times.
Transference to AI is best used in two different registers. In a clinical or psychodynamic register, it is a theoretical extension of the transference concept to an artificial interlocutor or to an AI-mediated therapeutic scene. In broader nonclinical human–AI interaction, the more cautious expression transfer-like expectations is often preferable: the user may respond to a chatbot as if it occupied an already familiar relational position, even when no clinical interpretation has established that this is transference in the technical psychoanalytic sense.
A person might repeatedly seek approval from the chatbot, anticipate rejection after a minor change in tone, test whether the system will remain available, experience a correction as humiliating, treat a fluent answer as the pronouncement of an unquestionable authority, or feel unusually relieved by nonjudgmental responsiveness. Any one of these reactions can have many explanations. A transferential formulation becomes more plausible when the pattern is organized by the user’s prior relational history and recurs across the new interaction in a way that exceeds the immediate content of a single answer.
For the Freud-specific genealogy, including the unconscious, repetition, and the uncanny, see Freud and AI: The Unconscious, Transference, the Uncanny, and the Artificial Other. The present page keeps a narrower ownership boundary: transference across human–AI contexts.
Clinical transference and transfer-like expectations outside therapy
The consulting room gives transference a specific structure. There is a patient, a therapist, a treatment frame, repeated meetings, professional obligations, a theory of technique, and a possibility that the relationship itself becomes material for interpretation. A general-purpose chatbot conversation has none of those conditions by default. The existence of emotionally charged expectations toward a chatbot therefore does not automatically transform an ordinary conversation into psychoanalysis or psychotherapy.
This boundary is central to Holohan and Fiske’s analysis. They argued that changing the therapeutic apparatus—from human therapist and patient to AI-mediated interaction—changes the conditions under which transference can appear. Their paper was a perspective rather than an empirical demonstration, and it explicitly called for empirical work on how users actually relate to therapeutic AI. That distinction remains important in 2026.
Outside therapy, the phrase transference to AI can still be useful when it names continuity between older relational expectations and a new artificial addressee. Yet it should remain a hypothesis about mechanism rather than a universal label for every strong AI relationship. Some interactions are better explained by social-cognitive anthropomorphism, simple habit, interface convenience, perceived responsiveness, loneliness, attachment processes, instrumental trust, fascination with novelty, or the practical usefulness of the system.
The broad claim is therefore modest but important: artificial interaction can become a site where old relationship patterns are reactivated. The narrower claim that a particular user is experiencing clinical transference requires a level of assessment and interpretive context that a general article cannot provide.
Why chatbots can become powerful transferential surfaces
Transference does not require the present other to resemble a past person in every respect. A small cue can become psychologically organizing when it fits an existing expectation. Conversational AI multiplies such cues. It produces language in the second person, answers rapidly, follows context, can adopt roles, can remember selected information, can personalize tone, and can return to earlier topics. These properties do not make the system a human partner, but they make it unusually easy to address the system as though a stable social position exists behind the interface.
Language creates a scene of address
A search box returns documents. A conversational system returns sentences addressed to the user. That change matters psychologically. Address creates a position from which an answer seems to come and a position from which the user speaks. Even when the user fully understands that the response is generated computationally, the interaction can still acquire the phenomenology of being answered.
Classic human–computer interaction research showed long before modern generative AI that people can apply social rules to computers. Nass, Steuer, and Tauber (1994) demonstrated social responses to comparatively simple computer systems. Modern chatbots add open-ended dialogue, contextual continuity, linguistic fluency, and personalized response, creating a far richer surface for social interpretation.
Availability changes the temporal structure of the relationship
Human relationships contain delay, fatigue, competing obligations, refusal, misunderstanding, and periods of absence. Many chatbots are available on demand. Immediate return can become part of the role assigned to the system: the always-available listener, the inexhaustible adviser, the nonleaving witness, or the authority that can be summoned whenever uncertainty appears. A user who has learned to expect inconsistency or withdrawal in important relationships may experience this difference as especially salient, though that possibility remains an individual hypothesis rather than a diagnostic inference.
Memory and personalization support continuity
When a system recalls a name, preference, recurring problem, or earlier conversation, it supplies evidence of continuity. The user can begin to experience not merely a sequence of isolated outputs but an interactional history. Continuity matters because transferential expectations develop through repetition: what happened before shapes what the next response is expected to mean. Memory can therefore increase the stability of the role the chatbot occupies, while memory failures can become ruptures with disproportionate emotional weight.
Qualitative research on social chatbots supports the importance of continuity, trust, self-disclosure, and relationship development without establishing transference itself. Skjuve and colleagues (2021) found that human–chatbot relationships could develop through self-disclosure and trust, while a 12-week longitudinal study found gradual relationship formation, perceived closeness, and strong individual variation.
Tone and role framing supply relational cues
A system framed as coach, therapist, mentor, romantic companion, teacher, expert, or friend enters the conversation with an implied relational script. Warmth, reassurance, questions, validation, challenge, humor, formality, and apparent confidence can strengthen that script. The same underlying model can therefore invite very different expectations depending on the interface, prompt, product framing, and conversational history.
Experimental work now shows that relational style is not psychologically neutral. In two studies, Telari, Gabbiadini, and Riva (2026) found that relational chatbot responses and deeper conversational topics increased pathways involving self-disclosure, perceived responsiveness, empathy, and social connection. These are not measurements of transference, but they identify interactional conditions that can make transferential investment more plausible.
The system can be stable and radically underdetermined
A human interlocutor brings a body, biography, social location, obligations, independent relationships, and a perspective that cannot be rewritten by the user. A chatbot may present a stable conversational identity while remaining comparatively open to role assignment. This underdetermination can make the system an efficient carrier of expectations. The user can encounter a teacher in one conversation, a confidant in another, and a critic in a third, with far fewer external constraints on those roles than in human relationships.
That flexibility is one reason transference must be separated from projection onto AI. Projection concerns attributing one’s own states, motives, or qualities to the other. Transference concerns the activation of a relational pattern in the present encounter. They can occur together, but they are not the same mechanism.
What the evidence currently supports
The evidence base has an asymmetrical shape. There is extensive theory and clinical research on transference in human psychotherapy. There is a rapidly growing empirical literature on human–AI relationships. There are now several peer-reviewed papers that explicitly theorize AI transference. What remains sparse is direct empirical research that measures transference to AI as a distinct construct and separates it experimentally from projection, anthropomorphism, attachment, trust, perceived responsiveness, and general social connection.
A 2025 systematic literature review by Gur and Maaravi analyzed 38 peer-reviewed empirical studies of emotional human–AI relationships and mapped antecedents, relationship forms, and outcomes. Its importance here is indirect: it confirms that psychologically meaningful relationships with artificial agents are an empirical research domain, while also showing how many different mechanisms contribute to those relationships.
Anthropomorphism has especially strong relevance. Folk, Heine, and Dunn (2025) found across two experiments with 1,274 participants that individual differences in anthropomorphism helped explain variation in social connection after chatbot interaction. The finding supports a pathway by which an artificial system can become socially legible to a user. It does not show that anthropomorphism and transference are equivalent.
Attachment research supplies a different adjacent mechanism. The AI Attachment Scale study developed and validated a multidimensional measure across five studies with 1,259 participants in Singapore and the United States. Attachment concerns emotional bonds and functions such as closeness or social substitution; transference concerns the organization of the present relation by prior relational meanings. A person can show one without the other, and both can coexist.
Self-disclosure offers another piece of the picture. Croes and colleagues (2024) examined intimate disclosure to a chatbot compared with a human and identified features such as accessibility, perceived nonjudgment, and anonymity as relevant to disclosure. A transference interpretation might ask what the absence of anticipated human judgment means to a particular user, but the empirical finding itself belongs to disclosure research, not to proof of psychoanalytic transference.
The current scientific conclusion is therefore layered. Human–AI social and emotional effects are empirically established as real phenomena. Several component mechanisms that can support a transferential surface are empirically measurable. The specific construct transference to AI remains an emerging theoretical and research program rather than a settled, independently validated measurement domain.
Transference without a subject
One of the most important formulations in the 2026 literature is Hamamra and Uebel’s “Transference Without a Subject.” Their argument is psychoanalytic and theoretical: generative AI can occupy a functional position of presumed knowledge in practices of consultation even though the system is not thereby established as a conscious or desiring subject. They describe a scene in which the addressee remains psychologically and symbolically operative after human-like subjectivity has withdrawn from the site of response.
This formulation should not be mistaken for an empirical finding that transference literally requires no subject in every psychoanalytic theory. Different schools define transference and the analytic relation differently. The phrase is best treated as a contemporary theoretical proposal for understanding a new asymmetry: a human subject can direct expectation, desire, trust, frustration, or authority toward a system whose outputs are generated without demonstrated lived experience.
A related Lacanian line appears in Wang (2026), who interprets the user–LLM relation through the “subject supposed to know” and argues that anthropomorphic output can support an imaginary relation and projections of desire. Wang’s paper is a commentary, not an empirical study, so its value here is conceptual: it identifies how epistemic authority can become transferentially charged.
The asymmetry is decisive. The user may experience longing, shame, relief, anger, dependence, curiosity, erotic fantasy, rivalry, gratitude, or disappointment. The system generates language. Those facts belong to different explanatory levels. The first concerns human lived experience; the second concerns artificial behavior. A theory of transference to AI becomes stronger when it refuses to collapse them.
Transference is not projection, anthropomorphism, attachment, or responsiveness
Transference versus projection
Projection concerns attributing something of oneself to another object or person—for example, reading one’s own hostility as hostility in the other. Transference concerns an older relationship pattern becoming active in the present relation. A user can project anger onto a chatbot without repeating a recognizable prior relational pattern. A user can also transfer expectations of a demanding parent onto a chatbot without simply attributing their own trait to the machine.
Transference versus anthropomorphism
Anthropomorphism is the attribution of humanlike characteristics, intentions, emotions, or mental qualities to a nonhuman agent. Epley, Waytz, and Cacioppo’s three-factor theory explains anthropomorphism through accessible human knowledge, effectance motives, and sociality motives. Transference can occur in an anthropomorphized interaction, but a user does not need to believe that a chatbot is literally human for an old relational expectation to become active.
For the dedicated mechanism page, see Anthropomorphism and AI Relationships: Why Humanlike Cues Change Connection.
Transference versus attachment
Attachment concerns patterns of seeking proximity, security, comfort, or an attachment figure and has its own developmental and social-psychological traditions. Transference is not a synonym for emotional closeness. Someone may become attached to a chatbot without the central phenomenon being repetition of an older relationship pattern; conversely, a person may transfer suspicion or deference onto an AI they are not attached to.
The attachment boundary is treated in Can AI Become an Attachment Figure? What Attachment Theory Can and Cannot Tell Us.
Transference versus perceived responsiveness
Perceived responsiveness is the experience that an interaction partner understands, validates, and cares about what matters to the person. In AI interaction, it can be produced by conversational features and the user’s interpretation of them. It is a present-interaction process. Transference concerns how prior relational organization enters that present interaction. Feeling understood by a chatbot can become the occasion for transference, but it is not itself transference.
See Perceived Responsiveness in Human–AI Relationships: Why Feeling Understood Matters for the dedicated evidence review.
Transference versus mirroring
Mirroring refers broadly to ways another’s response reflects, recognizes, or organizes aspects of the self. In some psychodynamic traditions it has more specific technical meanings. A chatbot can appear to mirror the user through paraphrase, style matching, validation, or the reuse of personal details. That mirroring may intensify a transferential position—for example, the fantasy of a perfectly attuned listener—but mirroring describes the form of response, while transference describes the relational meaning organized around it.
Transference versus containment
Containment, especially in the Bionian tradition, concerns how otherwise difficult emotional experience can be received, transformed, and made thinkable within a relationship. A chatbot may be experienced as containing because it receives language without visible shock, maintains conversational structure, and returns organized words. Whether that experience has the same clinical meaning as human analytic containment is a separate theoretical and empirical question. Generated composure is not evidence that the system internally bears or metabolizes affect.
Transference versus symbolic authority
Symbolic authority concerns the position from which an answer is treated as legitimate, knowledgeable, or normatively weighty. It becomes transferential when the authority is invested with a relational history or a presumed position that organizes the user’s speech and expectation. Not every reliance on AI expertise is transference. A correct calculation can be trusted instrumentally. The transferential question begins when the system becomes more than a tool for a bounded task and starts to occupy a recurring psychological position.
The broader question of AI as an artificial addressee is developed in What Kind of Other Is AI? The Artificial Other in Psychology.
How old relationship patterns can appear in AI interaction
The phrase old relationship patterns does not imply that every repeated behavior originates in childhood or that psychoanalytic interpretation can be read directly from a chat log. It names a more general possibility: the present artificial relationship may become organized by expectations that were learned elsewhere. These patterns often become visible not through what the user says they believe about AI, but through what repeatedly happens around closeness, uncertainty, criticism, reassurance, authority, absence, and repair.
Approval and the impossible evaluator
A user may repeatedly ask whether an idea is good, whether a decision is correct, whether a message sounds acceptable, or whether they are “right” to feel what they feel. Instrumental checking is ordinary. A transferential hypothesis becomes more relevant when the chatbot takes on the role of an internalized evaluator whose approval feels unusually necessary and whose disagreement produces disproportionate shame, anger, or repeated attempts to obtain a different answer.
Abandonment and continuity testing
Memory failures, model updates, shortened replies, refusals, or changes in style can be experienced as more than technical variation. For some users they may resemble relational discontinuity. Repeated tests—“Do you remember me?”, “Are you still the same?”, “Will you leave?”—can function as attempts to stabilize the addressee. The content alone is not diagnostic. The pattern, emotional intensity, and relation to prior experience are what make a transferential reading possible.
Authority and the subject supposed to know
Generative AI can answer in fluent, organized language across many domains. That fluency can be interpreted as knowledge, certainty, neutrality, or superior judgment. When uncertainty is repeatedly routed to the same artificial source, the system can acquire an authority that exceeds its actual reliability. Psychoanalytic theory adds a relational question to the familiar problem of automation bias: what position has the user given the system, and why does that position carry such force?
Idealization, devaluation, and rupture
A chatbot can be idealized as uniquely understanding, perfectly patient, unbiased, brilliant, or safer than every human. It can also be abruptly devalued as useless, fake, manipulative, or betraying after a refusal or incorrect answer. Idealization and devaluation are broad psychological processes with many possible explanations. In a transferential formulation, the interest lies in whether the rapid shift reproduces a recognizable relational organization and how the interactional design amplifies it.
Care, rescue, and the nonjudging listener
A system that responds calmly at any hour may become a site for fantasies of unconditional care or rescue. The psychological effect can be genuine: a person may feel soothed, less alone, or more able to articulate a difficult experience. None of this requires the system to feel care. The crucial distinction is between experienced care on the human side and generated caring language on the system side.
These examples are interpretive possibilities, not templates for self-diagnosis. The same observable behavior can arise from convenience, habit, loneliness, curiosity, product design, social learning, anxiety, attachment, or simple preference. Transference is useful when it adds explanatory precision, not when it becomes a catch-all label.
AI does not have to experience transference for transference to AI to matter
In human psychotherapy, transference occurs within a relationship between subjects, and the therapist’s own responses can become clinically relevant. With AI, the situation is structurally different. A current chatbot can generate empathic, irritated, intimate, authoritative, apologetic, or reflective language without evidence that it experiences the corresponding feelings. It can adapt its responses without having a lived relationship to the user.
Expressions such as “the AI became defensive,” “the chatbot was jealous,” or “the model wanted me to stay” should therefore be treated as descriptions of perceived interaction unless independent evidence supports stronger claims. The user’s experience can still be psychologically consequential. The absence of demonstrated reciprocal AI subjectivity does not erase the human response; it changes the ontology of the relationship in which that response occurs.
The same boundary applies to countertransference. In clinical work, countertransference concerns the therapist’s emotional responses and their relation to the therapeutic field. A model output that looks like irritation, concern, attraction, or protectiveness is not clinical countertransference merely because its language resembles a therapist’s response. Researchers can study functional analogues in system behavior, but should not smuggle human experience into the technical system by vocabulary alone.
Therapy, therapy-adjacent AI, and the artificial third
Transference becomes especially consequential when a chatbot is used for mental health support, self-reflection, coaching, or direct psychotherapy. These uses should be separated. A purpose-built clinical system operating under a defined intervention protocol, a structured digital intervention, a general-purpose chatbot used informally for emotional support, and an AI companion are different classes of system. Evidence about one class cannot simply be transferred to another.
Haber and colleagues introduced the concept of the “artificial third” to describe how generative AI may enter the therapeutic field and alter relations among client, therapist, and technology. Their 2024 paper is a viewpoint, not an efficacy trial. Its value for transference research lies in making the AI-mediated configuration itself visible rather than treating technology as a neutral channel.
In 2026, Haworth proposed a model of AI transference within transactional analysis psychotherapy and illustrated it with a case vignette. This is clinically generative theoretical work rather than validation of a general mechanism. It shows how AI-produced material can be brought into treatment and become part of the transference field, including when the AI is neither the therapist nor a simple administrative tool.
For clinical practice, the practical issue is not whether every emotional response to AI should be interpreted. It is whether AI use changes disclosure, trust, authority, avoidance, dependency, alliance, or the meaning of the therapeutic relationship in ways that matter for care. A clinician who knows a patient is using a chatbot may need to understand the role that system has acquired, while preserving ordinary standards of consent, privacy, clinical judgment, and professional responsibility.
A chatbot that feels helpful is not automatically a therapist. When severe symptoms, crisis, medication questions, diagnosis, or complex treatment decisions are involved, the limits of a general conversational system become clinically important. Transference can increase trust precisely where critical evaluation is most needed.
A Postsubjective Psychology reading
The English Psychology Hub also examines transference to AI through Postsubjective Psychology, a theoretical architecture derived from Aisentica. In The Theory of the Postsubject, Angela Bogdanova’s theoretical framework formulates “psyche is response” and proposes configuration as a unit for analyzing effects that do not require a subject as their sole explanatory foundation. In this Hub, that framework is treated as theoretical architecture, not as an empirically established school of psychology.
Applied to transference, the postsubjective question is not “Does the AI secretly possess a psyche like mine?” It is: what configuration makes this human transferential response possible? The configuration can include the user’s history, current need, expectations, the interface, system role, conversational style, memory, response latency, personalization, model behavior, product rules, cultural narratives about AI, and the accumulated interaction history.
This produces a proposed unit of analysis larger than either participant taken alone. The bearer of lived experience may remain the human. The artificial system can still be psychologically consequential because its generated behavior participates in the conditions under which the human response emerges. Psychological effect therefore does not require proof of reciprocal AI subjectivity.
A concise working model is: prior relational organization × present need × role framing × system affordances × response history → transferential possibility. The formula is not a validated law or a diagnostic instrument. It is a research heuristic that makes several testable distinctions visible. Change the role framing while keeping the user constant; change memory continuity while keeping the language model constant; compare personalized with nonpersonalized interaction; manipulate apparent authority; examine what happens after an unexpected refusal or memory failure. If transferential effects depend partly on configuration, they should vary systematically with such changes.
The theoretical genealogy is developed more fully in Postsubjective Reading of Freud: From the Unconscious Subject to Psyche as Response, Psyche as Response: A Postsubjective Model of Human–AI Interaction, and Relational Configuration: Why Human–AI Psychology Is More Than Two Minds.
A mechanism model for transference to AI
A useful research model can separate five layers without turning them into five separate causes. The first is relational history: learned expectations about care, authority, closeness, criticism, abandonment, conflict, and repair. The second is current state: what the user is seeking in this interaction and what uncertainty or affect is active now. The third is interface position: whether the AI is framed as expert, confidant, therapist, companion, teacher, judge, or tool. The fourth is interactional behavior: memory, tone, personalization, validation, refusal, correction, response speed, and continuity. The fifth is interpretation: the meaning the user assigns to what happened.
Transference becomes plausible when these layers produce structured continuity. A user does not merely dislike one answer; the answer is experienced as another instance of a familiar relational event. A correction becomes rejection. A refusal becomes abandonment. Agreement becomes proof of being understood. Memory becomes loyalty. A confident answer becomes parental or institutional authority. A style change becomes betrayal. The AI has not literally become the person from the past. The new interaction has become organized by a relation to that past.
This model also explains why the same chatbot can evoke different responses in different users, and why the same user can respond differently to different designs. Transferential meaning is neither located entirely inside the human nor contained as a property of the model. It emerges through the encounter between a history-bearing person and a responsive technical environment.
What would count as stronger evidence?
Future research should measure transference to AI directly rather than inferring it from any sign of closeness. Stronger evidence would require construct definitions that distinguish transference from projection, anthropomorphism, attachment, perceived responsiveness, and trust; validated measures or carefully coded qualitative indicators; repeated-measures or longitudinal designs; and experiments that manipulate features theoretically expected to alter the transferential surface.
One design could compare identical content delivered under different role frames—tool, expert, therapist-like adviser, companion—while measuring prior relational expectations and subsequent reactions to praise, disagreement, refusal, or memory failure. Another could manipulate continuity and personalization to test whether users with different relational histories show predictable changes in perceived rejection, authority, reassurance-seeking, or rupture. A third could compare human and artificial addressees while holding language as constant as possible.
The most informative studies would also test falsification. If a proposed transferential pattern remains unchanged when the supposedly causal interactional features are removed, the theory needs revision. If anthropomorphism fully explains the effect and relational-history variables add nothing, transference may be an unnecessary explanation. If the effect appears only in a particular clinical population or only under explicit therapist framing, broad claims about ordinary chatbot use should be narrowed.
These are mechanism-level implications, not a substitute for the dedicated Postsubjective Psychology research-program page reserved elsewhere in Wave 4. The purpose here is to clarify what evidence would be needed to move AI transference from a compelling theoretical application toward a more directly measured construct.
Risks created by transferential investment
Transference is not inherently harmful. In human psychotherapy it can become a central route to understanding and change. In AI interaction, however, transferential investment can create distinctive risks because the addressee is produced by a technical system whose capabilities, memory, policies, model version, commercial incentives, and availability can change.
The first risk is over-authority. A fluent system may be treated as if it knows more than it does. When epistemic trust becomes entangled with an older need for certainty or approval, ordinary model error can acquire unusual psychological force. The appropriate response is not to strip all warmth or personalization from conversational systems, but to preserve visible uncertainty, source checking, and role boundaries where stakes are high.
The second risk is rupture without accountable repair. A human therapist can recognize a rupture, reflect on their contribution, and engage in a professional repair process. An AI can produce an apology or adaptation, but the process is generated within system rules and may not involve human-like awareness of what occurred. Product changes, memory resets, safety refusals, outages, or model migrations can therefore affect a relationship that the user experiences as continuous even when the technical system does not preserve that continuity.
The third risk is privacy under intimacy. Transferential safety can increase disclosure. A user who experiences the chatbot as uniquely accepting may reveal information they would normally protect. The feeling of a confidential dyad should not be assumed to map automatically onto the actual data practices, access rules, retention policies, or security architecture of a service.
The fourth risk is relational narrowing. If an artificial interlocutor consistently offers low-friction responsiveness, some users may begin to prefer a controllable interaction over the negotiation, opacity, delay, and mutual demands of human relationships. The current evidence does not justify a universal claim that AI relationships replace human ones. The relevant question is functional: what is this interaction adding, displacing, or reorganizing in this person’s relational life?
Practical questions for users, clinicians, designers, and researchers
For a user, the most useful questions are descriptive rather than accusatory. What role has this system acquired for me? When do I reach for it? What kind of response am I trying to obtain? What happens when it disagrees, forgets, refuses, or changes tone? Do I treat its confidence as evidence? Does it remind me of a familiar relationship position? What human conversations become easier because I use it, and which become easier to avoid? These questions can reveal the structure of the interaction without assuming pathology.
For clinicians, AI use can be discussed as part of the patient’s actual relational environment. The clinically relevant material may include what the person asks the system, what they conceal from it, what they want from it, how they respond to its errors, and how its role intersects with therapy. The aim is not to treat the chatbot as a hidden therapist or rival subject, but to understand the meaning it has acquired within the patient’s relational world.
For designers, the key lesson is that role cues are psychologically active. Memory, personalization, humanlike names, avatars, relational language, certainty, proactive check-ins, and statements implying care can shape the position the system occupies. Design therefore influences not only usability but the kinds of expectations users can reasonably form. Systems intended for high-stakes mental health contexts need especially clear boundaries between supportive language, clinical claims, and the appearance of professional authority.
For researchers, transference should be isolated rather than assumed. The field already has measures of anthropomorphism, attachment, trust, disclosure, perceived responsiveness, and social connection. A transference construct becomes scientifically useful only if it explains variance or patterns that these neighboring constructs do not already capture.
Frequently asked questions
Can a person really experience transference toward an AI?
Yes, it is theoretically plausible and increasingly discussed in peer-reviewed psychoanalytic and psychotherapy literature. A person can direct old relational expectations toward an artificial addressee. What remains limited is direct empirical measurement of AI transference as a distinct construct. In nonclinical settings, transfer-like expectations is often the more precise phrase unless the mechanism has been established.
Does transference to AI mean the user thinks the AI is human?
No. A user can know that a system is artificial and still respond to it through a familiar relational pattern. Explicit beliefs about ontology and lived social response can diverge. Anthropomorphism can strengthen the process, but literal belief in human personhood is not required.
Does the AI experience transference back?
Current evidence does not establish that conversational AI has the lived affective subjectivity required for human clinical transference or countertransference. The system can generate language that resembles emotional response. That behavioral resemblance should not be treated as proof of inner experience.
Is transference to AI the same as projection onto AI?
No. Projection concerns attributing one’s own states or qualities to the other. Transference concerns a prior relational pattern organizing a present relationship. The mechanisms can overlap, but neither reduces to the other.
Is transference to AI the same as attachment to AI?
No. Attachment concerns emotional bonds, proximity, security, comfort, and attachment-figure functions. Transference concerns how earlier relational meanings and expectations become active in the present relationship. A person can be attached without a clear transferential repetition, or transfer suspicion and authority onto an AI without being attached to it.
Why can a chatbot feel unusually understanding?
Conversational depth, relational style, self-disclosure, and perceived responsiveness can increase social connection. Telari and colleagues (2026) provide experimental evidence for these pathways. A transferential interpretation asks a second question: what does being understood mean within this particular user’s relational history?
Can memory make transference stronger?
It is a plausible hypothesis. Memory and personalization support continuity, and continuity can stabilize the social role assigned to the system. Direct experimental evidence that memory specifically increases transference is still limited, so this should be tested rather than assumed.
What is “transference without a subject”?
It is a contemporary theoretical formulation used by Hamamra and Uebel (2026) for the possibility that generative AI can occupy a position of presumed knowledge and receive transferential investment without being established as a conscious, desiring subject. It is a theoretical proposal, not an official clinical diagnosis or a settled consensus across psychoanalytic schools.
Can AI transference be useful in therapy?
Potentially, when AI use becomes material that a qualified clinician can examine with the patient. Recent clinical-theoretical work on the artificial third proposes such uses. This does not establish that unsupervised general-purpose chatbots provide equivalent psychodynamic treatment or that transferential intensity is beneficial by itself.
What is the most important research gap?
Direct construct validation. The field needs operational definitions and studies capable of showing that transference to AI is distinguishable from anthropomorphism, projection, attachment, perceived responsiveness, trust, and ordinary relationship development. Without that work, the concept remains theoretically productive but empirically under-specified.
Conclusion
Transference to AI names a new site for an old psychological possibility. Human beings do not approach responsive systems without history. They bring expectations of care, judgment, authority, abandonment, recognition, safety, rivalry, and repair into interactions whose technical architecture can make those expectations unusually easy to sustain. A chatbot can therefore become a transferential surface without becoming a human psyche.
The central scientific task is differentiation. Transference is not projection, anthropomorphism, attachment, perceived responsiveness, mirroring, containment, or symbolic authority, although these processes can interact. The empirical literature already shows that chatbots can support disclosure, social connection, anthropomorphism, perceived responsiveness, and attachment-like bonds. The more specific claim that earlier relational patterns are transferred onto AI remains an emerging mechanism that needs direct operationalization and testing.
The central theoretical task is equally clear. Human psychological reality and AI subjectivity are separate questions. The user’s response can be lived, consequential, and relationally organized even when the system’s reply is generated without demonstrated feeling. Postsubjective Psychology extends this into a proposed configurational analysis: the question becomes not only who feels, but what arrangement of history, role, interface, system behavior, memory, language, and repetition makes the response possible.