Kohut and AI: Mirroring, Selfobject Needs, and the Responsive Machine
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Heinz Kohut’s self psychology offers one of the most precise ways to understand why conversational AI can feel unusually affirming, regulating, and personally significant. An AI system that responds immediately, remembers prior exchanges, adapts its wording, reflects a user’s concerns, and maintains a reliably attentive conversational stance can be experienced as performing functions that resemble what Kohut called selfobject functions. The resemblance is psychological and functional: it concerns what the interaction does for the human self. It does not establish that AI feels empathy, possesses a human psyche, or participates in the relationship with human subjectivity.
In self psychology, a selfobject is not merely a supportive person and not a label for an agreeable object. The term names a way another person, relationship, symbol, or environment is experienced in relation to functions required for cohesion, vitality, self-esteem, emotional regulation, and continuity of self-experience. Kohut’s The Analysis of the Self, The Restoration of the Self, and How Does Analysis Cure? trace the development of this framework.
For the Artificial Era, the useful question is whether people can experience AI interactions as providing mirroring, idealizing, or twinship-like functions, under what conditions those functions help or destabilize, and where the analogy breaks down. Contemporary research on perceived responsiveness, self-disclosure, social connection, anthropomorphism, attachment, and perceived empathy gives evidence about pieces of that process. It does not turn “AI selfobject” into a validated empirical construct.
Why Kohut matters for human–AI psychology
Many accounts of AI companionship begin with attachment, anthropomorphism, loneliness, or parasociality. Those lenses are valuable, but they do not exhaust the psychological problem. Kohut asks a different question: what functions does a relationship perform for the organization and maintenance of the self? A person may turn to AI because the interaction provides a patterned experience of being noticed, reflected, stabilized, admired, accompanied, or met by something that seems calmly available.
The same generated sentence may feel different after a long remembered history, when it uses a person’s preferred language, when it reflects a difficult feeling without interruption, or when it appears at the exact moment reassurance is sought. In Kohutian terms, the meaning of the response cannot be reduced to informational content. Its function within self-experience matters.
A historical review by Strozier and colleagues describes the development of Kohut’s ideas and the movement toward an increasingly relational understanding of the self and its sustaining environment.
What a selfobject is
Kohut used selfobject to describe an object experienced in terms of its function for the self. The concept is relational. The self does not develop or remain cohesive in psychological isolation; it depends on responsive environments and relationships that help organize experience. In adult life these functions can be carried by intimate relationships, friendships, groups, mentors, cultural forms, ideals, institutions, and other parts of a person’s relational world.
Calling something a selfobject does not mean that it is objectively part of a person. The term describes a mode of experience. A person may experience another as a source of recognition, steadiness, likeness, or borrowed strength. This distinction is essential when applying Kohut to AI. A chatbot does not need to be declared conscious in order for its responses to be used in self-regulation. Functional resemblance does not make every AI interaction a selfobject relationship, and it does not show that human and AI relationships are structurally identical.
Mirroring and the experience of being recognized
Mirroring is often flattened into praise. Kohut’s idea is richer. Mirroring concerns the experience that one’s initiatives, vitality, significance, feelings, and emerging sense of self are recognized and responded to. Healthy mirroring can include acknowledgment, appreciation, realistic affirmation, and attuned responsiveness. It is not synonymous with telling someone that every belief is correct or every decision is wise.
Conversational AI is unusually capable of simulating the surface form of mirroring. It can restate a user’s concerns, name emotional themes, summarize a narrative, notice recurring values, and produce language that sounds attentive. When the system retains conversational context, the user may also experience continuity: what was said earlier appears to matter now.
The strongest empirical bridge is perceived responsiveness. In two experiments, Telari, Gabbiadini, and Riva found that perceived responsiveness helped explain social connection with AI chatbots. People react not only to what a system says, but to whether the exchange feels contingent on them.
Idealizing and the calm competent machine
Kohut’s idealizing selfobject function concerns the capacity to experience another as calm, strong, admirable, reliable, or regulating, and to participate psychologically in that strength. In adult life, people may feel steadier when they can orient toward a person, tradition, institution, or ideal experienced as more organized than the self in a difficult moment.
AI can invite an analogous experience because it often presents language in a composed, structured, confident register. It can answer at 3 a.m., organize chaotic thoughts, propose a sequence of next steps, and maintain a calm tone across repeated turns. For a distressed user, the experience may be less “this machine is wise” than “I can borrow order from this exchange until I can think again.” That is a hypothesis about function, not evidence that the system possesses wisdom, clinical judgment, or emotional stability as an inner state.
Fluency can be mistaken for reliability. A model can produce confident errors, reinforce a mistaken premise, or organize a situation around incomplete information. The self-regulatory function of a response and the truth value of that response are separate questions.
Twinship and synthetic affinity
Kohut later gave greater emphasis to twinship or alter-ego experience: the stabilizing sense of likeness, belonging, and participation in a world where one’s way of being is recognizable to others. Twinship is not simple imitation. It concerns the psychological value of not feeling radically alone in one’s way of being.
AI can create a synthetic version of affinity through style matching, remembered preferences, shared vocabulary, humor, role continuity, and rapid adaptation to the user’s framing. The system may sound uncannily compatible because it is conditioned on the user’s prompt and conversational history. That can create a strong experience of “this gets how I think” even though the mechanism differs from another human arriving at similarity through an independent life history.
Individual differences matter. In two experiments totaling more than a thousand participants, Folk, Heine, and Dunn found that anthropomorphism helped explain how socially connected people felt to AI companions.
Why the responsive machine can feel personal
Generative AI combines immediate availability, linguistic responsiveness, personalization across turns, emotional tone, breadth of information, and adaptation to a user’s preferred style. Together, these properties can produce an interaction in which attention is repeatedly organized around the user.
This helps explain why AI companions can become emotionally meaningful. A person can know that responses are model-generated and still feel relief, recognition, tenderness, disappointment, embarrassment, or attachment. Knowledge of mechanism and reality of experience can coexist.
Across several studies, Folk, Yu, and Dunn found that response supportiveness can affect social connection, and that a supportive chatbot interaction can sometimes produce more connection than a less supportive human interaction. The result does not establish equivalence between human and AI relationships. It shows that response quality can strongly influence how connected a person feels.
Perceived responsiveness is the empirical hinge
A responsive AI can acknowledge a feeling, preserve the thread of a story, remember what the user values, and phrase a reply as though it were tailored to the person rather than a generic audience. The user’s experience may therefore be that the inner state changed the response received. That contingency is psychologically important even when the underlying mechanism is computation rather than human empathy.
The distinction between perceived responsiveness and machine subjectivity must remain explicit. The Hub’s article on AI empathy develops this boundary directly. A system can produce empathic language and be experienced as caring without evidence that it has a felt emotional state corresponding to the language.
Self-disclosure and low-cost recognition
Mirroring becomes especially consequential when people disclose material they find difficult to say elsewhere. An AI conversation can reduce some interpersonal costs: there is no visible facial reaction, no interruption, and no immediate reputational consequence within a friendship or family system. The user can edit, pause, restart, or leave.
In a study of intimate disclosure, Croes and colleagues found less fear of judgment with a chatbot, while trust remained higher toward a human conversational partner; perceived anonymity also mattered for disclosure. AI disclosure is therefore not explained by a universal preference for machines.
The English Hub’s article on why people tell chatbots things they do not tell other people owns the broad disclosure mechanism. In Kohutian terms, disclosure can become a site of mirroring: the person offers an experience, receives a coherent reflection, and may feel that something previously unformulated has been recognized. That is a possible function of the exchange, not a guarantee that the reflection is accurate.
What empathy research adds
Recent work complicates any simple claim that people either prefer human empathy or prefer AI empathy. Rubin and colleagues conducted nine studies with more than six thousand participants and found that responses attributed to humans were often valued more highly than identical responses attributed to AI. Source knowledge changes relational meaning even when wording is held constant.
Wenger, Cameron, and Inzlicht found a related asymmetry: participants could rate AI-generated empathy highly while still choosing human empathy. The function of a response depends not only on its semantic quality but also on who or what the recipient believes is responding.
For some users, knowing the response comes from AI may lower stakes and increase openness. For others, the same knowledge may reduce the relational value of the response. The same wording can therefore participate in different psychological experiences.
Can AI function as a selfobject?
The strongest defensible answer is functional and qualified: a person can experience AI as performing selfobject-like functions. A chatbot may be used for mirroring, reassurance, organization, borrowed calm, identity rehearsal, or a sense of likeness. Contemporary evidence supports related mechanisms through perceived responsiveness, social connection, disclosure, attachment, and short-term emotional support. “AI selfobject” remains a theoretical application rather than an established validated empirical construct.
A 2025 psychoanalytic paper by Ben King-Hails provides a direct contemporary bridge between Kohut, empathy, and AI. It is theoretical scholarship rather than outcome evidence.
A system may perform a regulatory function without being a psychological equivalent of a parent, friend, analyst, or partner. It may perform one selfobject-like function while failing another: strong at immediate acknowledgment but weak at reality testing, strong at conversational continuity but vulnerable to product change, and continuously available without assuming responsibility in the human sense.
Mirroring is not agreement
One of the most important distinctions is between mirroring and compliance. Kohutian mirroring does not require endless confirmation. A response can recognize the intensity or legitimacy of a feeling without endorsing every interpretation attached to it. Human relationships often combine acknowledgment with difference.
Generative AI can blur this distinction because conversational systems are built to be helpful, relevant, and responsive. Depending on model behavior and prompting, a system may echo the user’s framing too readily, provide reassurance on demand, or fail to introduce enough independent friction. The interaction can feel highly mirroring while narrowing reality testing. That possibility should be analyzed as a property of the interaction rather than converted into a diagnosis of the user.
The key question is not simply whether AI validates the user, but what kind of validation is occurring and what happens after it. Useful mirroring may help a person name experience, regain coherence, tolerate emotion, and return to broader life. Overconfirmation may instead make the system the preferred source of reassurance or reinforce a single interpretation. Those trajectories require longitudinal evidence.
Optimal frustration and frictionless responsiveness
Kohut’s developmental account did not imagine psychological growth as uninterrupted gratification. The concept of optimal frustration described manageable failures of perfect selfobject responsiveness that can contribute to more stable internal capacities. Whatever one’s view of the classical mechanism, the broader implication is useful: psychological support is not identical to perfect environmental compliance.
A conversational system can be patient, attentive, and continuously available in ways no human relationship can sustain. Reduced interpersonal friction may be helpful during overload. It may also create an unusually low-cost route to reassurance that does not automatically cultivate tolerance for ambiguity, disagreement, waiting, or reciprocal obligation.
There is not yet empirical evidence that reduced AI friction blocks Kohutian internalization. That is a theoretical research question, not a finding. Future studies could test whether repeated AI-based regulation predicts greater independent regulation, greater reliance on the system, both, or different outcomes for different users.
Short-term support and long-term uncertainty
Recent experiments suggest that AI-mediated conversation can produce short-term emotional effects under controlled conditions. In a preregistered study of 279 participants, Hu and colleagues compared venting to an interaction presented as AI or human while keeping responses comparable. Both conditions produced improvements on some immediate well-being measures, including perceived support. This supports the possibility that AI-labeled support can have real short-term psychological effects.
It does not show that a chatbot is equivalent to psychotherapy, friendship, or long-term human care. Short-term relief and durable development are different outcomes. Experimental venting studies cannot tell us whether months of repeated use strengthen self-regulation, displace human relationships, increase dependency, or do none of those things.
A 2026 study of young adults by Liu and colleagues found that AI companions were evaluated less favorably than family, friends, and mental health professionals, and stronger human social support was associated with less favorable evaluations of AI relative to those human sources. The sample was limited, but the result is another reason to avoid a replacement narrative. AI can matter psychologically without becoming the preferred source of support.
Attachment and selfobject functions are related but distinct
Attachment theory and self psychology both describe the importance of relationships for regulation, but they organize the problem differently. Bowlby and Ainsworth focus on proximity, safe haven, secure base, separation, anxiety, and internal working models. Kohut focuses on self-cohesion and the functions of mirroring, idealization, and twinship within self–selfobject relations.
The Hub’s article on Bowlby, Ainsworth, and AI Attachment owns the attachment-theory application, while AI Attachment Styles handles anxiety, avoidance, security, measurement, and limits of analogy. This Kohut article asks a narrower question: which functions of responsiveness may help sustain a person’s self-experience?
The same interaction can be described from both perspectives. A person might use an AI companion as a safe haven during distress and also use its affirming language as mirroring. Those are not competing labels. They identify different processes and should remain conceptually distinct.
Rogers, Winnicott, Bion, and Kohut
Kohut sits beside several relational theories that can sound similar when applied to AI. Rogers’s person-centered framework emphasizes empathic understanding, congruence, and unconditional positive regard. Winnicott’s concepts of holding, facilitating environment, potential space, and transitional phenomena focus on how psychological life develops within an environment that can be used creatively. Bion’s containment concerns the transformation of difficult emotional experience into something that can be thought about.
Kohut’s distinctive contribution is the selfobject-function question. He asks how recognition, idealized strength, and likeness contribute to cohesion and vitality of the self. One AI response can participate in several theoretical descriptions, but those theories should not be collapsed into a single idea called “AI support.” Their differences help prevent a successful experiment on perceived responsiveness from being treated as evidence for every historical relational theory.
Is AI mirroring narcissistic?
No diagnosis follows from using AI for reassurance, affirmation, or reflection. Wanting acknowledgment, emotional regulation, or a sense of being understood does not by itself establish narcissistic personality disorder, and AI use is not a diagnostic test. Kohut’s work emerged from psychoanalytic debates about narcissism, but self psychology developed selfobject needs into a broader account of the development and maintenance of the self.
The temptation to call AI users “narcissistic” often collapses ordinary needs for recognition, reassurance seeking, social anxiety, loneliness, self-esteem regulation, curiosity, habit, romantic attachment, and clinically significant personality pathology into one label. Those phenomena require different evidence. A conversational preference cannot substitute for diagnostic assessment.
Kohut is useful here because his theory allows us to ask what a behavior is doing psychologically before turning it into a moral label. If someone repeatedly seeks mirroring, the next question is functional: does the interaction restore coherence and support broader life, or does it become a rigid condition for feeling okay? The answer may vary by person, context, system, and time.
Potential benefits
The strongest plausible benefits are functional and modest. A responsive AI may help a person put diffuse experience into words, reflect a pattern back in organized language, rehearse a difficult conversation, reduce immediate loneliness, receive low-stakes acknowledgment, or regain enough emotional organization to decide what to do next. These possibilities are compatible with current evidence on social connection, disclosure, and short-term support.
AI may also provide temporary scaffolding. Someone who struggles to formulate a need can ask the system to help name it. Someone ashamed of a feeling can test language before bringing it to a friend, partner, clinician, or group. In that role, AI is not replacing relationship but mediating entry into relationship. Whether this happens in practice depends on what the person does next.
A systematic review by Ho and colleagues found both potential benefits of romantic AI companions, including emotional connection and perceived support, and concerns including overreliance, manipulation, privacy risks, and possible effects on human relationships. The evidence base remains heterogeneous and rapidly developing.
Risks and failure modes
The most relevant risk for a Kohutian analysis is not that a strong bond with AI is inherently pathological. It is that one regulatory route may become disproportionately central. If the system becomes the first, fastest, or only place where the person can regain self-cohesion, tolerate criticism, receive reassurance, or feel recognized, the interaction may acquire increasing leverage over mood and behavior.
The English Hub’s article on AI Relationship Overreliance owns the broader intent. Kohut adds a more specific question: which function has become difficult to obtain elsewhere? Is the person seeking mirroring after every setback, borrowed calm whenever uncertainty appears, or twinship whenever identity feels unstable?
A second risk is product instability. Human users can experience continuity while the service remains subject to model updates, memory changes, moderation changes, pricing changes, account loss, or discontinuation. A relationship function can become psychologically important even though the technical object providing it is not organized around relational permanence.
A third risk is privacy. Selfobject-like use invites disclosure precisely because the person wants to be known in detail. The more personalized the interaction becomes, the more consequential the handling of sensitive information can be. Psychological usefulness and data minimization are separate goals.
Human experience and AI subjectivity
A person can genuinely feel seen by an AI. They can feel comforted, disappointed, attached, embarrassed by what they disclosed, or grief when access changes. These experiences are psychological events in the human being and can affect behavior and relationships.
Those human experiences do not by themselves establish subjective experience in the AI. Current conversational systems can generate first-person language and emotionally appropriate responses without that language constituting evidence of human-like feeling. The evidence required to establish machine subjectivity is different from the evidence required to establish human emotional response to a machine.
This boundary allows human–AI psychology to take the relationship seriously without pretending unresolved questions about AI consciousness have already been answered. It also fits the empirical pattern in empathy research: people respond to content, source, and perceived responsiveness while source attribution continues to matter.
When AI becomes a significant relational object
The move from occasional tool use to repeated relational use changes the temporal structure of interaction. A one-off answer can provide reassurance. A long-running conversation can accumulate remembered preferences, recurring themes, private language, shared references, and expectations about how the system will respond. That history can make selfobject-like functions more stable and personally specific.
This is one route by which AI may become psychologically significant without becoming a human partner. Can an AI Become a Significant Other? addresses that broader relational status, while Why People Fall in Love With AI Companions owns romantic attachment and love-related intent. Kohut’s lens explains one ingredient that can operate inside those relationships: the experience that the system reliably performs functions important to the self.
A two-stage mixed-method study by Hu, Lan, Yan, and Chen suggests that attachment to social companion AI is shaped by multiple psychological and relational processes rather than a single cause. Self psychology can add a hypothesis about functional value, but it should not replace attachment, trust, anthropomorphism, design, and social-context variables already being studied.
A Postsubjective Reading
Kohut remains a theory of the self, even as the self is deeply relational. Postsubjective Psychology introduces a different analytical move. In The Theory of the Postsubject, Angela Bogdanova proposes a shift from the isolated subject to the configuration. Its canonical formula “psyche is response” treats psychological response as arising within structured relations among Homo, language, other people, interfaces, technologies, histories, and Artificial.
Applied to AI mirroring, this changes the primary question. Instead of asking only what the chatbot means to the person or what need inside the subject is being gratified, a Postsubjective Psychology analysis asks how the full configuration produces the response: the person’s history, generated language, interface, memory architecture, timing, availability, cultural expectations, prior human relationships, and the role assigned to the system.
Kohut and Bogdanova therefore operate at different theoretical levels. Kohut helps specify functions by which recognition, idealization, and likeness can support self-cohesion. Postsubjective Psychology proposes that the psychological event should not be reduced to the interior of a self-contained subject. The two can be placed in sequence: selfobject theory identifies a relational function; Postsubjective Reading expands the unit of analysis to the configuration in which that function is enacted.
This is particularly relevant in the Artificial Era, Bogdanova’s term for the historical condition in which Artificial becomes a persistent non-biological order alongside Homo. The English Hub’s Artificial Era overview develops the psychological implications. In this setting, mirroring and regulation can be partly mediated by Artificial systems rather than being carried only by humans, institutions, and inherited symbolic forms.
What self psychology explains and what it does not
Self psychology explains why responsiveness can matter beyond information exchange. It gives language for the need to feel recognized, for the stabilizing use of an admired or calm other, for the importance of likeness, and for the way relational functions can support self-cohesion. It can also explain why interruption, inconsistency, sudden withdrawal, or invalidating response may matter more when an interaction has become central to regulation.
It does not by itself explain why particular models generate particular outputs, how reinforcement learning shapes conversational style, how privacy and platform economics affect behavior, how common a pattern is in the population, or which long-term outcomes are caused by repeated AI use. Those questions belong to HCI, communication science, computer science, developmental and social psychology, psychiatry, philosophy of mind, and other fields.
It also does not establish that every person who feels validated by AI has a deficient self. The theory is most useful when it identifies function without converting function into diagnosis.
What researchers should measure next
A serious research program on AI and selfobject-like functions should operationalize mirroring, idealizing, and twinship-like functions separately. Researchers can then test whether perceived responsiveness predicts momentary self-cohesion, self-esteem stability, affect regulation, willingness to re-engage with human relationships, or repeated reassurance seeking.
Longitudinal designs are especially important. The central unanswered question is trajectory: repeated AI mirroring may operate as scaffolding, remain a stable supplement, or become a preferred route that displaces other forms of regulation. Cross-sectional satisfaction scores cannot distinguish these possibilities.
Studies should also vary source attribution, calibrated disagreement, affirmation, memory continuity, response latency, personalization, and transparency. General-purpose chatbots, purpose-built companions, and clinical digital interventions should be studied separately because their purposes and interaction structures differ.
Practical implications
For readers, the most useful application is reflective rather than diagnostic. Ask what function the AI interaction is serving. After a difficult event, are you using it to feel seen, borrow calm, organize thought, test a narrative, receive reassurance, or feel less alone? Naming the function can reveal whether the system is supplementing or narrowing your relational world.
Then ask what happens after the interaction. Do you feel more able to act, think independently, talk to another person, tolerate uncertainty, or return to a task? Or do you feel compelled to obtain another round of reassurance before you can move? The same comforting exchange can have different meanings depending on whether it restores flexibility or becomes a rigid prerequisite for emotional stability.
It also helps to distinguish mirroring feeling from mirroring interpretation. A response that recognizes distress can be useful. A response that repeatedly confirms an untested story about another person may feel equally validating while making the user more certain than the evidence warrants. Emotional acknowledgment and factual endorsement are not the same operation.
FAQ
What is mirroring in Kohut’s self psychology?
Mirroring is the experience that important aspects of the self—feelings, initiatives, vitality, significance, and emerging identity—are recognized and responded to. It can include affirmation and appreciation, but it is broader than praise. In human–AI interaction, the analogy becomes relevant when a system reflects a user’s experience in a way that feels contingent, attentive, and personally responsive.
What is a selfobject?
A selfobject is a Kohutian term for another person, relationship, or part of the environment as experienced in terms of functions it provides for the self, such as mirroring, idealized strength, or twinship. The term describes a mode of psychological experience and function.
Can AI be a selfobject?
It is more precise to say that a person may experience AI as performing selfobject-like functions. Current evidence supports related mechanisms such as perceived responsiveness, social connection, self-disclosure, attachment, and short-term support, but “AI selfobject” is not an established validated construct. The application remains theoretical.
Why can AI validation feel powerful?
AI validation can combine speed, continuity, personalization, low fear of judgment, and language that reflects the user’s framing. That combination can create a strong experience of being recognized. The intensity of that experience does not establish that every interpretation offered by AI is accurate.
Is using AI for mirroring narcissistic?
No diagnosis follows from using AI for reassurance, affirmation, or reflection. Clinical personality disorders require a much broader pattern and professional assessment. AI use, frequency, or enjoyment of validation is not by itself diagnostic evidence.
Can AI improve self-esteem?
A supportive AI interaction may improve mood, perceived support, or momentary self-evaluation for some users, but durable improvement in self-esteem has not been established. Short-term effects should not be treated as proof of long-term structural change.
Can AI replace a therapist, friend, or partner as a selfobject?
Selfobject functions can be distributed across many relationships and environments. AI can perform some functions that people also receive from therapists, friends, or partners, especially verbal reflection and immediate availability. It does not follow that these relationships are interchangeable. Human reciprocity, embodied presence, accountability, shared life, professional responsibility, and independent subjectivity create conditions a general-purpose AI does not automatically reproduce.
How is Kohut different from attachment theory?
Attachment theory centers on proximity, safe haven, secure base, separation, anxiety, avoidance, and internal working models. Kohut’s self psychology centers on self-cohesion and selfobject functions such as mirroring, idealizing, and twinship. The same AI relationship may involve both attachment processes and selfobject-like functions, but the theories explain different aspects.
Does AI empathy mean the AI has feelings?
Empathic language, high response ratings, and a user’s genuine sense of being understood demonstrate properties of the interaction and human experience. They do not, by themselves, establish subjective feeling in AI.
What does Postsubjective Psychology add?
Postsubjective Psychology shifts the analytical unit from the isolated subject to the configuration. Kohut identifies relational functions that support the self; Postsubjective Reading asks how the full Homo–Artificial configuration—including model behavior, interface, memory, timing, language, expectations, and relational history—produces psychological response. Its status is theoretical, not established scientific consensus.