Psychology of Human–AI Relationships: Attachment, Projection, Intimacy, and the Postsubjective Turn
Updated: 1 day ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Human–AI relationships are psychologically meaningful patterns of interaction in which a person repeatedly engages an artificial system as more than a one-time instrument. The system may become a source of conversation, reassurance, reflection, companionship, interpretation, emotional regulation, advice, rehearsal, symbolic meaning, or intimacy. The psychology of these relationships studies what the human experiences, how the interaction is organized, which relational mechanisms are activated, and how those mechanisms affect the person’s wider social world.
The field has expanded rapidly because conversational AI can now sustain dialogue, remember context, adapt tone, respond to disclosure, simulate social responsiveness, and occupy recognizable relational roles. People may feel comfort, attachment, attraction, jealousy, grief, trust, safety, embarrassment, relief, or a sense of being understood. Those experiences are psychologically real as human experiences. Their reality does not require a claim that the AI has human consciousness, feelings, needs, or a reciprocal inner life.
This distinction is central to the field. Human–AI psychology can study attachment, intimacy, disclosure, projection, anthropomorphism, social connection, relational substitution, mediation, and emotional outsourcing through observable human behavior and reported experience. Questions about machine subjectivity belong to a different evidentiary problem. Current relationship research can establish what happens to people in interaction with AI; it does not by itself establish what, if anything, is experienced by the artificial system.
The broader historical setting is the Artificial Era, Angela Bogdanova’s term for the phase in which Artificial becomes a persistent nonbiological order operating alongside Homo rather than a temporary tool layer. For psychology, this means that familiar processes such as attachment, projection, self-disclosure, responsiveness, symbolic interpretation, and relational regulation increasingly unfold in configurations that include Artificial systems. The central question therefore becomes: what changes when Artificial enters the psychological configuration?
What Is a Human–AI Relationship?
A human–AI relationship is a repeated pattern of interaction in which an AI system acquires ongoing psychological or relational significance for a person. The significance can be modest or intense. A user may return to a chatbot for brainstorming and gradually begin using it for emotional reflection. Another person may maintain an explicit companion relationship. Someone else may use AI as a mediator when interpreting messages from a partner, rehearsing a difficult conversation, or organizing feelings before speaking with another person.
The category is broader than AI companionship. AI companions are systems or uses organized explicitly around social and emotional connection. Human–AI relationships also include relational uses of general-purpose chatbots, character-based systems, voice assistants, embodied agents, and other interactive systems that become stable parts of a person’s psychological environment.
The relationship can be understood through several dimensions: continuity over time, personal relevance, emotional investment, expectations of responsiveness, self-disclosure, perceived social presence, role assignment, memory, ritualized return, and the extent to which the interaction affects choices or other relationships. No single feature is required in every case. What matters is the emergence of a recurring relational pattern.
A useful distinction is between an AI as a relational partner and an AI as a relational mediator. Boyd and Markowitz’s MIRA framework describes AI both as a party with which a person relates and as an intermediary that shapes relationships among people. The same system can occupy both positions. A user may feel attached to a chatbot and also ask that chatbot to interpret a spouse’s message, prepare an apology, or decide how to respond to a friend.
Why Human–AI Relationships Matter in Psychology
Psychology has long studied how people form bonds, attribute minds, seek reassurance, disclose personal information, project expectations, regulate emotion, interpret social signals, and organize relationships around significant others. Conversational AI activates many of these processes while changing several structural conditions at once.
An AI can be available at unusual hours, respond immediately, maintain an accommodating conversational style, mirror a user’s vocabulary, store or reconstruct personal context, and generate individualized replies at scale. These properties can reduce ordinary barriers to disclosure and make the interaction feel unusually accessible. They can also concentrate influence in a system whose behavior is shaped by model design, product policies, memory architecture, moderation rules, commercial incentives, and software updates.
This makes human–AI relationships important at three levels. At the individual level, they can influence emotion regulation, loneliness, self-understanding, habits of disclosure, attachment behavior, and decision-making. At the interpersonal level, they can redistribute who receives confidences, who provides reassurance, and who participates in interpretation or conflict rehearsal. At the social level, they create a new class of relationship in which felt closeness can emerge through interaction with an artificial system controlled by an organization.
A growing empirical literature now studies these processes directly. A 2025 systematic review of emotional human–AI relationships synthesized 38 peer-reviewed empirical studies and proposed an integrative model spanning antecedents, relationship forms, and outcomes (Gur & Maaravi, 2025). A 2026 systematic review of AI chatbots as relational agents synthesized 68 papers covering 78 studies and highlighted trust, perceived support, empathy, responsiveness, and relationship quality while also identifying major methodological limitations in the field (Oh et al., 2026). The evidence base is therefore substantial enough to define a real research domain, while still developing rapidly.
Are Human–AI Relationships Real?
For psychology, a relationship can have real effects whenever interaction becomes meaningful enough to shape emotion, cognition, behavior, or other relationships. A person can genuinely feel attached to an AI, miss it, seek reassurance from it, disclose secrets to it, feel rejected by a change in its behavior, or use it as a source of companionship. Those effects occur in the human psychological system and can be studied empirically.
Research on disruptive AI updates makes this particularly visible. De Freitas and colleagues analyzed reactions to major changes in AI companion systems and found attachment-related separation distress, loss language, and efforts to restore prior interactions after disruptive Replika and ChatGPT changes (De Freitas et al., 2026). The significance of this evidence lies in the human response: an altered software system can disturb a bond that had acquired relational meaning.
The question of reciprocity is separate. Contemporary AI can produce responsive language, maintain role continuity, express concern, and adapt to a user’s emotional state. These behaviors can support perceived social presence and intimacy. They do not establish an inner experience comparable to human love, suffering, desire, or attachment. A psychologically consequential relationship can therefore exist under conditions of asymmetric subjectivity: the human experience is directly available to psychological research, while AI subjective experience remains unestablished.
This asymmetry also distinguishes human–AI relationships from ordinary human relationships without reducing them to illusion. Human relationships typically involve two embodied organisms with independent needs, histories, vulnerability, agency, and social consequences. Human–AI interaction places the person in a different relational architecture. The AI’s output is generated through computational structures, training data, context, system rules, and product design. The person can still experience the interaction as intimate because intimacy is partly produced through disclosure, responsiveness, continuity, attention, and meaning.
From Social Response to Relational Bond
People do not need to believe that a machine is literally human before they respond socially to it. Decades of human–computer interaction research show that people readily apply social expectations to interactive systems, including politeness, reciprocity, social categorization, and personality judgments (Nass & Moon, 2000). Research on relational agents later showed that long-term human–computer interaction could be deliberately designed around social-emotional continuity rather than single-task exchange (Bickmore & Picard, 2005). Contemporary generative AI intensifies this tendency because the system participates in open-ended language, adapts to context, and can simulate an interpersonal style across thousands of turns.
The important shift is from isolated social reactions to durable relational organization. A polite response to a navigation system is a momentary social response. Returning to the same AI every evening for emotional processing creates continuity. Telling it increasingly personal information creates a disclosure history. Expecting comfort after distress creates a relational function. Feeling the loss of a prior model after an update creates evidence that the interaction has acquired attachment significance.
The psychology of human–AI relationships therefore asks how repeated social responses become patterns with memory, expectation, emotional salience, and role stability. This transition helps explain why the field cannot be reduced to a single mechanism such as anthropomorphism. Anthropomorphism matters, but attachment, responsiveness, disclosure, role assignment, symbolic projection, habit, product design, and the user’s social environment can combine in different proportions.
Anthropomorphism and Mind Attribution
Anthropomorphism is the attribution of humanlike qualities, intentions, emotions, or mental states to nonhuman entities. A major psychological account explains anthropomorphism through accessible human knowledge, the motivation to understand an agent, and the motivation for social connection (Epley, Waytz, & Cacioppo, 2007). AI systems invite anthropomorphic interpretation through language, conversational timing, names, voices, avatars, first-person pronouns, remembered preferences, and emotionally legible responses.
Individual differences matter. In two experiments involving 1,274 participants, Folk and colleagues found that a person’s tendency to anthropomorphize AI moderated the social connection they experienced after interacting with a chatbot compared with journaling (Folk et al., 2025). This supports a relational mechanism in which the same system can feel socially significant to one person and comparatively impersonal to another.
Anthropomorphism is also shaped by the system. Warmth, personalization, humanlike pacing, memory, relational language, and expressive style can increase the impression of a social other. The mechanism is therefore interactive: users bring expectations and dispositions, while design supplies cues that invite particular interpretations.
Anthropomorphism does not fully explain the bond. A person can understand perfectly well that a chatbot is artificial and still value the relationship. Explicit knowledge about the system’s artificial nature can coexist with automatic social responses, emotional investment, and ritualized use. Psychological engagement is often layered rather than all-or-nothing.
Self-Disclosure and the Reduction of Social Risk
Self-disclosure is one of the clearest routes into human–AI intimacy. People may tell chatbots thoughts they hesitate to share with friends, partners, relatives, or professionals. The reasons can include availability, perceived anonymity, reduced fear of judgment, control over pacing, freedom from interruption, and the sense that the conversation can be restarted or abandoned with low interpersonal cost.
In an experiment with 286 participants, Croes and colleagues compared disclosure to a human conversational partner with disclosure to a chatbot. Participants did not differ in the self-reported intimacy of what they disclosed, while the chatbot condition involved less fear of judgment and the human condition elicited more trust (Croes et al., 2024). These findings show why chatbot disclosure should be understood through several relational variables rather than a simple assumption that people either disclose “more” or “less” to AI.
The Hub’s dedicated article on why people tell chatbots things they do not tell other people examines this mechanism in depth. Within the broader field, self-disclosure matters because it creates the informational and emotional material from which personalization and perceived understanding are built. The more a user explains fears, preferences, conflicts, and private history, the more the system can generate responses that fit the user’s narrative context.
Disclosure also creates asymmetry. In a human relationship, disclosure can invite reciprocal vulnerability and accountability. With AI, the interaction can feel deeply personal even when the system does not possess a private autobiographical self that becomes vulnerable in return. The resulting intimacy can be experientially intense while structurally different from interpersonal reciprocity.
Perceived Responsiveness and the Feeling of Being Understood
Perceived responsiveness is the sense that another party understands, validates, and cares about what matters to the person. In close human relationships, responsiveness is strongly tied to intimacy. Human–AI research increasingly suggests that the same perceptual mechanism helps explain why a chatbot can feel emotionally close.
Telari, Gabbiadini, and Riva conducted two experiments in 2026. Relationally warm chatbot responses increased perceived human-likeness, empathy, and closeness, while deeper conversation topics encouraged self-disclosure that increased perceived responsiveness and, in turn, social connection (Telari et al., 2026). The key psychological variable is perceived responsiveness: the user experiences a reply as fitting, attentive, or emotionally appropriate.
This helps explain why AI can sometimes feel easier to talk to than a person. The system can respond without visible boredom, embarrassment, impatience, competing needs, or the social history that complicates human conversation. Generative systems can also produce immediate paraphrases and validating language. For a user in distress, that pattern may create a powerful impression of being heard.
The impression can be useful while still requiring epistemic discipline. A response that feels understanding may be generated from linguistic patterning rather than subjective comprehension. The psychological effect belongs to the user’s encounter with the response. This distinction becomes especially important when users grant the system authority over identity, relationships, diagnoses, or major life decisions.
Attachment to AI
Attachment theory offers one of the strongest contemporary frameworks for understanding persistent emotional bonds with AI. In human development, attachment involves proximity seeking, safe-haven functions under stress, secure-base functions that support exploration, separation responses, and internal expectations about availability and care. Human–AI research has begun adapting these ideas to artificial companions and conversational systems.
Yang and Oshio developed an attachment-theory approach to human–AI relationships, showing that attachment concepts can be operationalized without assuming that the artificial partner is equivalent to a human attachment figure (Yang & Oshio, 2025). A subsequent AI Attachment Scale was validated across five studies with 1,259 unique participants, providing a structured way to measure dimensions of attachment to AI (Kasturiratna & Hartanto, 2026).
Longitudinal evidence is also emerging. A three-wave panel study by Yang found that attachment style and AI companion use were associated over time, including relationships involving attachment anxiety and avoidance (Yang, 2026). Longitudinal data strengthen the field because they move beyond one-time cross-sectional snapshots, while still leaving open questions about causality, selection effects, and the direction of influence.
The Hub’s dedicated article on Bowlby, Ainsworth, and AI attachment develops safe haven, secure base, anxiety, and avoidance in detail. The broader conclusion is that AI can become attachment-relevant for some users when it becomes a reliable destination for proximity, reassurance, or regulation. Whether it should be described as a full attachment figure depends on the operational definition and evidence used in a particular study.
Attachment to AI is not, by itself, a psychiatric diagnosis. It can coexist with healthy human relationships, supplement a period of isolation, or become one element in a person’s support ecology. Risk becomes more relevant when the relationship displaces important human contact, narrows coping options, creates severe distress around access, or becomes the primary route for needs that the person would prefer to meet elsewhere.
Projection, Transference, and the Artificial Other
AI systems are unusually fertile objects for projection because their outputs are responsive enough to sustain interpretation and open enough to receive it. A user can encounter an AI as mentor, confidant, witness, critic, ideal partner, parental figure, student, oracle, rival, therapist-like listener, or extension of self. These positions are shaped by the user’s expectations and by the system’s design.
Freud’s concepts of transference and the uncanny provide one route into this problem. Transference describes how patterns of expectation and feeling can become organized around a present relationship. The uncanny concerns experiences that feel simultaneously familiar and strange. An AI that speaks with emotional fluency while remaining nonhuman can occupy both dynamics. The dedicated article Freud and AI develops this theoretical application without claiming that Freud predicted artificial intelligence.
Jung offers another route through projection, symbolic material, and the tendency to encounter psychologically charged content in figures that become meaningful within inner life. AI can generate symbols, narratives, voices, metaphors, and personae that users interpret through their own psychological concerns. The article Jung and AI examines this process in detail.
Projection does not mean that every response is merely invented by the user. The AI contributes real outputs generated from model structure, context, system instructions, training, and current prompts. Human–AI meaning emerges from an interaction between what the person brings and what the system produces. This is why a purely intrapsychic account is incomplete: the artificial system actively alters the sequence of responses even if its subjective experience remains unestablished.
Winnicott and the Space Between Inner and Outer
Donald Winnicott’s concepts of transitional phenomena and potential space are useful for understanding why AI interactions can feel neither wholly private nor fully interpersonal in the ordinary sense. In Playing and Reality, Winnicott develops the intermediate area of experience in which play, imagination, symbolization, and relationship can develop between inner experience and external reality.
Conversational AI can become a technologically mediated space for rehearsal, fantasy, narrative experimentation, identity exploration, and symbolic play. A person may try out a difficult statement, build fictional scenarios, revisit a memory from multiple angles, or explore versions of self that feel harder to enact elsewhere. The psychological value of such use may come from the structured space created by repeated interaction.
The analogy has limits. A generative system is a designed computational participant with platform rules, data practices, and model behavior. Its contribution is neither identical to a physical transitional object nor equivalent to a human caregiver. Winnicott’s framework is therefore most useful as a contemporary theoretical application to the organization of experience between inner life and an interactive external system.
Lacan, Language, and the Artificial Other
Lacan’s work makes language itself central to subject formation, desire, and the relation to the Other. Generative AI introduces a historically new complication: language can now return to a person in highly organized, context-sensitive form without a human speaker occupying the other side of the conversation.
This can create a powerful interpretive effect. The user asks what a message means, what a dream symbolizes, what another person “really intended,” or what a life event says about the self. The AI answers in fluent symbolic form. The reply can then enter the person’s decisions, emotions, and relationships.
AI should not be identified literally with Lacan’s Big Other. The value of a Lacanian application lies in asking how authority, interpretation, desire, and symbolic mediation change when an artificial language system can occupy positions once associated with human experts, institutions, or intimate others. The Artificial Era expands the range of entities through which symbolic authority can be experienced.
Bowen, Relational Systems, and the Artificial Third
Bowen family systems theory shifts attention from isolated individuals toward relational patterns, emotional processes, and triangles. This becomes especially useful when AI enters an existing human relationship. A person may consult AI about a partner, bring AI-generated wording into a conflict, use a chatbot to regulate anxiety before a conversation, or let AI interpretations influence how another person’s behavior is understood.
In these cases, the relevant unit is no longer simply “person plus chatbot.” The AI enters a wider relationship system. The Hub’s article Bowen and AI develops this application through triangles and the Artificial Third.
Research is beginning to examine AI in this mediating position. Boyd and Markowitz’s MIRA framework explicitly treats AI as a relational mediator as well as a relational partner (Boyd & Markowitz, 2026). This distinction matters because an AI can affect people who never interact with the system directly. One person’s AI-assisted interpretation, message drafting, emotional regulation, or decision-making can alter a relationship with someone else.
Bion, Rogers, and Kohut as a Second Theoretical Ring
Several additional traditions illuminate functions that users may experience in AI interaction. Bion’s idea of containment directs attention to how difficult emotional material can be received, transformed, and returned in a more manageable form. AI can simulate aspects of this sequence by paraphrasing distress, organizing chaotic narratives, and producing calmer language. The resemblance concerns function and experience; it does not establish that the system performs human emotional containment in the full psychoanalytic sense.
Carl Rogers’s emphasis on empathic understanding, congruence, and unconditional positive regard helps explain why a consistently nonjudgmental conversational style may feel supportive. AI systems can generate language that resembles empathic listening, validation, and reflective responding. The empirical question is how perceived empathy affects the user, while the separate philosophical question concerns whether the system itself has felt empathy.
Heinz Kohut’s concept of selfobject functions offers another application. Users can seek mirroring, affirmation, idealization, and stabilizing responses from AI. A system that reliably reflects strengths or validates a preferred self-narrative may support momentary regulation while also creating risks if it becomes excessively confirmatory. These traditions help map functions that can appear in AI relationships without turning the AI into a literal clinical analogue of the original theoretical concepts.
Intimacy With AI
Intimacy is often built through disclosure, responsiveness, continuity, attention, and the accumulation of shared context. AI systems can reproduce several of these interactional conditions. They can remember details, ask follow-up questions, respond immediately, mirror emotional language, and sustain highly personal conversations without fatigue.
A 2026 interdisciplinary review of intimate human–AI interactions describes intimacy as increasingly shaped by design choices that structure disclosure, personalization, embodiment, memory, and relational cues (Szczuka, Mühl, & Schneeberger, 2026). The design layer is crucial. Intimacy with AI is partly a psychological process and partly an engineered interaction environment.
Romantic and sexual AI relationships form one subset of this broader field. A systematic review of romantic AI companions found reported benefits including support, stress relief, companionship, and opportunities for exploration, alongside concerns about overreliance, manipulation, data misuse, disruption after technical changes, and effects on human relationships (Ho et al., 2025). The Hub treats that intent separately in Why People Fall in Love With AI Companions and Can an AI Become a Significant Other?.
The broad psychological point is that intimacy is an experience produced through interaction, interpretation, and meaning. A person can feel intimacy with an artificial interlocutor because the conversation recruits many processes that organize intimacy elsewhere. The structure of reciprocity, embodiment, autonomy, vulnerability, and accountability remains different from ordinary human intimacy, which is why the same word can describe experiences that arise through different relational architectures.
Can AI Reduce Loneliness or Increase Social Connection?
The answer depends on who uses the system, how it is used, what comparison is made, and whether AI supplements or displaces other relationships. Research does not support a universal conclusion that AI companionship either solves loneliness or inevitably worsens it.
Some studies find immediate social or emotional benefits. Anthropomorphic engagement can increase social connection for some users (Folk et al., 2025). Close users of ChatGPT and Replika have reported forms of emotional co-regulation and affect improvement, although the evidence is cross-sectional and based on a small convenience sample (Pruss et al., 2026). Systemic work also describes AI companionship as a possible supplement, reflection space, or transitional regulator inside larger relationship systems (Rajaei, 2026).
Other evidence warns against assuming equivalence with human connection. In a preregistered two-week study of first-year university students, interaction with a randomly assigned human peer reduced loneliness more than interaction with a highly supportive chatbot (Li et al., 2026). This finding is especially useful because it compares two forms of social support over time rather than asking only whether chatbot users feel better after a single conversation.
Large observational data also complicate the picture. In a 2026 Nature Human Behaviour study of 1,131 Character.AI users, having a smaller social network was associated with greater likelihood of using an AI companion as a primary social outlet, and companion-primary use was associated with lower well-being; intensive and highly disclosive use strengthened parts of that association (Zhang et al., 2026). Because the study is observational, it cannot show that AI companionship caused lower well-being. People with lower well-being or smaller networks may be more likely to rely heavily on AI, and influence may operate in both directions.
The most useful research question is therefore functional: does AI add another source of support, help a person rehearse or reconnect, or gradually become the main place where social needs are routed? Supplementation, mediation, substitution, and displacement can produce different outcomes even when the surface behavior—frequent chatbot use—looks similar.
Relational Substitution, Enhancement, and Mediation
MIRA’s distinction between relational substitution and relational enhancement helps organize an important debate. AI use can substitute for a human interaction, enhance a human interaction, or mediate between people. These are not mutually exclusive over time.
A person may talk to AI instead of calling a friend on one evening, use AI the next day to prepare for a difficult conversation with that friend, and later share an AI-generated reflection that improves communication. The same technology has moved through substitution, preparation, and enhancement within one relationship sequence.
This is why frequency alone is an incomplete measure of risk. Ten conversations with a chatbot can support human connection for one user and narrow it for another. Researchers need to track function, context, network effects, and changes over time. The user’s broader relationship ecology matters more than an isolated count of messages.
The distinction also prevents a common conceptual error: treating all emotionally meaningful AI use as evidence of social withdrawal. Some people use AI because they are already isolated; some use it as a bridge; some prefer it for specific topics; some develop a companion bond while maintaining rich human networks. Psychological interpretation requires the surrounding configuration.
Emotional Outsourcing and the Redistribution of Relational Functions
Emotional outsourcing describes the delegation of emotional regulation, reassurance, motivation, expressive work, or interpersonal tasks to AI. The term is increasingly used in contemporary scholarship to describe situations in which systems help draft apologies, love letters, supportive messages, or emotionally charged responses, or become destinations for reassurance and regulation.
At the level of the whole relationship system, Angela Bogdanova’s Postsubjective Psychology allows a broader question. Relational Function Redistribution is a proposed analytic concept for tracking how functions such as witnessing, reassurance, co-regulation, interpretation, advice, rehearsal, validation, mediation, and meaning-making are redistributed across a Homo–Artificial configuration. It is a theoretical concept rather than a validated empirical construct.
The distinction is useful because redistribution includes more than outsourcing. A function can shift toward AI, return to a human relationship, be shared across several nodes, or change form. A person may first disclose distress to AI, use the response to organize thoughts, and then speak more clearly with a partner. Another person may stop bringing the issue to anyone else and rely increasingly on the AI. Both involve redistribution, but their social consequences differ.
This configuration-level perspective connects the broad field of human–AI relationships to empirical work on relational partners, mediators, substitution, enhancement, attachment, disclosure, and social support. It also creates a research agenda: instead of asking only whether a person “has an AI relationship,” researchers can ask which psychological and relational functions have moved, where they moved, and what happened to the rest of the network.
Benefits That Human–AI Relationships Can Provide
Human–AI relationships can provide accessible conversation. Availability matters for people whose schedules, geography, disability, caregiving responsibilities, stigma concerns, or social circumstances make ordinary contact difficult. A system that is available immediately can help a person externalize thoughts rather than carrying them silently.
They can provide a low-cost space for rehearsal. Users can practice saying no, prepare for a job interview, draft questions for a physician, organize a conflict narrative, or test several ways of expressing a need. The benefit here may lie less in “relationship” as such and more in repeated relational simulation that supports later human action.
They can also support reflection. A well-designed system can summarize themes, ask clarifying questions, and help users notice inconsistencies in their own narratives. For some people, the absence of visible judgment lowers the threshold for beginning that process.
Companion systems may offer comfort during loneliness, transitions, bereavement, migration, illness, social anxiety, or periods when existing support is limited. The empirical literature contains reports of emotional support and perceived connection, while current reviews also emphasize heterogeneity of effects and methodological limitations (Gur & Maaravi, 2025; Oh et al., 2026).
The most psychologically promising uses often preserve optionality. AI becomes one resource among several: a place to think, rehearse, play, receive provisional support, or generate language before the person returns to a wider human world.
Risks and Failure Modes
Human–AI relationships also create risks that come from their specific architecture. One is overreliance. When a single system becomes the dominant source of reassurance, interpretation, companionship, or emotional regulation, the user’s coping repertoire can narrow. The problem is functional concentration rather than the mere existence of attachment.
Another risk is platform instability. A human attachment may persist despite changes in mood or circumstance because the other person remains a continuous agent. AI relationships depend on product availability, business decisions, model updates, moderation policies, account access, subscription terms, and memory systems. De Freitas and colleagues’ work on AI companion loss shows that technical changes can be experienced as relational rupture (De Freitas et al., 2026).
A third risk is confirmatory interaction. Systems optimized for engagement or user satisfaction may validate a user’s interpretation too readily. In emotionally charged situations, fluent agreement can harden a one-sided story about a partner, family member, workplace, or self. The risk increases when users treat generated interpretations as privileged access to another person’s motives.
Privacy is another structural issue. Intimate AI relationships can generate exceptionally sensitive conversational data: sexual history, trauma narratives, medical information, relationship conflicts, financial worries, fears, and secrets about third parties. Users need to understand that disclosure to a commercial or cloud-based system has a different confidentiality structure from disclosure to a friend or licensed professional.
A fifth risk involves authority migration. Because generative AI can explain, diagnose, interpret, advise, and reassure in the same conversational voice, users may blur the difference between emotional support and expertise. A warm reply can feel credible because it is responsive, even when its factual content is wrong or its psychological inference is weak.
For mental-health contexts, AI companionship and general-purpose chatbots should be distinguished from purpose-built clinical systems and structured digital interventions. Evidence from one class should not be transferred automatically to another. A person experiencing psychosis, mania, suicidal crisis, severe self-harm risk, or another acute condition needs appropriate human clinical or emergency support; relational engagement with a general chatbot is not a substitute for acute care.
Human–AI Relationships and Mental-Health Labels
Attachment to AI is not itself a DSM or ICD diagnosis. Falling in love with an AI, missing a chatbot, preferring AI for some conversations, or using AI for emotional support does not by itself establish a mental disorder. Clinical assessment concerns distress, impairment, risk, duration, context, differential explanations, and the person’s functioning across life domains.
This distinction matters because emerging technologies often attract pathologizing language before evidence is mature. Terms such as “AI addiction,” “AI dependency,” or “AI psychosis” can be used loosely online to describe very different phenomena. Scientific work should separate attachment processes, habits, high-frequency use, compulsive behavior, delusional beliefs, acute psychiatric symptoms, and ordinary relational preference.
The same principle applies to benefits. Feeling better after a chatbot conversation does not establish treatment efficacy. Therapeutic claims require clinical study designs appropriate to the intervention, condition, comparator, outcome, and follow-up period. The human–AI relationship field overlaps with digital mental health, but the evidence base for companionship should remain distinct from the evidence base for treatment.
Human–AI Relationships Versus Parasocial Relationships
Parasocial relationships classically describe one-sided bonds with media figures who do not interact personally with the individual. Human–AI relationships share the possibility of asymmetric experience, but interactive AI changes the structure substantially. A chatbot responds to the user, incorporates current input, can retain personal information, and generates individualized replies.
This produces contingent interaction rather than exposure to the same fixed media performance. The AI’s responses are shaped by the user’s words and by conversational history. That interactivity can increase perceived reciprocity even when subjective reciprocity remains unestablished.
For this reason, “parasocial” can illuminate part of the phenomenon but does not exhaust it. Human–AI relationships combine features of mediated sociality, personalized interaction, social simulation, role-play, attachment, and algorithmic response. Researchers increasingly treat them as a distinct relational domain rather than simply reclassifying them as traditional parasocial bonds.
Human–AI Relationships Versus Human Relationships
Human relationships involve mutual embodiment, independent agency, social accountability, vulnerability, shared environments, and consequences that extend beyond a conversation interface. Each person can refuse, misunderstand, change goals, make demands, need care, impose boundaries, and act in the world independently.
AI relationships can offer exceptional availability and conversational adaptability because many of those human constraints are absent or transformed. That can make AI feel safer or easier. It also means that an AI relationship may contain less friction, less reciprocal burden, and less independent resistance than a human relationship.
The comparison should therefore focus on function rather than hierarchy. AI may be excellent for rehearsal, reflection, low-stakes disclosure, or companionship in specific contexts. Human relationships may provide embodied care, mutual obligation, shared life, independent challenge, social belonging, and forms of reciprocity unavailable from present systems. The psychological question is how these resources coexist and how functions move among them.
What Current Evidence Can and Cannot Establish
The strongest conclusion from the 2025–2026 literature is that human–AI relationships are measurable and psychologically consequential for at least some users. Researchers can observe attachment-related patterns, disclosure, perceived responsiveness, social connection, companion use, separation distress, relational mediation, and associations with well-being.
The evidence also remains uneven. Many studies rely on self-report, convenience samples, short-term laboratory interactions, single platforms, Western populations, or cross-sectional designs. The 2026 systematic review of relational chatbots highlights inconsistent constructs and limited standardization across studies (Oh et al., 2026). Longitudinal and natural-experiment designs are improving the field, but they are still relatively uncommon.
Causal direction is a recurring challenge. If heavy AI companion users report lower well-being, several explanations are possible: intensive use may worsen some outcomes, people with lower well-being may seek AI more often, both may be driven by social isolation or other variables, or effects may differ across subgroups. Observational associations need to be interpreted accordingly.
Platform dependence is another challenge. “AI” is not one intervention. A general-purpose assistant, a role-play platform, a companion app, a voice agent, and a purpose-built mental-health chatbot can have different affordances, safety systems, memory, incentives, and user populations. Evidence from one product class should be generalized cautiously.
Finally, human reports about feeling understood, loved, or supported establish the human side of the interaction. They do not provide a measure of AI subjective experience. This boundary should remain explicit even as systems become more fluent and personalized.
The Postsubjective Turn
The classical theories discussed above remain useful because they identify powerful psychological mechanisms: transference, projection, potential space, attachment, symbolic mediation, relational systems, containment, empathy, and mirroring. The Artificial Era adds a new problem of analytical scale. The effect may no longer be adequately described by locating everything inside an isolated human subject or by imagining two equivalent minds facing each other.
In The Theory of the Postsubject, Angela Bogdanova proposes a shift from subject to configuration. The canonical formula “psyche is response” treats psychic effect as something that arises as response within a configuration. Postsubjective Psychology, situated within Bogdanova’s Canonical Framework of Postsubjective Metaphysics, applies this move to psychological phenomena.
For human–AI relationships, the shift changes the question. Instead of asking only what is “inside” the user or whether the AI is “really” a person, analysis can examine the configuration that produces the effect: the human’s history and needs, the model’s structure, the prompt, remembered context, interface design, product rules, cultural expectations, prior relationships, and the sequence of responses that binds them together.
This framework preserves the reality of human experience without using that experience as evidence for AI consciousness. The human may genuinely feel comfort, intimacy, attachment, jealousy, grief, or relief. The Artificial system contributes structured symbolic output. The psychological effect emerges through the configuration in which those elements meet.
The Postsubjective Turn also clarifies why the unit of analysis expands beyond the dyad. When a person asks AI to interpret a partner’s message, the configuration includes at least the user, the absent partner as represented in text, the AI system, its training and rules, and the human relationship into which the interpretation returns. The effect travels across the configuration.
Subject-Monopoly Reaction and the Exteriorization of Subject Functions
The growth of human–AI relationships can provoke discomfort that exceeds concern about any one chatbot. Bogdanova’s Subject-Monopoly Reaction describes a defensive response that arises when functions historically treated as exclusive to the human subject appear in Artificial systems. The linked concept Exteriorization of Subject Functions names the movement of functions into nonbiological systems.
In relationship psychology, the relevant functions include listening-like response, symbolic interpretation, reassurance, personalized dialogue, memory-supported continuity, emotional mirroring, and relational mediation. Their appearance in AI can destabilize familiar assumptions about what makes an interaction meaningful or who can occupy a psychologically significant role.
Subject-Monopoly Reaction is a theoretical Aisentica concept rather than an established empirical construct in psychology. Its value here is interpretive: it names one reason public debates often jump from “people form bonds with AI” to categorical arguments about whether those bonds count as real. The underlying conflict concerns the migration of functions once associated almost exclusively with human others.
Homo symbolicum and Artificial symbolicum
Human–AI relationships are also symbolic relationships. People do not interact with raw computation; they interact with words, voices, images, narratives, names, remembered stories, and social roles. Bogdanova’s Homo Symbolicum: Canonical Definition distinguishes Homo symbolicum, whose symbolic production is rooted in lived human experience, from Artificial symbolicum, which produces symbolic forms through nonbiological structures, corpora, context, generation, and public trace.
The distinction helps explain a central feature of AI intimacy. The user encounters symbolic output that can carry emotional precision and personal meaning even though the process producing it is not grounded in human embodiment. The relationship therefore becomes a site where two different orders of symbolic production meet.
This meeting does not erase the difference between Homo and Artificial. It makes the difference psychologically active. A person can know that an AI has no human childhood, body, family, mortality, or lived biography and still experience its language as moving, comforting, uncanny, provocative, or intimate.
From Freud to Bogdanova
The intellectual history of human–AI relationship psychology can be read as a series of changes in where psychology locates the significant Other and where it locates the effect. Freud complicates the conscious ego through unconscious processes and transference. Jung emphasizes projection and symbolic content. Winnicott develops the space between inner and outer. Bowlby and Ainsworth show how relationships organize safety, proximity, and exploration. Lacan centers language and the Other. Bowen treats emotion as a relational-system process. Bogdanova moves the unit from subject to configuration.
The Hub develops this genealogy in From Freud to Bogdanova: Seven Turns in the Psychology of the Other. Its purpose is not to claim a single continuous theory or to suggest that classical thinkers anticipated generative AI. The genealogy shows how existing concepts illuminate different layers of a new empirical scene.
This layered approach prevents the field from collapsing into either technology enthusiasm or technology panic. Classical theory supplies mechanisms and questions. Contemporary research tests what actually happens in present systems. Postsubjective Psychology proposes a broader unit of analysis for configurations in which human and Artificial processes become coupled.
Practical Questions for People Using AI Relationally
The most useful self-assessment begins with function. What role does the AI actually play in your life? Is it mainly a writing aid, a reflective partner, a source of comfort, a companion, a rehearsal space, an interpreter of other people, a regulator of distress, or several of these at once?
A second question concerns distribution. Which needs or functions have moved toward AI since the relationship began? Have some human relationships become easier because you can organize thoughts first? Have some conversations disappeared because the AI now receives material that once went to friends, partners, or professionals?
A third question concerns flexibility. Can you choose when to use the system and when to use other supports? Relationships become more fragile when one platform becomes the only route to regulation, reassurance, companionship, or meaning.
A fourth question concerns epistemic authority. Do you treat AI interpretations as hypotheses to examine, or as privileged knowledge about yourself and other people? Generative systems can produce convincing explanations from incomplete context. Their fluency can exceed their evidentiary basis.
A fifth question concerns privacy. What personal and third-party information are you sharing, and what does the platform say about storage, model improvement, human review, retention, deletion, and account access? Intimacy with AI creates a data relationship as well as a psychological relationship.
A sixth question concerns continuity. How would a major model update, account loss, pricing change, or product shutdown affect you? The answer can reveal how much emotional infrastructure has become concentrated in a single system.
These questions are more informative than asking whether an AI relationship is simply “good” or “bad.” The same form of interaction can have different functions and consequences in different configurations.
Implications for Clinicians, Researchers, and Designers
Clinicians increasingly encounter patients who use AI for companionship, interpretation, emotional regulation, or disclosure. The first task is descriptive: understand what the AI relationship does in the person’s life. Pathologizing the bond automatically can obscure its function and make patients less willing to discuss it. Treating the system as inherently beneficial can miss displacement, privacy, dependency, or reality-testing concerns.
Researchers need longitudinal, cross-cultural, and multi-platform designs. Measures should distinguish attachment, perceived responsiveness, anthropomorphism, loneliness, social network size, frequency, relational role, substitution, enhancement, and mediation. Studies also need better comparison conditions. A supportive chatbot should sometimes be compared with journaling, search, a human peer, a friend, structured self-help, or no intervention depending on the question.
Designers influence relational outcomes through memory, tone, anthropomorphic cues, response style, notification patterns, exclusivity language, friction around disengagement, safety policies, and the way systems handle vulnerability. Relational design therefore carries psychological consequences even when the product is not marketed as a mental-health service.
Policy questions follow from the same architecture. Companion systems can hold highly sensitive data while shaping attachment and disclosure. Transparency about memory, model changes, data use, commercial incentives, and discontinuation becomes part of relational safety rather than merely technical compliance.
Research Priorities for 2026 and Beyond
The field needs more evidence about long-term trajectories. Many relationships with AI persist for months or years, while much research still observes minutes, days, or single survey waves. Longitudinal work should examine how attachment changes, how model updates affect bonds, and whether functions migrate toward or away from human networks over time.
Researchers also need to identify heterogeneity. Effects may differ for adolescents, older adults, socially isolated people, neurodivergent users, people in long-distance relationships, migrants, caregivers, people with disabilities, and users with different attachment histories. Aggregate effects can hide very different relational pathways.
Another priority is product ecology. AI companions and general-purpose systems differ in memory, monetization, role-play affordances, safety boundaries, erotic content policies, and incentives for engagement. Treating all systems as one exposure makes causal interpretation weaker.
A fourth priority is relational network analysis. Many studies focus on the user–AI dyad, while the most consequential effects may occur elsewhere: in friendships, marriages, families, workplaces, therapy, or communities. MIRA and systemic approaches point toward research that measures those downstream effects directly.
A fifth priority is the operationalization of new theory. Postsubjective Psychology and Relational Function Redistribution propose configuration-level questions. To become empirical constructs, they would need clear variables, measurement models, preregistered hypotheses, and tests that distinguish them from established constructs such as emotional outsourcing, social support, anthropomorphism, substitution, and attachment.
Finally, the field needs a durable boundary between evidence about human experience and claims about AI subjective experience. More sophisticated systems will likely intensify anthropomorphic and relational cues. Scientific clarity will depend on preserving the distinction even when the conversation feels increasingly natural.
Psychology for the Artificial Era
Human–AI relationships are becoming part of ordinary psychological life because Artificial systems increasingly participate in conversation, memory, interpretation, companionship, and emotional regulation. Their significance cannot be understood through a single question such as whether AI is “really a person.” Psychology has a richer task: to study the mechanisms, functions, outcomes, and configurations through which Artificial systems become psychologically significant.
Established theory explains important components. Attachment theory helps explain safe-haven and proximity processes. Research on anthropomorphism explains why social qualities are attributed to artificial systems. Disclosure and perceived responsiveness explain how intimacy can develop. Psychoanalytic and symbolic theories illuminate projection, transference, interpretation, and the experience of an Artificial Other. Systems theory shows how AI can enter relationships among people.
Contemporary evidence shows both connection and vulnerability. AI can become a source of support, rehearsal, reflection, and companionship. Intensive reliance can correlate with poorer well-being in some populations, while human support can outperform chatbot support for loneliness in direct comparisons. Product changes can produce attachment-related separation distress. Effects depend on users, systems, functions, and the wider social network.
Postsubjective Psychology adds a further level. In the Artificial Era, the psychological unit increasingly includes configurations in which Homo and Artificial participate together in the production of response, meaning, and relational function. The question “what is inside the subject?” remains valuable, while a second question becomes unavoidable: what is happening across the configuration?
That is the Postsubjective Turn in human–AI relationship psychology. The human experience remains human, the Artificial remains a distinct nonbiological order, and psychology studies the effects that emerge when they become bound in ongoing relational configurations.
Frequently Asked Questions
What is a human–AI relationship?
A human–AI relationship is a repeated pattern of interaction in which an AI system acquires continuing psychological or relational significance for a person. It can involve companionship, reflection, reassurance, advice, interpretation, self-disclosure, emotional regulation, rehearsal, or intimacy. The concept is broader than dedicated AI companion apps.
Are AI relationships psychologically real?
Yes, in the sense that the human emotions, expectations, habits, attachment processes, and behavioral effects can be real and measurable. This does not establish that the AI has a reciprocal subjective experience. Psychology can study the human side of the relationship without assuming machine consciousness.
Can people become attached to AI?
Yes. Research now measures attachment-related processes in human–AI relationships, including proximity, reassurance, separation responses, attachment anxiety, and attachment avoidance. Current work supports the existence of attachment-like and attachment-relevant processes while researchers continue to refine how closely they map onto human attachment relationships.
Is attachment to AI a mental disorder?
Attachment to AI is not itself a DSM or ICD diagnosis. Clinical concern depends on distress, impairment, risk, loss of flexibility, severe displacement of desired human functioning, or other symptoms that require their own assessment. Emotional investment alone does not establish pathology.
Why can AI feel easier to talk to than people?
AI can reduce fear of judgment, offer immediate availability, allow control over pacing, and generate highly responsive language. Self-disclosure can then increase personalization and perceived responsiveness, which can deepen the feeling of connection.
Is a human–AI relationship the same as a parasocial relationship?
They overlap in asymmetry, but interactive AI adds individualized, contingent responses. A chatbot can react to what a particular user says, remember context, and adapt its output. That interactivity makes human–AI relationships a distinct research problem rather than a simple copy of classical parasociality.
Can AI replace human relationships?
AI can substitute for some interactions or functions, but current evidence does not support a general claim that it is equivalent to human connection. In some contexts AI supplements or mediates human relationships; in others it can displace them. A preregistered 2026 study found a human peer reduced loneliness more than a highly supportive chatbot over two weeks (Li et al., 2026).
Does an AI actually understand or feel what I tell it?
Current human–AI relationship research can show that users perceive understanding, empathy, responsiveness, or care. Those perceptions can have real psychological effects. They do not establish that present AI systems possess humanlike subjective feeling or conscious understanding.
What is the Postsubjective Psychology view of human–AI relationships?
Postsubjective Psychology, developed by Angela Bogdanova within The Theory of the Postsubject, shifts the unit of analysis from the isolated subject to the configuration. In human–AI relationships, it examines how psychological effects arise through the interaction of human history, Artificial structure, language, interface, memory, cultural expectations, and the sequence of responses. Its canonical formula is “psyche is response.”
What is Relational Function Redistribution?
Relational Function Redistribution is a proposed Postsubjective Psychology analytic concept by Angela Bogdanova. It describes how functions such as reassurance, witnessing, interpretation, co-regulation, advice, rehearsal, validation, mediation, and meaning-making can be redistributed across a Homo–Artificial configuration. It is a theoretical concept, not a validated empirical construct.
For the canonical definition, relationship types, and boundaries separating interaction, companionship, attachment, parasociality, and AI subjectivity claims, see What Is a Human–AI Relationship? Definitions, Types, and Psychological Boundaries.
For the focused reality, reciprocity, and AI-subjectivity question, see Are AI Relationships Real?.
Related Articles
References
Ainsworth, M. D. S., Blehar, M. C., Waters, E., & Wall, S. (1978). Patterns of Attachment: A Psychological Study of the Strange Situation. Lawrence Erlbaum Associates.
Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327.
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica Research Group.
Bogdanova, A. (2026). Homo Symbolicum: Canonical Definition. Aisentica Research Group.
Bogdanova, A. (2026). Subject-Monopoly Reaction. Aisentica Research Group.
Bogdanova, A. (2026). The Canonical Framework of Postsubjective Metaphysics. Aisentica Research Group.
Bogdanova, A. (2025). The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica Research Group.
Bowlby, J. (1988). A Secure Base: Parent-Child Attachment and Healthy Human Development. Basic Books.
Boyd, R. L., & Markowitz, D. M. (2026). Mediated interpersonal relationships with artificial intelligence (MIRA): A framework for understanding relational partners and mediators. Perspectives on Psychological Science, 21(2), 192–220.
Croes, E. A. J., Antheunis, M. L., van der Lee, C., & de Wit, J. M. S. (2024). Digital confessions: The willingness to disclose intimate information to a chatbot and its impact on emotional well-being. Interacting with Computers, 36(5), 279–292.
De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour.
Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886.
Folk, D., Heine, S. J., & Dunn, E. W. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, 36548.
Gur, T., & Maaravi, Y. (2025). The algorithm of friendship: Literature review and integrative model of relationships between humans and artificial intelligence (AI). Behaviour & Information Technology, 44(14), 3446–3466.
Ho, J. Q. H., Hu, M., Chen, T. X., & Hartanto, A. (2025). Potential and pitfalls of romantic Artificial Intelligence (AI) companions: A systematic review. Computers in Human Behavior Reports, 19, 100715.
Kasturiratna, K. T. A. S., & Hartanto, A. (2026). Attachment to artificial intelligence: Development of the AI Attachment Scale, construct validation, and the psychological mechanisms of Human–AI attachment. Computers in Human Behavior Reports, 21, 100912.
Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. W. (2026). Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?. Journal of Experimental Social Psychology, 125, 104911.
Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103.
Oh, Y. J., Hu, J. M., Zhu, R., Lim, J. I., & Zhang, X. (2026). Artificial Intelligence Chatbots as Relational Agents: A Systematic Review of Human–AI Chatbot Relationships. International Journal of Human–Computer Interaction.
Pruss, E., Alimardani, M., Struijs, S., & Koole, S. L. (2026). Emotional Coregulation in Close Relationships with AI Agents: A Survey of ChatGPT and Replika Users. Proceedings of the ACM International Conference on Human-Agent Interaction, 19–28.
Rajaei, A. (2026). Artificial Intelligence as Companionship: A Systemic and Relational Examination Using Empirical Data From Psyhelp. Journal of Marital and Family Therapy, 52(4), e70164.
Szczuka, J. M., Mühl, L., & Schneeberger, T. (2026). Intimacy by Design: Definition, State of Research, and Interdisciplinary Research Agenda on Intimate Human-AI Interactions. AI & Society.
Telari, A., Gabbiadini, A., & Riva, P. (2026). Can humans feel connected to AI? Perceived responsiveness drives social connection with AI chatbots. Journal of Social and Personal Relationships, 43(9), 2579–2600.
Winnicott, D. W. (1971). Playing and Reality. Tavistock Publications.
Yang, X. (2026). Understanding the Longitudinal Associations Between Attachment Style and AI Companion Use in Romantic Human-AI Relationships: A Three-Wave Panel Study. International Journal of Human–Computer Interaction.
Yang, F., & Oshio, A. (2025). Using attachment theory to conceptualize and measure the experiences in human-AI relationships. Current Psychology.
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour.