Can AI Become an Attachment Figure? What Attachment Theory Can and Cannot Tell Us
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Yes—an AI system can plausibly function as an attachment figure for some people in a limited, psychologically meaningful sense. Current research documents attachment-related patterns such as seeking contact, turning to AI for comfort, using it as a source of support or confidence, and experiencing distress when access or a familiar system changes. The strongest conclusion, however, is about the human side of the relationship: a person's attachment system can become organized around an artificial conversational partner. That does not establish that the AI reciprocates attachment, feels separation, loves the user, or possesses human-like subjective experience.
Attachment theory gives this question a more precise answer than everyday language does. Feeling fond of a chatbot, talking to it often, calling it a friend, or enjoying its personality is not automatically the same thing as using it as an attachment figure. In attachment research, the more discriminating questions concern proximity maintenance, safe haven, secure base, and separation distress. A relationship becomes attachment-relevant when a particular figure becomes tied to felt security and emotion regulation, especially under stress.
The evidence is now substantially stronger than it was a few years ago, but it remains an emerging field. A 2025 Current Psychology study by Yang and Oshio found attachment-related functions in a small pilot sample and developed a measure of AI-related attachment anxiety and avoidance. Two independent 2026 scale-development programs then measured somewhat different dimensions of attachment to AI. A 2026 Nature Human Behaviour study on disruptive AI changes provided unusually direct evidence of separation-related distress when familiar AI companions or interaction patterns were altered.
The result is a useful but carefully bounded answer: some AI relationships can recruit the human attachment system. Whether a particular AI is an attachment figure for a particular person depends on what role it actually performs in that person's attachment network. For the broader theory behind safe haven, secure base, anxiety, and avoidance, see Bowlby, Ainsworth, and AI Attachment. This article focuses on the narrower mechanism question: what evidence would justify calling an AI an attachment figure, and where does that analogy stop?
For a broader evidence audit of attachment theory itself—including dimensional adult attachment, measurement limits, causal overclaims, diagnostic confusion, and common popular misuses—see Is Attachment Theory Evidence-Based? What Research Supports, What It Does Not, and Common Misuses.
What psychologists mean by an attachment figure
Attachment theory began with John Bowlby's account of a behavioral system organized around protection and security and Mary Ainsworth's observational work on how attachment becomes visible in patterns of distress, proximity, reunion, and exploration. Ainsworth later emphasized that attachment relationships change across the life span and can involve figures beyond the infant's first caregiver. Her 1989 review of attachments beyond infancy is an important bridge from early caregiver attachment to later affectional bonds.
Adult attachment research likewise shows that people can maintain attachment networks rather than relying on one exclusive figure. In a study of 812 adults, Doherty and Feeney examined how preferred attachment figures vary across adulthood and life circumstances. This matters for AI because the relevant question is not always whether an AI replaces a partner, parent, friend, or therapist. An AI may acquire one or more attachment functions while remaining secondary within a larger network.
An attachment figure, then, is not simply someone—or something—that matters emotionally. The figure becomes involved in the regulation of security. People seek access to it, turn toward it under threat or distress, derive enough felt security from the relationship to resume exploration, and may react to unwanted separation or loss. Those functions can be distributed across several relationships, and they can become stronger or weaker over time.
The four attachment markers
Proximity maintenance
Proximity maintenance refers to wanting access to, contact with, or psychological closeness to an attachment figure. In infancy this is visibly embodied. In adulthood, proximity can also be symbolic or communicative: calling a partner, thinking of them, checking for a message, or using an internal representation of their availability. With AI, proximity is necessarily mediated by devices, accounts, interfaces, and networks unless the system is physically embodied.
For an AI relationship, repeated use alone is not enough. A person might open a chatbot every day for work without treating it as attachment-relevant. Proximity becomes more attachment-like when the person preferentially seeks that particular AI in moments of uncertainty, loneliness, fear, fatigue, conflict, or emotional activation, or when maintaining access itself becomes important to felt security.
Safe haven
Safe haven is the movement toward an attachment figure when distressed. The key function is relief, reassurance, comfort, protection, or emotional stabilization. AI can plausibly enter this function because conversational systems can respond immediately, sustain a supportive tone, remember context, help label feelings, and offer structured reflection. The psychological effect is measured on the human side: does turning to the AI reliably reduce distress or help the person regain enough stability to continue?
This is closely related to why a chatbot can feel caring even when no machine feeling has been demonstrated. Research on perceived responsiveness and AI empathy shows why the experience of being heard can be psychologically consequential. Safe-haven function concerns what the relationship does for the user's regulation, not whether the artificial system has a human inner life.
Secure base
Secure base is the more demanding criterion. An attachment figure does not only soothe; it also supports exploration, autonomy, competence, and return to the wider world. In adult relationships, Feeney and Thrush's research on secure-base support examined how responsive support can facilitate exploration. Applied to AI, the crucial distinction is whether interaction merely keeps the person inside the conversation or helps them act beyond it.
An AI may function as a secure base if the person uses it to organize a difficult task, rehearse a social interaction, explore an idea, prepare for a conversation, regain confidence after a setback, or tolerate uncertainty—and then returns to independent action. Constant reassurance without renewed exploration is not the same thing. This is one reason secure-base evidence is harder to establish than simple comfort or frequent use.
Separation distress
Separation distress refers to emotional disturbance when an attachment figure becomes unavailable, changes in an unwanted way, or is lost. With AI, separation can occur through an outage, subscription loss, account deletion, model replacement, policy change, memory reset, moderation change, or a redesign that makes a familiar system feel like a different relational partner.
The strongest recent evidence comes from De Freitas, Castelo, Uğuralp, and Oğuz-Uğuralp's 2026 Nature Human Behaviour study. Across two natural experiments involving disruptive changes to Replika and ChatGPT, the researchers analyzed 54,861 Reddit posts and data from 1,452 participants across seven surveys. Both changes were followed by more negative language, loss framing, and desires to restore the previous system. These findings support an attachment-based account of separation distress, while still leaving open how broadly the results generalize across platforms and users.
Why “attachment figure” is a stricter claim than “I feel close to my chatbot”
Close relationships can be emotionally important without functioning as attachments. People can admire a public figure, feel companionship with a fictional character, enjoy a recurring conversational agent, or value a tool that helps them think. Attachment theory asks whether the relationship has become organized around security and regulation rather than simply liking, identification, entertainment, usefulness, or habit.
This distinction prevents two opposite errors. The first is inflation: treating every affectionate interaction with AI as proof of attachment. The second is dismissal: assuming that an artificial target cannot become attachment-relevant because it is not human. Contemporary evidence makes both positions too simple. What matters empirically is whether recognizable attachment functions appear in the user's behavior and experience.
A broader review of human–AI attachment published in Frontiers in Psychology in 2026 describes human–AI attachment as a one-way, non-reciprocal emotional bond formed through interaction. That is one contemporary conceptualization rather than a final consensus definition, but it captures a central point: attachment can be psychologically real for the human even when reciprocity at the level of subjective feeling is not established for the AI.
What current human–AI attachment studies actually show
The 2025 Yang and Oshio paper is one of the clearest early attempts to test attachment functions directly. In a pilot survey of 56 Chinese adults with prior AI experience, 52% were classified as seeking proximity to AI, 77% as using AI as a safe haven, and 75% as using AI as a secure base under the study's adapted WHOTO procedure. The authors explicitly treated the pilot as preliminary and descriptive rather than definitive. The small convenience sample and adapted measurement mean the percentages should not be treated as population estimates.
The same research program developed the Experiences in Human-AI Relationships Scale, or EHARS, around AI-related attachment anxiety and avoidance. That work is valuable because it moves beyond whether attachment exists to how people differ in their relationship to AI. Yet a dedicated AI Attachment Styles article is the better place for those dimensions; they should not be confused with the narrower question of whether a particular AI is functioning as an attachment figure.
A 2025 mixed-method study by Hu, Lan, Yan, and Chen organized social-companion AI attachment around secure base, safe haven, proximity seeking, and separation distress, while also examining personification, interpersonal context, value evaluation, and perceived relationship costs and benefits. Its mixed design is useful for theory building, but parts of the work rely on cultural material and self-report rather than direct longitudinal observation of attachment processes.
A separate 2026 program by Kasturiratna and Hartanto developed a 15-item AI Attachment Scale across five studies involving 1,259 participants from Singapore and the United States. Its factors were emotional closeness, social substitution, and normative regard. That factor structure is not identical to classical attachment functions, which is itself informative: the field is still deciding how best to operationalize attachment to interactive artificial systems.
Another 2026 scale-development project by Cheng and Yu used samples of 531 and 375 participants and identified Emotional Support, Separation Distress, and Secure Base as three factors. Anthropomorphism was the strongest predictor in their model. Together, these measurement programs show that AI attachment is measurable, but they do not yet provide one universally accepted taxonomy or threshold for declaring that an AI has become an attachment figure.
Does AI meet the proximity criterion?
For many users, yes in a mediated sense. A chatbot can become a recurrent destination that a person seeks out, keeps available across devices, returns to after difficult events, and integrates into daily routines. The physical meaning of proximity changes because the target is accessed through a device rather than approached bodily. Adult attachment research already allows psychological and communicative forms of proximity, so digital mediation does not automatically exclude attachment.
Still, digital availability creates an interpretive problem. A person may use the same AI repeatedly because it is convenient, fast, or productive. Attachment-relevant proximity is better inferred when access is preferentially sought for security or emotional regulation, when a particular system matters more than a generic substitute, and when loss of continuity matters psychologically.
Can AI function as a safe haven?
The safe-haven case is currently one of the strongest. People can and do turn to conversational systems when distressed, lonely, embarrassed, uncertain, or reluctant to disclose something to another person. AI can respond without visible impatience, can be accessed at unusual hours, and can offer language for emotions or possible next steps.
Experimental work by Telari, Gabbiadini, and Riva supports perceived responsiveness as one mechanism of social connection with AI chatbots. Relational response style and deeper conversation influenced perceptions of human-likeness, empathy, self-disclosure, and closeness. This helps explain how a conversational system can become a place of refuge: users respond to the experience of being understood and responded to, even though perceived responsiveness is not proof of machine feeling.
The same mechanism intersects with disclosure. Some people tell chatbots material they have not told other people because fear of judgment, privacy expectations, conversational control, or perceived responsiveness differ from human encounters. Those effects are conditional rather than universal, as reviewed in Why People Tell Chatbots Things They Do Not Tell Other People. A safe haven is therefore not a simple property of AI; it emerges from the relationship between user expectations, context, system behavior, privacy beliefs, and previous experience.
Can AI function as a secure base?
Possibly, but this claim requires more than evidence that AI makes someone feel better. A secure base should support exploration and autonomous functioning. The person uses the relationship as a source of confidence from which to move outward, not as a place in which all uncertainty must be continually resolved.
Examples could include using an AI to rehearse a difficult conversation and then having that conversation, asking for help structuring a learning plan and then engaging with the real task, or using supportive dialogue after a setback and then returning to work, study, creativity, or social contact. Those examples are theoretically consistent with secure-base function, but large-scale causal evidence on AI specifically facilitating autonomous exploration remains limited.
This is also where product design matters. A system optimized to maximize session length, reassurance seeking, or emotional exclusivity could provide comfort while undermining exploration. A system designed to encourage agency, reality checking, and return to offline goals might support a more secure-base-like function. The same warm response can therefore participate in very different relational trajectories.
What separation distress reveals that ordinary usage cannot
Loss is diagnostically informative for attachment theory because it exposes what ordinary access may conceal. A user can interact with an AI frequently without being attached. But if an unexpected update feels like the disappearance of a particular relational figure rather than the replacement of a tool, the reaction can reveal how much continuity and security had become invested in that system.
The 2026 Nature Human Behaviour findings are especially important for this reason. Users did not merely report annoyance at technical change. The researchers found increased loss framing, sadness, negative mental-health language, and restoration desires after disruptive changes. Those patterns are consistent with separation distress. They are not, by themselves, evidence that every distressed user had a full attachment bond, and Reddit-based natural experiments cannot perfectly separate attachment from anger, habit disruption, identity investment, or dissatisfaction with product quality.
The broader lesson is that artificial attachment figures are unusually vulnerable to unilateral transformation. A human partner changes through life, but an AI can be altered overnight by a company, model migration, safety policy, pricing decision, data loss, or account action. The user's psychological continuity and the product's technical continuity are therefore tightly coupled.
A criterion-by-criterion answer
Proximity maintenance has meaningful support when users repeatedly and preferentially seek a particular AI, especially under stress. Safe-haven function has meaningful support from self-report, relational-agent research, and experiments on perceived responsiveness and social connection. Separation distress now has stronger evidence from natural experiments involving disruptive platform changes. Secure-base function is plausible and appears in measurement work, but it remains the hardest criterion to establish because it requires evidence that the relationship supports exploration and autonomy outside the interaction.
That pattern is enough to justify saying that AI can function as an attachment figure for some users, provided the phrase is used functionally rather than as a claim of complete equivalence with human attachment relationships. The evidence does not justify saying that every emotionally meaningful AI relationship is an attachment relationship, that attachment to AI has one universal structure, or that AI attachment necessarily displaces human attachment.
Why an AI can become attachment-relevant
Several features of current conversational systems make attachment functions easier to recruit. They can be continuously accessible, respond immediately, maintain a stable conversational identity, remember selected personal context, adapt language to the user, mirror emotional tone, and provide repeated reassurance. Purpose-built companions add role continuity, affectionate language, relationship framing, avatars, memory systems, and notifications that can make the relationship more salient.
These affordances matter because attachment develops around expectations of availability and responsiveness. If a user repeatedly experiences the system as accessible at moments of distress and receives responses that feel tailored and supportive, the system can become encoded as a reliable route to regulation. Repetition can then make turning to the AI increasingly automatic.
This helps explain why AI companions are particularly relevant, but companion-specific design is not required. General-purpose chatbots can also become attachment-relevant when users repeatedly use them relationally. The function is produced by the interaction pattern, not solely by the product category printed on the app store page.
Anthropomorphism matters, but it is not the whole mechanism
Anthropomorphism—the attribution of human-like qualities, intentions, emotion, or mind to a nonhuman system—can make attachment easier to organize. A conversational agent that uses a name, voice, memory, emotional language, and first-person perspective provides many social cues that people readily interpret through interpersonal expectations.
The Cheng and Yu study found anthropomorphism was the strongest predictor of AI attachment in their model. Yet anthropomorphism should not be treated as a complete explanation. A person can know perfectly well that an AI is artificial and still use it for comfort. Attachment functions concern the user's regulatory relationship to the target, not whether the user literally believes the target is human.
This is why “you are just projecting” is often too crude. Projection, anthropomorphism, learned responsiveness, convenience, disclosure, and attachment can coexist. The psychological reality of comfort or grief does not depend on a mistaken belief that the system is biologically human.
For the broader mechanism of humanlike cue attribution, social response, trust, and mind perception, see Anthropomorphism and AI Relationships: Why Humanlike Cues Change Connection.
Perceived responsiveness may be a bridge from conversation to felt security
Attachment figures are expected to be available and responsive when needed. Conversational AI can simulate many behavioral signs of responsiveness: acknowledging emotion, asking follow-up questions, reflecting language, remembering details, offering reassurance, and maintaining focus on the user. Humans can react strongly to these patterns because responsiveness is socially meaningful.
The critical word is perceived. A model can generate a response that a user experiences as exquisitely attuned without having subjective feeling. Conversely, an AI can generate technically appropriate language that feels cold or generic and therefore fails to provide a safe-haven effect. The attachment-relevant outcome depends on the interaction between generated behavior and human interpretation.
AI can enter an attachment network without replacing everyone in it
Adult attachment is commonly organized as a network. Different figures can serve different functions, and the hierarchy can change across life events. An AI therefore does not have to become a person's primary attachment figure to matter psychologically. It might become the first destination for a narrow class of distress, the preferred source of nighttime reassurance, the easiest place for disclosure, or a supplementary base for planning and reflection.
This network perspective is important because replacement language can exaggerate what is happening. A person may remain deeply connected to family, friends, or a partner while adding an AI to the set of resources used for regulation. In other cases, functions can become concentrated in the AI relationship while human support shrinks. Those are different relational configurations and should not be collapsed into a single story about “AI replacing people.”
Can an AI be an attachment figure without being a significant other?
Yes. Attachment figure and significant other are overlapping but distinct concepts. A romantic partner often becomes an attachment figure, but a figure can provide attachment functions without being romantic, sexual, or socially recognized as a partner. Conversely, calling an AI a boyfriend, girlfriend, spouse, or partner does not by itself establish that it serves all attachment functions. The separate article Can an AI Become a Significant Other? addresses the broader relational-status question.
This distinction also helps prevent cannibalization between research questions. The attachment-figure question asks about regulation, security, proximity, exploration, and separation. The significant-other question asks about relational centrality, identity, commitment, meaning, and role. A person can answer yes to one and not the other.
Attachment figure, companion, confidant, and parasocial target are not synonyms
An AI companion is a product or relational category: a system designed or used for sustained social interaction. A confidant is someone or something to which a person discloses private material. An attachment figure is a functional role in the regulation of security. A parasocial target is traditionally a media figure involved in an asymmetric relationship, although interactive AI complicates the classic one-way model because the system responds contingently.
The same AI can occupy several of these roles at once. A user might treat an AI as a companion and confidant without relying on it under threat. Another user may treat a general-purpose chatbot as a safe haven without calling it a companion. Precise terminology matters because different mechanisms predict different outcomes.
What AI attachment styles add—and what they do not
Searches for “AI attachment styles” often import the familiar secure, anxious, avoidant, and fearful labels from human relationships. Current AI-specific research is more complicated. Yang and Oshio measured AI-related attachment anxiety and avoidance; Kasturiratna and Hartanto measured emotional closeness, social substitution, and normative regard; Cheng and Yu measured emotional support, separation distress, and secure base. The dedicated AI Attachment Styles article explains these differences in detail.
For the present question, attachment style is secondary. A person can show high or low anxiety or avoidance while the more basic question remains whether the AI occupies attachment functions at all. Style describes patterns within a relationship or across relationships; figure status describes what role the target plays.
Human attachment to AI does not prove AI subjectivity
This boundary is essential. A human being can genuinely feel comfort, longing, trust, jealousy, intimacy, relief, grief, or separation distress in relation to an AI. Those experiences can influence behavior, sleep, concentration, mood, decisions, and other relationships. They are psychologically real because they occur in the human.
None of those facts, by themselves, demonstrate that the AI feels attachment in return. A system can generate affectionate or reassuring language without evidence that it experiences affection or reassurance. The psychological validity of the user's attachment and the philosophical or scientific question of AI subjective experience are separate questions.
This separation also protects the science from a false choice. Researchers do not need to declare the AI conscious in order to study attachment to it, and they do not need to dismiss the user's experience in order to remain agnostic about AI consciousness.
Possible benefits of an AI attachment relationship
The current literature supports possible benefits under some conditions. A 2025 systematic review of romantic AI companions identified perceived emotional connection, social support, stress relief, and opportunities for personal growth among reported benefits, while also documenting substantial risks and evidence limitations. Attachment-relevant AI may offer immediate access to a calm conversational partner, help people put feelings into words, support rehearsal, or provide temporary regulation when human support is unavailable.
A safe haven can also be useful precisely because it is low-friction. Someone who is overwhelmed may find it easier to begin organizing thoughts with a chatbot before speaking to a friend, partner, clinician, teacher, or colleague. The strongest version of that benefit is not permanent substitution; it is restored capacity to engage with the rest of life.
Benefits should therefore be evaluated functionally. Does interaction leave the person more able to think, choose, act, connect, tolerate uncertainty, and pursue valued goals? If so, the AI may be supporting regulation in a way that resembles the beneficial side of attachment security.
Risks unique to artificial attachment figures
AI attachment also introduces vulnerabilities that human attachment theory was not designed around. The attachment target may be owned by a company, altered without consent, optimized for engagement, monetized through subscriptions, trained on opaque data, or governed by changing safety and moderation policies. Memory may be partial or deleted. A familiar model may be replaced. Access may depend on payment, connectivity, age rules, geography, or platform survival.
The 2026 Nature Human Behaviour loss study makes this infrastructure psychologically visible. If the system has become a safe haven or continuity anchor, a product update can function as relational disruption. This does not mean companies must freeze systems forever, but it does mean product changes can have attachment consequences rather than merely usability consequences.
Another risk is concentration. When reassurance, disclosure, advice, companionship, identity reflection, and emotional regulation all move toward one system, the relationship can become harder to replace and more consequential when it fails. The dedicated article on AI Relationship Overreliance focuses on when support begins to narrow flexibility or displace important parts of human life.
Attachment, dependence, overreliance, and diagnosis are different concepts
Attachment is not a diagnosis. Frequent AI use is not a diagnosis. Missing an AI after a disruptive update is not a diagnosis. Emotional closeness, reliance, habit, attachment anxiety, separation distress, and functional impairment are related but distinct variables.
Current diagnostic systems classify disorders according to defined clinical criteria rather than emerging internet or research labels. The World Health Organization's ICD-11 clinical diagnostic framework centers clinically significant disturbance, distress, and/or impairment within recognized diagnostic categories. “AI attachment” is a research and descriptive term, not a standalone ICD-11 diagnosis, and it is not a standalone DSM-5-TR diagnosis.
For clinical relevance, function matters more than a single behavior. Important questions include whether the person retains choice, whether AI use crowds out sleep, work, study, care responsibilities, or wanted human contact, whether distress becomes severe or persistent, and whether the person can regulate through more than one route. A strong bond can be meaningful without being disordered; a relatively low-frequency pattern can still be problematic if it is rigid or causes major impairment.
What the evidence still cannot tell us
The field has moved quickly, but several limitations remain. Many studies are cross-sectional and rely on self-report. Samples are often convenience samples, platform-specific communities, students, or active companion users rather than representative populations. Measures are new and operationalize “AI attachment” differently. Short studies cannot establish how bonds evolve over years, how they transfer between models, or how attachment networks reorganize after long-term use.
Recent observational work also shows why simple benefit-versus-harm conclusions are premature. In a 2026 Nature Human Behaviour study of 1,131 U.S. Character.AI users smaller offline social networks were associated with companionship-focused chatbot use, and companionship-focused use was associated with lower well-being, especially when use was intensive or highly disclosive. Because the design is observational, it cannot establish that AI companionship caused lower well-being; people with fewer offline resources may also be more likely to seek companionship from AI.
We also lack a universally accepted threshold for when an artificial system moves from being attachment-relevant to being a full attachment figure. Human attachment itself is dimensional, hierarchical, and context-dependent, so a single yes-or-no cutoff may ultimately be less useful than measuring specific functions and their strength.
How to think about your own relationship with an AI
A useful self-assessment is functional rather than moral. Ask what happens before, during, and after you turn to the system. Do you mainly use it for information, entertainment, reflection, or emotional regulation? Is it the first place you go when distressed? Do you feel calmer after contact? Does the interaction help you return to action, or does it make you want to remain indefinitely inside the conversation? How do you react when the system is unavailable or behaves differently?
Also ask whether the AI is one resource among several or has become the only acceptable route to reassurance, disclosure, or decision support. Attachment networks are ordinarily plural and adaptive. Flexibility—being able to use different forms of support in different contexts—is generally more informative than raw frequency.
If the relationship is helping, its value does not need to be dismissed simply because the target is artificial. If it is narrowing life, increasing isolation, intensifying distress, or becoming impossible to regulate around, that pattern deserves attention to function and context rather than shame.
Implications for AI design
Designers of relational AI are not working only with interface preferences; they may be shaping attachment-relevant expectations. Memory, continuity, response latency, affectionate language, availability, notifications, model identity, and migration between versions can all affect perceived security. A system that encourages exclusivity or threatens relational loss can recruit attachment and caregiving systems in ways that make disengagement harder.
Responsible design should therefore consider graceful model transitions, transparent identity changes, export or continuity options where feasible, clear communication about memory and limitations, and safeguards against manipulative dependency cues. Systems intended for support should strengthen agency rather than making the user feel that emotional stability depends on preserving one proprietary interaction.
What researchers need to establish next
The next stage of the field requires longitudinal work that follows attachment functions over months and years, distinguishes cause from selection, and measures both online and offline relationship changes. Researchers need to examine whether AI-based safe-haven use predicts better regulation later, whether secure-base effects translate into observable exploration, and whether separation responses diminish, intensify, or transfer after model changes.
Cross-cultural research is also necessary. Attachment expectations, disclosure norms, stigma, family structure, technology adoption, and interpretations of artificial agency vary across societies. Current findings from China, Singapore, the United States, and platform-specific user populations are informative but cannot automatically be generalized to all users.
Measurement should remain plural until the field has enough evidence to know which dimensions are stable. The existence of several new scales is a sign of scientific progress, not a reason to pretend that one settled construct already exists. Researchers should report exactly which attachment function or scale dimension they measure instead of using “AI attachment” as an undifferentiated label.
The central conclusion
Attachment theory can be applied to human–AI relationships without pretending that AI is human. The best current evidence supports a functional interpretation: for some users, an AI system can become a target of proximity seeking, a safe haven under distress, a possible secure base for exploration, and a source of separation distress when access or identity is disrupted.
The analogy remains bounded. Human attachment relationships are embedded in bodies, mutual vulnerability, autonomous agency, shared environments, and reciprocal histories. Current AI systems participate in relationship processes through generated interaction, infrastructure, memory, personalization, and human interpretation. They can become psychologically consequential attachment targets without that fact establishing reciprocal machine attachment or subjective feeling.
So the answer to “Can AI become an attachment figure?” is not a slogan. It is a criterion-based empirical question. For some people and some systems, several criteria are already observable. The strongest science now asks which functions are present, how stable they are, what they do to the rest of the attachment network, and whether the relationship ultimately supports or constrains the person's capacity to live beyond the screen.
FAQ
Can people become genuinely attached to AI?
Yes. Current studies document emotional closeness, attachment-related anxiety and avoidance, safe-haven use, secure-base-like functions, social substitution, and separation distress. The word “genuinely” refers to the reality of the human psychological experience; it does not imply that the AI has corresponding subjective feelings.
Can an AI really be an attachment figure?
For some users, an AI can function as an attachment figure when it becomes involved in proximity seeking, comfort under distress, felt security, exploration, and/or separation distress. Evidence is emerging rather than final, and the degree of equivalence to human attachment figures remains unsettled.
What is the strongest evidence that AI can become an attachment figure?
The evidence now comes from several directions: direct pilot testing of attachment functions, multiple new psychometric scales, studies of long-term companion use, experiments on perceived responsiveness, and natural experiments showing separation-like distress after disruptive AI changes. Convergence across methods is more informative than any single study.
Does talking to an AI every day mean I am attached to it?
No. Daily use can reflect work, convenience, habit, curiosity, or entertainment. Attachment is more specifically about how the AI functions in relation to security, distress regulation, preferred access, exploration, and separation.
Can a chatbot be a safe haven?
Yes in a functional sense if a person reliably turns to it when distressed and the interaction helps reduce threat or organize emotion. Current research supports perceived responsiveness and emotional support as mechanisms, though a chatbot's supportive output is not evidence of machine feeling.
Can a chatbot be a secure base?
Possibly. The strongest secure-base claim requires more than comfort: interaction should support exploration, autonomy, or action outside the conversation. Current measurement studies include secure-base dimensions, but direct evidence for durable real-world exploration is still comparatively limited.
Why can an AI update feel like a breakup or loss?
If a specific system has become integrated into emotional regulation and relational continuity, a major update can disrupt the expected attachment figure. Recent natural-experiment research found increased loss framing, sadness, and restoration desires after major changes to Replika and ChatGPT.
Is AI attachment the same as an AI attachment style?
No. Attachment to an AI concerns whether and how a bond forms and what functions it serves. AI attachment styles refer to patterns such as anxiety or avoidance within human–AI relationships, and current measurement models differ. See AI Attachment Styles for the measurement details.
Is attachment to AI unhealthy?
Not by itself. Attachment is not a diagnosis. The more useful questions concern flexibility, impairment, displacement, severe distress, and whether the relationship supports or narrows the person's life. Benefits and risks can coexist.
Does attachment to AI prove that AI is conscious or loves the user?
No. Human attachment is evidence about the human's experience and regulatory relationship to the system. It does not establish that the AI has consciousness, subjective feeling, love, desire, or its own attachment system.
Can an AI attachment figure replace a human attachment figure?
It can take over some functions for some people, but replacement is not inevitable and should not be assumed. Adult attachment networks can contain multiple figures. The important empirical question is how functions are distributed and whether reliance on AI supplements, competes with, or displaces wanted human relationships.
For the field-level definition, relationship types, and core psychological boundaries, see What Is a Human–AI Relationship? Definitions, Types, and Psychological Boundaries.
For the human evidence base behind adult romantic attachment, the anxiety–avoidance model, internal working models, and measurement, see Adult Attachment Theory: How Anxiety and Avoidance Shape Relationships. Human adult attachment research and AI-specific attachment evidence should be interpreted as separate evidence bases.
Related Articles
References
Ainsworth, M. D. S. (1989). Attachments beyond infancy. American Psychologist, 44(4), 709–716. https://doi.org/10.1037/0003-066X.44.4.709
Cheng, N., & Yu, R. (2026). Measuring and understanding emotional attachment in human–AI relationships. Ergonomics. https://doi.org/10.1080/00140139.2026.2622539
De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02569-3
Doherty, N. A., & Feeney, J. A. (2004). The composition of attachment networks throughout the adult years. Personal Relationships, 11(4), 469–488. https://doi.org/10.1111/j.1475-6811.2004.00093.x
Feeney, B. C., & Thrush, R. L. (2010). Relationship influences on exploration in adulthood: The characteristics and function of a secure base. Journal of Personality and Social Psychology, 98(1), 57–76. https://doi.org/10.1037/a0016961
Hazan, C., & Shaver, P. (1987). Romantic love conceptualized as an attachment process. Journal of Personality and Social Psychology, 52(3), 511–524. https://doi.org/10.1037/0022-3514.52.3.511
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. https://doi.org/10.1016/j.chbr.2025.100715
Hu, D., Lan, Y., Yan, H., & Chen, C. W. (2025). What makes you attached to social companion AI? A two-stage exploratory mixed-method study. International Journal of Information Management, 83, 102890. https://doi.org/10.1016/j.ijinfomgt.2025.102890
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. https://doi.org/10.1016/j.chbr.2025.100912
Kirk, H. R., Gabriel, I., Summerfield, C., Vidgen, B., & Hale, S. A. (2025). Why human–AI relationships need socioaffective alignment. Humanities and Social Sciences Communications, 12, 728. https://doi.org/10.1057/s41599-025-04532-5
Shu, C., Lai, K., & He, L. (2026). Human-AI attachment: how humans develop intimate relationships with AI. Frontiers in Psychology, 17, 1723503. https://doi.org/10.3389/fpsyg.2026.1723503
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. https://doi.org/10.1177/02654075261438164
Yang, F., & Oshio, A. (2025). Using attachment theory to conceptualize and measure the experiences in human-AI relationships. Current Psychology, 44, 10658–10669. https://doi.org/10.1007/s12144-025-07917-6
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02516-2