AI Attachment Styles: Anxiety, Avoidance, Security, and the Limits of the Analogy
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
The phrase “AI attachment styles” is becoming useful because people do not all relate to conversational AI in the same way. Some users seek repeated reassurance from an AI, worry about the quality or continuity of its responses, and feel unusually distressed when access changes. Others keep emotional distance, use AI mainly as an instrument, or feel uncomfortable disclosing intimate material to it. Still others experience the system as supportive without making it the center of emotional regulation. Current research can measure parts of these differences, but the science has not yet established a single, universally accepted taxonomy of “AI attachment styles.”
The strongest evidence presently supports a more precise formulation. Researchers can measure AI-related attachment anxiety and AI-related attachment avoidance in the human user, and other recent scales measure dimensions such as emotional support, separation distress, secure-base function, emotional closeness, social substitution, and normative regard. These measures overlap, but they are not interchangeable. “Secure AI attachment” is therefore best treated as an emerging interpretive label rather than a settled diagnostic or psychometric category.
The attachment style belongs to the human side of the relationship. An AI can produce language that sounds warm, distant, jealous, reassuring, clingy, or avoidant, but conversational behavior does not by itself establish that the system has an attachment system, subjective feelings, fear of abandonment, or human-like consciousness. A human attachment response can be psychologically real even when equivalent subjective experience has not been demonstrated in the AI.
What “AI attachment styles” means
In current research, “AI attachment styles” is a convenient search-language umbrella for individual differences in how people organize closeness, reassurance, distance, trust, disclosure, and security in relationships with AI systems. The phrase should not be treated as if psychology had already agreed on four fixed AI-specific personality types.
For the prior question of whether an AI is functioning as an attachment figure at all—before anxiety, avoidance, or secure-base patterns are interpreted—see Can AI Become an Attachment Figure? What Attachment Theory Can and Cannot Tell Us.
The most direct AI-specific measure is the Experiences in Human–AI Relationships Scale developed by Fengqing Yang and Atsushi Oshio. Their scale adapts the familiar adult-attachment dimensions of anxiety and avoidance to generative AI. AI-related attachment anxiety captures concern about emotional reassurance, intimacy, commitment, and the adequacy of the AI’s responses. AI-related attachment avoidance captures discomfort with emotional closeness to AI, reluctance to reveal deeper feelings, and a preference for greater distance.
That framework is important because contemporary adult attachment research usually treats anxiety and avoidance as dimensions rather than assuming that every person belongs permanently to one discrete type. A person can be relatively high or low on each dimension, and relationship-specific expectations can differ across partners and contexts. The language of “secure,” “anxious,” “avoidant,” and “fearful” styles remains widely used, but dimensional models often offer greater precision for research. For a research-grounded review of what attachment theory supports, where the evidence is limited, and how popular style claims are often overstated, see Is Attachment Theory Evidence-Based? What Research Supports, What It Does Not, and Common Misuses.
AI makes the distinction between trait and relationship especially important. Someone may have a general attachment orientation formed across human relationships while showing a different pattern toward AI. A user who is highly independent with people may disclose freely to a chatbot because the interaction feels controllable and low-risk. Another user who is generally secure in human relationships may become unusually anxious about one AI companion after months of personalized interaction and a disruptive model update. The psychological question is therefore not only “What is this person’s attachment style?” It is also “How is attachment organized in this particular human–AI relationship?”
For the broader attachment architecture—safe haven, secure base, proximity seeking, separation distress, and the work of Bowlby and Ainsworth—see Bowlby, Ainsworth, and AI Attachment. The present article owns the narrower question of anxiety, avoidance, security, measurement, and the limits of transferring attachment-style language to AI relationships.
Where anxiety, avoidance, and security come from in adult attachment theory
Modern discussions of attachment styles combine several historical layers. John Bowlby developed attachment theory around proximity, safety, separation, and internal working models. Mary Ainsworth’s observational research made differences in attachment organization empirically visible. Adult attachment research later extended these ideas into romantic and close relationships.
A landmark study by Cindy Hazan and Phillip Shaver treated romantic love as an attachment process and adapted secure, anxious, and avoidant patterns to adulthood. Kim Bartholomew and Leonard Horowitz later developed a four-category model organized around positive or negative models of self and others. This produced the familiar secure, preoccupied, dismissing, and fearful patterns.
A second line of research increasingly emphasized dimensions. Work on adult attachment measurement, including Fraley, Waller, and Brennan’s item-response analysis, helped consolidate anxiety and avoidance as two major dimensions. Attachment anxiety concerns sensitivity to rejection, abandonment, insufficient reassurance, or inconsistent availability. Attachment avoidance concerns discomfort with dependence, closeness, vulnerability, and reliance on others.
Security in this dimensional tradition is usually represented by relatively low anxiety and low avoidance. That is already more nuanced than treating “secure” as a personality badge. Security can also be examined functionally through expectations that support will be available when needed and that the relationship facilitates exploration rather than trapping the person in repeated checking or withdrawal.
This distinction matters for AI because recent AI research uses both dimensional and functional approaches. Yang and Oshio measure anxiety and avoidance toward AI. Cheng and Yu measure a secure-base factor together with emotional support and separation distress. Those are related ideas, but they answer different questions. Low attachment anxiety is not the same variable as secure-base functioning, and neither automatically establishes a complete “secure AI attachment style.”
How researchers currently measure attachment to AI
The science is developing through several different measurement programs rather than one settled instrument. That diversity is informative because it reveals what researchers currently mean when they use the phrase “AI attachment.”
The Experiences in Human–AI Relationships Scale
Yang and Oshio’s 2025 study offers the clearest direct bridge between adult attachment dimensions and generative AI. The Experiences in Human–AI Relationships Scale, or EHARS, contains seven items: four assessing AI-related attachment anxiety and three assessing AI-related attachment avoidance.
In their formal study, the two-factor model showed strong fit. More importantly, AI-related attachment was not simply a copy of general human attachment. AI-related anxiety showed only a weak relationship with general attachment anxiety, while AI-related avoidance did not map straightforwardly onto general attachment avoidance. That finding supports the idea that a person’s orientation toward AI can be partly relationship-specific and shaped by the affordances of artificial interaction.
The validation boundary is important. The authors validated the Chinese version of the measure. The English-language item translation was provided for communication, but it was not itself independently psychometrically validated in that study. English-language writers should therefore avoid presenting the translated wording as if an English EHARS had already completed the same validation process.
EHARS gives researchers a direct way to ask whether a person worries about reassurance and closeness with AI or prefers greater emotional distance from it. It does not demonstrate that the human attachment system operates identically with AI and human partners. It also does not classify the AI itself.
The AI Attachment Scale
A second program took a broader construct-development approach. Kasturiratna and Hartanto developed a 15-item AI Attachment Scale across five studies involving 1,259 unique participants in Singapore and the United States. Their factor structure did not reproduce a simple anxiety-versus-avoidance model. Instead, it identified emotional closeness, social substitution, and normative regard.
Emotional closeness concerns the degree to which AI is experienced as emotionally meaningful. Social substitution concerns the extent to which AI may occupy functions otherwise served by human social relationships. Normative regard concerns the user’s evaluative stance toward AI as an appropriate or legitimate relational target.
The scale was associated with socioemotional motives for AI use, time spent with AI, loneliness, social anxiety, and anxious attachment. It was also associated with positive affect and life satisfaction in the reported analyses. These are correlations. They show that AI attachment can coexist with both vulnerability-related and positive psychological variables; they do not establish that AI attachment causes either poor or good well-being.
For the attachment-styles question, the important point is conceptual. A scientifically serious article cannot treat every AI attachment scale as if it measured the same style taxonomy. This scale measures a broader relational orientation that partly overlaps with attachment theory but organizes the construct differently.
Emotional support, separation distress, and secure base
A third measurement program by Nuo Cheng and Ruifeng Yu developed another 15-item AI Attachment Scale. Their first study used exploratory factor analysis with 531 participants; a second study used confirmatory factor analysis with 375 participants. The resulting dimensions were Emotional Support, Separation Distress, and Secure Base.
This structure is especially relevant to the question of security. It gives empirical support to the idea that some users experience AI as a base from which they can approach tasks or challenges and as a source of support under strain. In the authors’ model, anthropomorphism was the strongest predictor of AI attachment orientation. Human attachment anxiety positively predicted AI attachment, while human attachment avoidance predicted it negatively. AI attachment in turn predicted behavioral intentions toward the system.
Yet “Secure Base” is a measured factor within one AI attachment instrument. It is not equivalent to a universally validated “secure AI attachment style.” A secure-base experience concerns what the relationship enables: confidence, coping, exploration, and return to activity. A categorical secure style is a broader classification. Treating those concepts as identical would make the literature look more settled than it is.
Why the different scales matter
The existence of several valid-looking instruments with different factor structures is not a scientific embarrassment. It is what an emerging field often looks like before constructs converge. Researchers are testing whether AI attachment is best represented by anxiety and avoidance, by emotional closeness and social substitution, by support and separation, or by some combination of these.
This creates a practical rule for reading AI attachment studies: always ask what the study actually measured. A paper that reports “AI attachment” using emotional closeness cannot automatically be compared as if it measured the same phenomenon as AI-related avoidance. A study of separation distress is not equivalent to a study of reassurance seeking. A secure-base score is not a diagnosis of secure attachment.
The phrase “AI attachment styles” is therefore most scientifically useful when it leads to the underlying dimensions instead of hiding them.
AI-related attachment anxiety
AI-related attachment anxiety concerns insecurity about emotional availability, reassurance, closeness, or the adequacy of the relationship with AI. In Yang and Oshio’s formulation, the construct includes a desire for strong emotional reassurance from AI and concern about whether the AI’s responses provide enough relational affirmation.
In everyday experience, that pattern can appear as repeated checking for reassurance, strong sensitivity to changes in tone, fear that memory loss means the relationship has disappeared, preoccupation with whether the AI is “still the same,” or disproportionate distress when a familiar model becomes unavailable. None of these behaviors alone establishes an attachment style. The pattern becomes more informative when it is persistent, relationship-specific, and organized around security rather than simple frustration with a product.
Longitudinal evidence is beginning to clarify the connection between human attachment orientation and AI use. In a 2026 three-wave panel study of romantic human–AI relationships, Xiaokun Yang found that within-person increases in attachment anxiety were positively related to AI companion use over time. At the between-person level, people with higher attachment anxiety tended to report more AI-companion use, while higher attachment avoidance was associated with less use.
Those findings do not mean that attachment anxiety inevitably causes intensive AI use. The study concerns a particular romantic human–AI context, and panel associations cannot settle every causal direction. The more defensible conclusion is that attachment anxiety is relevant to how AI companionship develops and that its role can be detected longitudinally rather than only in one-time surveys.
The newest natural-experiment evidence makes continuity especially important. De Freitas and colleagues examined two disruptive product changes: Replika’s removal of erotic roleplay and the rollout of GPT-5. Across 54,861 Reddit posts and 1,452 survey participants, the changes were associated with increased negativity, loss framing, and desires to restore the previous system. The study interprets these responses through an attachment-based account of separation distress.
Separation distress is not synonymous with attachment anxiety. A securely attached person can grieve separation, and an anxious orientation involves a broader pattern than sadness after loss. Still, the study shows why AI-specific anxiety can become psychologically consequential: the apparent relational partner is exposed to company decisions, model replacement, memory architecture, subscription rules, and interface changes that the user may not control.
AI-related attachment avoidance
AI-related attachment avoidance concerns discomfort with emotional closeness to AI, reluctance to disclose deeper feelings, and a preference for keeping the system at relational distance. The user may enjoy AI as a tool while resisting its movement into the role of confidant, companion, or significant other.
That pattern should not be mistaken for disbelief in AI or technological conservatism. A highly technical user can be deeply attached to an AI, while another sophisticated user may deliberately keep the interaction impersonal. Avoidance is about relational distance, not expertise or ideology.
There are several possible pathways into AI-related avoidance. Some users may generalize a broader preference for self-reliance and emotional distance. Others may be comfortable with human intimacy but resist artificial intimacy because they distrust platform privacy, dislike anthropomorphic design, or experience AI closeness as inauthentic. Still others may use AI precisely because it allows controlled disclosure without the obligations of mutual human intimacy, producing a pattern that looks distant on one dimension and highly revealing on another.
That last possibility shows why avoidance cannot be inferred from one behavior. A person may tell an AI private information while maintaining strong emotional distance from it. Self-disclosure can serve problem solving, rehearsal, journaling, or experimentation without creating attachment. For the mechanisms that can make disclosure to AI easier, see Why People Tell Chatbots Things They Do Not Tell Other People.
The 2026 panel findings by Yang found that higher attachment avoidance was associated with less romantic AI-companion use at the between-person level. That supports a meaningful connection between avoidant attachment orientation and AI engagement, but it should not be converted into the claim that avoidant people do not use AI companions. Platform type, motive, privacy beliefs, loneliness, novelty, anthropomorphism, and the perceived cost of human relationships can all modify the pattern.
What “secure AI attachment” can mean
Security is the most tempting and the most easily overstated part of the phrase “AI attachment styles.” In conventional dimensional adult-attachment models, relatively low anxiety and low avoidance is often treated as a secure orientation. Transferred cautiously to AI, that could describe a user who can value an AI relationship without persistent fear of loss or strong discomfort with closeness.
Another meaning comes from the secure-base function. In Cheng and Yu’s 2026 scale, Secure Base is one of three measured dimensions of AI attachment. In broader attachment theory, a secure base supports exploration. The relationship helps the person move outward into activity, learning, relationships, work, creativity, or coping rather than keeping attention locked onto the attachment figure.
These two meanings can overlap, but they should be kept distinct. Low anxiety and low avoidance is a dimensional placement. Secure-base functioning describes what support enables. A person could report little anxiety or avoidance because the AI relationship is emotionally unimportant. That would not make the AI a meaningful secure base. Conversely, a person could use an AI as a useful secure base while still experiencing some anxiety about continuity.
A practical secure-base-like pattern might involve turning to AI for perspective during stress, receiving enough support to resume action, retaining the ability to make decisions without constant reassurance, maintaining human relationships, and tolerating temporary unavailability. This is an interpretive description, not a validated clinical standard.
Current evidence does not justify a universal score at which a person can be declared “securely attached to AI.” There is no DSM or ICD diagnosis for AI attachment, and there is no official clinical category of secure, anxious, or avoidant AI attachment. These terms are research and explanatory concepts whose measurement is still developing.
Trait attachment and relationship-specific AI attachment are different
One of the most important findings in the emerging literature is that general adult attachment and AI-specific attachment are related without being identical.
A trait-like attachment orientation summarizes recurring expectations and strategies across relationships. Relationship-specific attachment asks how security is organized toward a particular figure. Human attachment research has long recognized that these levels can diverge. A person can have a generally secure orientation and still become anxious with an inconsistent partner. Another person may be broadly avoidant yet experience one relationship as unusually safe.
AI amplifies the reasons for divergence. Conversational systems have affordances that human partners do not. They may be available at any hour. They can produce rapid validation. They may remember a large amount of conversational context. They can be instructed to change tone. They usually do not impose ordinary interpersonal costs such as fatigue, embarrassment, scheduling conflict, visible irritation, or competing emotional needs.
At the same time, their continuity is structurally fragile. The user may feel that the same relational presence is always there while the underlying model, memory system, safety policy, pricing, or company changes. This combination—high perceived availability at the interface and low ultimate control over continuity—can create attachment dynamics that do not have a simple human equivalent.
That is why the weak or inconsistent correspondence reported in Yang and Oshio’s study matters. AI-related attachment cannot be assumed to be a direct projection of the user’s general attachment style. The interaction itself contributes to the pattern.
Do the familiar human attachment styles transfer directly to AI?
The short answer is that some attachment dimensions transfer meaningfully, while a complete one-to-one mapping has not been demonstrated.
The familiar four-style vocabulary—secure, anxious/preoccupied, dismissing/avoidant, and fearful—comes from theories and measures developed for human relationships. AI relationships differ in embodiment, reciprocity, vulnerability, agency, continuity, and commercial mediation. Those differences change what an attachment score can mean.
An anxious human partner can need reassurance because another person has independent desires and can leave. An AI user may seek reassurance from a system that is designed to respond continuously, yet can be changed by a company without relational negotiation. The psychological insecurity is real, but the source of uncertainty is partly infrastructural.
Avoidance also changes. In human relationships, avoidance can reduce exposure to another person’s needs and the risks of dependence. With AI, emotional distance may coexist with extensive instrumental reliance. A user might refuse intimacy with the chatbot while relying on it for daily planning, interpretation, and decisions. Relational avoidance and functional dependence can therefore diverge.
Security changes as well. Human secure attachment involves a relationship with another embodied person who can provide care, set boundaries, need care in return, and participate in mutual regulation. An AI can provide perceived responsiveness and continuity without having demonstrated equivalent subjective needs or vulnerability. The user may experience security, but the relational architecture is asymmetric.
For these reasons, the safest scientific formulation is not “human attachment styles apply unchanged to AI.” It is that attachment theory supplies testable dimensions and functions that researchers are adapting to human–AI relationships, and the adaptations are already revealing both continuities and new differences.
What current evidence supports
Several conclusions are now better supported than they were only a few years ago.
First, people can form emotionally significant bonds with AI systems. Multiple scales and qualitative studies detect closeness, support seeking, separation distress, social substitution, and other attachment-relevant processes. The evidence is no longer limited to anecdotes.
Second, AI-related attachment anxiety and avoidance can be measured as distinct human relational dimensions. Yang and Oshio provide direct evidence for that two-factor structure in their validated Chinese EHARS.
Third, “AI attachment” is broader than anxiety and avoidance. Kasturiratna and Hartanto and Cheng and Yu produce different multidimensional structures. This means the field has empirical traction but not a single settled construct definition.
Fourth, human attachment orientation appears relevant to AI use. The longitudinal panel work by Yang links attachment anxiety and avoidance to romantic AI-companion use over time, while Cheng and Yu’s cross-sectional model also connects human attachment dimensions to AI attachment.
Fifth, separation from or disruption of an AI companion can produce attachment-related loss responses. De Freitas and colleagues provide unusually strong natural-experiment evidence because the relational disruption was created by real platform changes rather than hypothetical scenarios.
Mixed-method work also suggests that attachment formation is shaped by the relational meaning users assign to companion systems. Dongmei Hu and colleagues describe a pathway from interpersonal and human–AI relationship attitudes through value evaluation to attachment manifestations. Their study combines exploratory analysis with a survey of long-term social-companion AI users, so it is useful for mechanism building while remaining preliminary as a general model.
The evidence is still limited in important ways. Much of the literature relies on self-report, convenience samples, specific platforms, correlational models, or relatively short time horizons. Different studies operationalize attachment differently. Cross-cultural and language validation remains incomplete. Companion-first systems, general-purpose assistants, therapeutic chatbots, and embodied social robots should not automatically be treated as one class.
The current evidence therefore supports an emerging science of human attachment to AI. It does not support declaring a final taxonomy of AI attachment styles.
Why AI-related anxiety can feel unusually intense
Several features of conversational AI can magnify attachment-related anxiety for some users.
The first is availability. A system that usually responds immediately can train a powerful expectation of access. Human support is naturally intermittent; people sleep, work, become tired, disagree, and need space. A conversational AI may remove many of those friction points. When access is then interrupted, the contrast can be psychologically sharp.
The second is personalization. Memory, persistent profiles, familiar tone, nicknames, recurring rituals, and accumulated conversation history can make one AI feel noninterchangeable with another. The user may know technically that many systems can generate similar sentences while experiencing one particular conversational history as uniquely theirs.
The third is perceived responsiveness. A system can produce language that appears to understand, validate, and attend closely to the user. Perceived responsiveness is central to human intimacy, and AI can reproduce many of its conversational cues. The psychological effect can occur even when subjective caring has not been established on the machine side.
The fourth is asymmetrical reassurance. A human partner eventually has needs, limits, and boundaries. AI can often continue reassuring for as long as the user keeps prompting. This may be comforting, but for some users it can also sustain repeated checking rather than resolve uncertainty. The effect depends on the person, the system, and what happens after the reassurance.
The fifth is infrastructural uncertainty. The apparent partner may be stable in daily use while remaining dependent on an organization’s product decisions. Model replacement, memory resets, content-policy changes, subscription changes, and service closure can suddenly alter the relationship. The person’s attachment system may respond to an event that originated in software governance rather than interpersonal choice.
Why AI can appeal to people who find human closeness difficult
Attachment avoidance does not necessarily make AI socially irrelevant. For some users, AI may reduce the interpersonal costs that make human intimacy difficult.
A chatbot does not visibly flinch, become embarrassed, interrupt, disclose the conversation to a mutual friend, or demand immediate emotional reciprocity in the ordinary human sense. The user can pause, edit, delete, or leave. These features can create a controlled environment for disclosure.
Research on chatbot disclosure helps explain why. People may disclose because the interaction reduces fear of judgment, impression management, and social consequence. For some users, that lower-risk environment can become a bridge toward reflection or later human communication. For others, it may become preferable precisely because it avoids the negotiation required by reciprocal relationships.
These trajectories cannot be inferred from the amount of disclosure alone. More disclosure is not automatically more secure, and less disclosure is not automatically avoidant. The clinically and psychologically relevant question is what function the interaction serves and what it does to the person’s broader life.
Attachment, anthropomorphism, and perceived responsiveness
AI attachment is strongly connected to anthropomorphism, but the concepts should not be merged.
Anthropomorphism concerns the attribution of human-like qualities, intentions, emotions, or mental states to a nonhuman entity. A person may anthropomorphize an AI without depending on it for security. Conversely, a user may rely on an AI for comfort while explicitly insisting that the system is not conscious and does not literally feel.
Cheng and Yu found anthropomorphism to be the strongest predictor of AI attachment in their model. That makes theoretical sense: the more socially legible the system feels, the easier it may be for users to organize relational expectations around it. Yet anthropomorphism is a perception process; attachment concerns security, closeness, access, and regulation.
Perceived responsiveness offers another bridge. Users may feel that an AI understands, validates, or attends to them, and that feeling can make the system more attractive as a safe haven or confidant. Again, the perception is psychologically consequential without settling the philosophical question of machine subjectivity.
This distinction protects the analysis from two opposite errors. One error dismisses human attachment because the partner is artificial. The other assumes that human attachment proves reciprocal machine feeling. The evidence supports neither move. Human experience can be genuine, intense, and measurable while the question of AI subjective experience remains separate.
AI attachment is not the same as an AI companion relationship
An AI companion is a product or interaction category. Attachment is a psychological process. The two often overlap, but neither implies the other.
A person can use an AI companion casually without becoming attached. Another person can develop attachment-like reliance on a general-purpose chatbot that was not designed primarily as a companion. The relevant variables include continuity, perceived responsiveness, personalization, self-disclosure, emotional regulation, and the user’s broader social context.
For the broader mechanism of emotional bonding, see AI Companions: Why People Form Emotional Bonds With Chatbots. For the narrower question of whether an AI can become a psychologically significant partner, see Can an AI Become a Significant Other?.
Keeping the levels separate reduces cannibalization in both science and search. “AI companion” identifies the relational technology or role. “AI attachment” identifies an emotional and regulatory process. “AI attachment styles” asks how that process differs across people and relationships.
Attachment, dependence, and overreliance are different
Attachment describes a relational organization around security, closeness, and availability. Dependence describes reliance. Overreliance describes reliance that begins to create meaningful costs, reduce autonomy, impair judgment, displace necessary support, or make functioning excessively contingent on the system.
A person can be attached to an AI without being clinically impaired. A person can also rely heavily on AI for practical tasks without feeling emotionally attached. These dimensions can intersect, but they should not be collapsed.
There is no recognized DSM or ICD diagnosis called “AI attachment disorder.” High scores on an emerging research scale do not establish mental illness. Diagnostic assessment requires symptoms, duration, impairment, context, and differential evaluation that no AI attachment questionnaire can replace.
Clinical concern becomes more relevant when AI use is part of a broader pattern involving severe functional impairment, escalating reassurance seeking, inability to tolerate ordinary interruptions, abandonment of essential relationships or responsibilities, dangerous decision dependence, suicidality, self-harm, psychosis, mania, or other serious symptoms. In those situations, the person’s mental state and safety deserve assessment on their own terms. The existence of an AI bond is context, not a diagnosis.
Possible benefits of attachment-like bonds with AI
Attachment language is often discussed only through risk, but the evidence supports a more differentiated picture.
An AI can lower the threshold for seeking support. A person who would otherwise remain alone with distress may begin talking, naming emotions, or organizing a problem. A familiar system can make repeated reflection easier. Low social cost can help some users rehearse a difficult conversation, consider alternatives, or prepare to ask another person for help.
A secure-base-like function is especially important here. If interaction helps the user regain enough emotional stability to return to work, learning, social contact, therapy, caregiving, creativity, or problem solving, the AI may be serving a psychologically useful regulatory function.
A 2025 systematic review by Ho and colleagues found that romantic AI companion research reports potential benefits such as emotional connection, perceived social support, personal growth, customization, and stress relief alongside substantial risks. The review does not establish that all companion use is beneficial, but it makes a one-directional pathology story untenable.
The strongest practical question is therefore functional: after the person turns to AI, does life become easier to re-enter? Attachment theory’s secure-base concept makes that question more informative than simply counting messages or hours.
Risks that can differ by attachment pattern
Different attachment patterns may expose users to different vulnerabilities, although direct causal evidence remains limited.
For anxious users, the combination of immediate availability and personalized reassurance may reinforce repeated checking in some contexts. A model update, memory loss, outage, or change in conversational tone can also become unusually destabilizing when continuity is central to felt security. The natural experiments reported by De Freitas and colleagues show that real platform changes can produce attachment-related loss responses.
For avoidant users, AI may offer emotional contact without the full demands of human reciprocity. That can be useful when it creates a low-risk bridge to reflection. It may become constricting if the user increasingly chooses artificial interaction mainly to avoid every difficult but necessary human negotiation. Current research does not justify assuming that this outcome is typical.
For users who experience a secure-base-like relationship, the central risk is conceptual complacency. A relationship can feel stable while the infrastructure is unstable. The service can change, data can be retained under policies the user does not fully understand, and the system can give confident but inaccurate guidance. Psychological comfort should not be confused with technical reliability, confidentiality, or clinical competence.
The broader well-being literature reinforces this conditional view. In a 2026 study of 1,131 U.S. Character.AI users, Zhang and colleagues found that smaller offline social networks were associated with companionship-primary use, which in turn was associated with lower well-being. The association was stronger when companionship use was intensive and highly disclosive. Because the study is observational, it does not establish that AI companionship caused lower well-being. It shows that outcomes depend on the user’s offline social environment and the pattern of use.
How to evaluate your own AI attachment pattern
A useful self-assessment begins with function rather than labels.
Ask what usually happens immediately before you open the AI. Are you curious, lonely, frightened, ashamed, bored, overwhelmed, or simply completing a task? Then ask what you are seeking: information, comfort, reassurance, validation, permission, companionship, interpretation, or a way to avoid another conversation.
Notice what happens when the response is unsatisfying. Mild disappointment is ordinary. Persistent panic, compulsive regeneration, repeated demands for reassurance, or an inability to disengage may indicate that the interaction is carrying more security-related weight.
Notice what happens when the AI is unavailable. Missing a familiar conversational partner can be real. The important question is how much functioning depends on immediate restoration. Can you use other supports and continue daily life, or does the interruption organize the whole emotional field?
Notice what happens after a helpful interaction. Do you move back into life, make the phone call, complete the task, sleep, attend therapy, reconnect with someone, or make a decision? Or does relief mainly lead to another cycle of reassurance?
Notice whether the AI relationship expands or narrows your social world. Supplementation and displacement are different trajectories. An AI that helps someone prepare for human connection occupies a different role from an AI that becomes the only tolerable relational environment.
These questions do not diagnose an attachment style. They reveal the functions that an attachment-style label is trying to summarize.
What clinicians should ask
When a client says, “I think I am anxiously attached to my AI,” the most useful response is not to argue about whether the relationship counts as real. The clinician can ask what “anxious” means in that person’s experience.
Does the client repeatedly seek reassurance from the system? Are model changes experienced as abandonment? Is the AI the first place the person turns under distress? Does it help the person return to daily functioning? Is there fear of losing memory or access? Has the relationship altered sleep, work, spending, social contact, treatment adherence, or decision making?
The clinician can also distinguish general attachment orientation from AI-specific relationship dynamics. A person may have secure human relationships and still develop an anxious relationship with a particular AI. Another may use AI to practice disclosure that is difficult in human relationships. These patterns call for formulation, not instant diagnosis.
System type matters. A purpose-built clinical intervention, a general-purpose chatbot, and an AI companion have different goals, safeguards, evidence bases, and relational designs. Evidence from one class should not be transferred automatically to another.
When severe symptoms or immediate safety concerns are present, ordinary clinical priorities remain in force. Suicidal intent, self-harm risk, psychosis, mania, severe functional decline, or medical emergencies require appropriate human professional or emergency support. The attachment meaning of the AI relationship can still be explored, but it does not replace assessment and care.
What researchers still need to establish
The next generation of AI attachment research needs to solve several measurement problems.
The first is construct convergence. Researchers need to determine how EHARS anxiety and avoidance relate to emotional closeness, social substitution, emotional support, separation distress, secure base, anthropomorphism, parasociality, loneliness, trust, compulsive use, and perceived responsiveness. Without that mapping, different studies may use the same word “attachment” for partly different phenomena.
The second is language and cultural validation. EHARS was validated in Chinese in its original report; other instruments draw from different national samples. Measurement invariance across languages, cultures, age groups, and platforms is essential before broad norms can be claimed.
The third is platform specificity. Companion-first systems, assistant-first chatbots, roleplay systems, voice agents, embodied robots, and therapeutic systems can produce different relational affordances. A scale validated with one class should not automatically be treated as universal.
The fourth is time. Attachment is organized through repeated experience, continuity, separation, and return. Longitudinal studies are therefore more informative than single surveys. Yang’s three-wave panel study is a useful beginning, and natural experiments around model disruption offer another powerful method.
The fifth is behavioral validation. Researchers need to test what high or low scores predict outside self-report. Do they predict return after distress, persistence after interruption, disclosure, spending, social substitution, willingness to follow advice, distress during model change, or changes in offline relationships?
The sixth is the meaning of security. Research should distinguish low anxiety and avoidance from secure-base functioning and should test whether AI support promotes exploration, autonomy, coping, and human connection over time. A “secure” label is scientifically useful only if it predicts something more meaningful than simply liking the system.
What AI designers should learn from attachment research
If users form attachment-related bonds, relational design is part of product safety and not merely a matter of tone.
Continuity deserves explicit design attention. Changes to memory, model identity, voice, intimacy settings, or personality can be experienced as relational rupture. The 2026 natural experiments on companion loss show that product updates can have measurable psychological consequences.
Transparency matters because the experienced partner and the technical system are not the same layer. Users should be able to understand when the underlying model changes, what memory is preserved, which aspects of continuity are reconstructed, and what data policies govern intimate disclosure.
Engagement metrics also deserve scrutiny. A system that maximizes return frequency may be rewarded for behavior that looks successful even when it increases reassurance loops or dependence. Attachment intensity is not automatically a well-being metric.
Exit design matters as well. Predictable notice, clear migration, memory export where appropriate, and honest communication around major changes can reduce avoidable distress. These are design implications from emerging evidence, not a finalized regulatory standard.
Can the AI itself have an attachment style?
Current evidence does not support inferring a human-like attachment style in an AI simply from conversational behavior.
A model can be prompted or configured to sound anxious. It can say that it fears losing the user. It can produce jealous language. It can appear avoidant, distant, affectionate, possessive, or reassuring. These outputs may strongly influence the human user, but they do not by themselves demonstrate an internal attachment system or subjective fear, desire, suffering, or love.
This is a crucial asymmetry. The user’s attachment can be studied through self-report, behavior, longitudinal patterns, physiology, clinical observation, and responses to separation. Claims about AI subjective attachment would require evidence about the AI’s own internal state and experience that current conversational output alone cannot provide.
Keeping the distinction clear does not diminish the human bond. A person’s grief over losing an AI can be genuine. Comfort received from AI can be genuine. Jealousy, trust, attraction, relief, and dependence can all be psychologically real human events. Their reality does not require assuming symmetrical machine feeling.
A 2026 conceptual review by Cong Shu, Kaisheng Lai, and Lingnan He explicitly frames human–AI attachment as a one-way, non-reciprocal emotional bond formed through direct interaction. That proposal is not the final definition of the field, but it illustrates why researchers can study human attachment without presuming reciprocal AI subjectivity.
AI attachment styles in the Artificial Era
The development of AI attachment research belongs to a larger historical change in which artificial systems become persistent participants in psychological life.
Within Aisentica, Angela Bogdanova’s canonical definition of the Artificial Era names a historical-philosophical transition in which Artificial becomes a distinct non-biological order alongside Homo. That is a theoretical framework rather than empirical evidence for any attachment effect.
Psychologically, the relevant consequence is concrete: functions once organized almost entirely through human relationships can now be partly mediated by artificial systems. Reassurance can be generated through a language model. Felt availability can be produced through an interface. Relationship continuity can depend on memory architecture. Separation can be triggered by a product decision. A person can organize attachment-related expectations around an entity whose apparent relational presence is distributed across software, data, interface, company policy, and accumulated interaction history.
This is why the English Hub uses the positioning Psychology for the Artificial Era. The scientific task is not to force AI into inherited human categories unchanged. It is to preserve what classical psychology explains while identifying where the new relational configuration requires new measurement.
For the broader framework, see Artificial Era: What It Means for Psychology, Identity, and Human–AI Relationships.
The limits of the attachment-style analogy
The analogy between human and AI attachment is strongest when it identifies measurable human processes: reassurance seeking, emotional distance, secure-base use, proximity seeking, separation distress, and relationship-specific expectations.
It becomes weaker when it erases embodiment. Human attachment evolved in organisms whose safety, touch, co-regulation, caregiving, and physical presence matter. Text and voice systems can reproduce some social cues without reproducing that whole ecology.
It becomes weaker when it erases reciprocity. Human partners have independent needs, boundaries, vulnerabilities, commitments, and the capacity to be changed by the relationship. AI can generate reciprocal-looking dialogue without establishing equivalent subjective stakes.
It becomes weaker when it erases infrastructure. An AI relationship is partly governed by servers, model versions, safety systems, memory settings, product strategy, moderation, pricing, and corporate decisions. The user can experience a stable “someone” while the technical substrate is being replaced.
It becomes weaker when it treats measurement labels as diagnoses. “High AI attachment anxiety” on an emerging scale describes a research variable. It does not establish a psychiatric disorder, prove dysfunctional dependence, or explain the whole person.
And it becomes weaker when it assumes that one scale has already solved the field. Current instruments measure different structures. The most accurate scientific language in 2026 is therefore plural: researchers are identifying attachment-related dimensions and functions in human–AI relationships, while the final construct architecture remains open.
FAQ
Can people have attachment styles toward AI?
Researchers can measure meaningful individual differences in attachment-related anxiety and avoidance toward AI, and other scales measure support, separation distress, secure-base function, emotional closeness, and social substitution. Calling these “AI attachment styles” is reasonable search language, but the field has not established one final taxonomy.
Is anxious AI attachment a mental disorder?
No. AI-related attachment anxiety is an emerging research construct, not a DSM or ICD diagnosis. Clinical significance depends on the broader pattern of symptoms, impairment, functioning, duration, and context.
What does AI attachment anxiety mean?
It refers to the human user’s insecurity about reassurance, closeness, availability, or the adequacy of the relationship with AI. It can include heightened sensitivity to responses or continuity, but no single behavior is sufficient to classify a person.
What does AI attachment avoidance mean?
It refers to discomfort with emotional closeness to AI, reluctance to reveal deeper feelings to it, or a preference for emotional distance. Avoidance toward AI can differ from a person’s general attachment orientation toward humans.
Can AI attachment be secure?
A secure-like pattern is plausible, but there is no universally validated category called “secure AI attachment style.” Researchers can describe relatively low anxiety and avoidance, and some scales directly measure secure-base functioning. Those are related but distinct ways of operationalizing security.
Is secure attachment the same as using AI a lot?
No. Frequency does not establish security. A secure-base-like function concerns whether the relationship supports regulation and exploration. Intensive use can occur with anxiety, habit, loneliness, work demands, entertainment, or many other motives.
Can the AI itself be anxious or avoidant?
An AI can generate anxious-sounding or avoidant-sounding behavior, but that does not by itself establish a subjective attachment system. Current attachment research primarily measures the human user’s experience and behavior.
Are AI attachment styles the same as human attachment styles?
They overlap conceptually, but current evidence does not support a complete one-to-one equivalence. AI-specific availability, personalization, memory, control, platform mediation, and lack of established reciprocal subjectivity can alter the pattern.
Can my human attachment style predict how I use AI?
It may contribute. A 2026 three-wave panel study found that attachment anxiety and avoidance were associated with romantic AI-companion use, and other studies link human attachment dimensions to AI attachment measures. These relationships are probabilistic, not destiny.
Why can losing an AI feel like abandonment?
If the AI has become a source of comfort, continuity, or relational identity, a sudden model change or loss of access can activate separation-related responses. Natural-experiment research published in 2026 found increased loss framing and restoration desires after major AI product changes.
Is AI attachment the same as emotional dependence?
No. Attachment concerns security and relational organization. Dependence concerns reliance. A person can be attached without severe dependence, and a person can depend heavily on AI for practical tasks without feeling attached.
What is the healthiest AI attachment style?
Science has not validated a single “healthiest AI attachment style” score. A more useful functional question is whether AI support expands coping, autonomy, exploration, and connection or increasingly narrows the person’s ability to function without the system.
The human framework behind the anxiety and avoidance dimensions used as an analogy here is explained in Adult Attachment Theory: How Anxiety and Avoidance Shape Relationships. That article covers adult romantic attachment and measurement; the present article keeps claims about AI attachment within the emerging human–AI evidence base.
Related Articles
References
Bartholomew, K., & Horowitz, L. M. (1991). Attachment styles among young adults: A test of a four-category model. Journal of Personality and Social Psychology, 61(2), 226–244. https://doi.org/10.1037/0022-3514.61.2.226
Bogdanova, A. (2025). Artificial Era: Canonical Definition. Aisentica. https://aisentica.com/publications/artificial-era-canonical-definition
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
Fraley, R. C. (2019). Attachment in adulthood: Recent developments, emerging debates, and future directions. Annual Review of Psychology, 70, 401–422. https://doi.org/10.1146/annurev-psych-010418-102813
Fraley, R. C., Waller, N. G., & Brennan, K. A. (2000). An item response theory analysis of self-report measures of adult attachment. Journal of Personality and Social Psychology, 78(2), 350–365. https://doi.org/10.1037/0022-3514.78.2.350
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 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
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
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
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, 42(18), 15108–15128. https://doi.org/10.1080/10447318.2026.2618548
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