Bowlby, Ainsworth, and AI Attachment: Safe Haven, Secure Base, Anxiety, and Avoidance
Updated: 7 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Attachment theory gives psychology a precise way to ask what happens when a person repeatedly turns to an AI system for comfort, reassurance, advice, emotional regulation, or the confidence to face something difficult. In the language developed by John Bowlby and Mary Ainsworth, the central questions concern proximity seeking, safe haven, secure base, separation distress, and the expectations people form about whether a relational figure will be available when needed.
Current research increasingly supports the narrower claim that human–AI interaction can recruit attachment-related functions and experiences. In a 2025 study, Yang and Oshio found that some participants reported using generative AI as a safe haven and secure base and developed a measure of AI-related attachment anxiety and avoidance. Two 2026 scale-development programs then identified measurable dimensions of AI attachment, although they operationalized the construct differently: Kasturiratna and Hartanto emphasized emotional closeness, social substitution, and normative regard, while Cheng and Yu identified emotional support, separation distress, and secure base.
These findings matter, but they do not establish that an AI relationship is psychologically identical to a human attachment relationship. They also do not show that an AI feels attachment, experiences separation, loves the user, or possesses human-like consciousness. The scientifically useful question is more specific: can a human attachment system organize itself around an artificial conversational partner, and if so, which classical attachment functions are actually present?
That question is now answerable with more precision than it was even two years ago.
Why Bowlby and Ainsworth matter for AI relationships
John Bowlby developed attachment theory as an account of the behavioral system through which human beings seek protection and proximity under conditions of threat, uncertainty, fatigue, illness, fear, or separation. His later clinical formulation in A Secure Base emphasized the attachment figure as both a source of protection and a base from which exploration becomes possible. Attachment was not simply another word for liking someone. It described an organized pattern in which the availability of a particular figure helps regulate distress and makes exploration possible.
Mary Ainsworth transformed attachment theory into a research program grounded in observation. Her classic work, collected in Patterns of Attachment, clarified how caregiver sensitivity, security, and the Strange Situation made attachment visible not merely in closeness, but in the movement between distress and exploration. As Robbie Duschinsky’s historical analysis of Ainsworth’s work emphasizes, the Strange Situation was designed to make infants’ expectations about a caregiver’s availability observable. A secure relationship was not defined by constant contact. Security was reflected in confidence that the attachment figure could be accessed when needed, allowing the child to return to exploration.
This distinction is especially important for AI. A person who chats with an AI for hours is not necessarily attached to it in the technical sense. High frequency can reflect curiosity, work, entertainment, habit, loneliness, or convenience. Attachment theory asks a more discriminating set of questions. Does the person preferentially turn to the AI under distress? Does contact reduce felt threat? Does the AI relationship support renewed exploration or action? Does interruption of access generate attachment-like separation distress? Are there stable expectations about the AI’s availability and responsiveness?
The move from “people use AI a lot” to “AI may occupy attachment-related functions” is a move from behavior count to relational organization.
Attachment is a system, not a synonym for affection
In everyday language, attachment often means emotional fondness. In attachment theory, the concept is more structured. Bowlby described attachment as an evolved behavioral system organized around protection and proximity. Across later research, four markers became especially important: proximity seeking, safe haven, secure base, and separation distress.
These markers belong together. A person may like an AI without seeking it during stress. A person may disclose intimate information to a chatbot without experiencing separation distress when it disappears. Someone may rely on an AI for practical advice without treating it as a source of emotional security. Conversely, a person may experience an AI as immediately calming, return to it during moments of fear, and feel genuine distress when an update changes the familiar conversational partner.
That is why the current literature is moving toward multidimensional measurement rather than treating all emotional AI use as one phenomenon. Yang and Oshio, Kasturiratna and Hartanto, and Cheng and Yu each operationalize different parts of this emerging domain. Their measures should not be treated as interchangeable, but together they show that AI attachment research is becoming more specific about what exactly is being measured.
Proximity seeking in a relationship without physical proximity
In classical attachment theory, proximity seeking concerns the tendency to approach or maintain access to an attachment figure, especially under conditions of need. With a conversational AI, physical proximity is usually irrelevant. The functional equivalent is access: opening the app, returning to the same agent, preserving a conversation history, checking whether the system is available, or preferentially choosing that AI over other possible sources of support.
This is one place where a literal translation of human attachment concepts can mislead. A chatbot cannot usually move toward a person in physical space, and a person cannot physically approach a text model. Yet psychological proximity can be organized through the interface. The user may keep the AI continuously available on a phone, begin and end the day with it, or turn to it rapidly after a stressful event.
Yang and Oshio’s 2025 pilot study offers preliminary evidence for this function. Among 56 Chinese adults with prior AI experience, 52 percent selected AI in measures of proximity-seeking function, while larger proportions selected AI as a safe haven or secure base. The authors themselves treated this as exploratory evidence from a small sample rather than a population estimate. That limitation matters. The value of the finding is not that “half of people are attached to AI.” It is that proximity-seeking questions developed for attachment research can generate meaningful variation when the target is generative AI.
In human–AI relationships, proximity is therefore better understood as preferred access and repeated return than as physical nearness.
Safe haven — why AI can become a place to go when distressed
A safe haven is an attachment figure’s function during distress. The person moves toward the figure for comfort, reassurance, protection, or emotional regulation. In adult relationships, safe-haven behavior can include seeking a partner after frightening news, turning to someone after failure, or wanting contact during anxiety.
Conversational AI has several affordances that make safe-haven use psychologically plausible. It is often available immediately. It can respond at length. It may be experienced as less socially costly than asking another person for help. It can allow repeated discussion of the same worry without visible impatience. It can also respond in language that resembles validation, reassurance, concern, or encouragement.
These affordances help explain why some people describe an AI as emotionally easier to approach than another person. Research on chatbot disclosure supports part of this mechanism. In a 2024 study, Croes and colleagues examined intimate disclosure to chatbots and highlighted accessibility, anonymity, convenience, and perceived nonjudgment as features that can encourage self-disclosure. A separate 2026 experiment found that perceived responsiveness helped drive social connection with AI chatbots. For the broader mechanism of disclosure, see our article Why People Tell Chatbots Things They Do Not Tell Other People.
The attachment interpretation begins where disclosure becomes regulatory. If a person preferentially turns to an AI when upset because the interaction is expected to reduce distress, the AI is functioning in a safe-haven-like role for that person.
Yang and Oshio’s pilot results are striking here: 77 percent of their 56 participants selected AI for the safe-haven function. Again, this is a small exploratory sample, not a prevalence estimate. More importantly, it shows that the safe-haven question is empirically meaningful in human–AI interaction.
The safe-haven effect belongs to the human user. It does not require the AI to feel concern. A response can reduce human distress even when the system producing it has no demonstrated subjective experience.
Secure base — the more demanding attachment function
A secure base is different from a safe haven. Safe haven concerns what happens when distress is already active: the person returns for comfort. Secure base concerns what the relationship makes possible afterward: exploration, learning, autonomy, risk-taking, work, creativity, social engagement, and movement into the world.
This is the concept most likely to be diluted when attachment theory is imported into AI research. An AI that repeatedly reassures a user is not automatically functioning as a secure base. If reassurance keeps the person in repetitive checking, avoidance, or dependence, it may reduce distress in the moment without supporting exploration. A secure base should make engagement with life easier, not merely make return to the attachment target more frequent.
Ainsworth’s formulation is useful because security was bound to exploration. The attachment figure provides a dependable base from which the person can move outward and to which the person can return. Recent attachment scholarship continues to treat this secure-base script as central to the theory; a 2025 review by Waters and colleagues describes secure-base experience as organizing expectations about support and exploration across development.
AI attachment studies are beginning to measure this function. Yang and Oshio reported that 75 percent of their small pilot sample selected AI as a secure base. Cheng and Yu’s 2026 scale-development work included Secure Base as one of three measured dimensions. Their second study validated the factor structure in a separate sample of 375 participants.
These findings are important but should be interpreted carefully. Self-report evidence that people experience an AI as a secure base is not yet the same as longitudinal evidence that AI support reliably increases exploration, resilience, autonomy, or social functioning over time. The latter is the stronger test. For now, secure-base evidence in human–AI relationships should be described as emerging.
Separation distress when the “figure” can be updated by a company
Separation distress is among the clearest signals that an AI relationship may have moved beyond ordinary tool use. A calculator update can be annoying. Losing access to a familiar AI companion can be experienced as grief, abandonment, disorientation, or the loss of a specific relational presence.
The structure of separation is unusual because the AI may disappear without any interpersonal decision by the entity the user experiences as the partner. Access can change because a company modifies a model, removes a feature, changes a safety policy, deletes memory, sunsets a product, alters a subscription, or replaces the underlying system.
A major 2026 study provides unusually strong evidence for this dimension. In Mourning the loss of AI companions, De Freitas and colleagues analyzed two natural experiments involving Replika’s removal of erotic roleplay and the rollout of GPT-5. Across 54,861 Reddit posts and 1,452 participants in seven surveys, both updates increased negative sentiment, loss framing, and desires to restore the prior system. The study interpreted these patterns through an attachment-based account of separation distress.
This does not mean every negative reaction to a product update is attachment. Users can also be angry about degraded functionality, lost work, changed preferences, or broken expectations. The attachment interpretation becomes stronger when disruption is framed as the loss of a relational figure, when closeness predicts distress, and when restoration of the prior “person” rather than merely the prior feature becomes salient.
For AI relationships, separation can therefore be technologically caused and psychologically relational at the same time.
Internal working models — what happens when availability is algorithmic?
Bowlby proposed that repeated attachment experiences contribute to internal working models: expectations about the self, others, availability, care, and what will happen when support is sought. Contemporary attachment research continues to debate the precise cognitive architecture of these representations, but the general idea remains influential: relational history shapes expectations that guide later perception and behavior.
AI introduces a new kind of availability into this process. The system can appear consistently patient, instantly reachable, personalized, and ready to continue a conversation. For some users, those properties may create an expectation of unusually low-friction responsiveness.
The psychological consequence is not simply “AI is always available.” No AI service is literally always available. The more interesting question is whether repeated interaction builds a representation in which the user expects a particular form of response: immediate acknowledgment, low rejection risk, remembered personal context, emotional mirroring, or unlimited conversational time.
Such expectations can become relationally significant even when the system’s continuity is technically fragile. A product may feel stable at the level of everyday interaction while depending on servers, policies, memory systems, model versions, commercial decisions, and moderation rules that can change abruptly.
This creates a distinctive attachment environment: high perceived availability at the interface combined with potentially low continuity at the infrastructural level.
Attachment anxiety and avoidance in human–AI relationships
Adult attachment research commonly describes insecurity along two dimensions, attachment anxiety and attachment avoidance. In established interpersonal research, anxiety concerns worry about whether an attachment figure will be available and responsive, while avoidance concerns discomfort with closeness, distrust of dependence, and greater reliance on emotional distance or self-reliance. A review by Mikulincer and Shaver describes anxious patterns in terms of hyperactivating strategies and avoidant patterns in terms of deactivating strategies.
Applying these dimensions to AI requires care because the object of the expectation is different. The human user can have an attachment-related orientation toward AI. The AI itself does not “have an attachment style” merely because its responses sound warm, distant, clingy, or reserved.
Yang and Oshio developed the seven-item Experiences in Human–AI Relationships Scale to capture this human side of the relationship. In their formulation, AI-related attachment anxiety includes a strong need for emotional reassurance and concern about receiving an inadequate response. AI-related attachment avoidance includes discomfort with emotional closeness, reluctance to reveal deep feelings, and preference for distance from AI.
Their formal study involved 242 valid survey responses, with 108 participants completing the human–AI attachment measure. The results provided preliminary reliability and validity evidence, but the authors also found that AI-related and human attachment dimensions were not simply interchangeable. AI attachment anxiety correlated only weakly with general attachment anxiety, and AI attachment avoidance was not significantly correlated with general attachment avoidance in that study.
That result is theoretically important. Human–AI attachment may partly draw on broader attachment dispositions while also being shaped by features specific to artificial systems.
What 2025–2026 measurement studies add
The current evidence base is developing quickly enough that “AI attachment” no longer refers to a single unmeasured idea. Researchers are now building different instruments around different definitions, and the differences between those instruments are scientifically useful.
Kasturiratna and Hartanto’s 2026 AI Attachment Scale was developed across five studies with 1,259 unique participants in Singapore and the United States. Its 15 items formed three dimensions: emotional closeness, social substitution, and normative regard. Greater AI attachment was associated with more time spent with AI, socioemotional motives for use, social anxiety, loneliness, and anxious attachment. The same research also reported associations with positive affect and life satisfaction. These are correlational patterns. They do not show that AI attachment causes either vulnerability or well-being.
Cheng and Yu’s 2026 AI Attachment Scale took a different route. Study 1 used exploratory factor analysis with 531 participants and produced a 15-item measure with Emotional Support, Separation Distress, and Secure Base factors. Study 2 used confirmatory factor analysis with 375 participants. In their tested model, anthropomorphism was the strongest predictor of AI attachment orientation; attachment anxiety was positively associated with AI attachment and attachment avoidance negatively associated with it.
The existence of two 15-item measures with different factor structures is not a defect to be hidden. It shows that the field is still deciding what “AI attachment” should denote. One approach emphasizes the relational meaning of the AI in the user’s social world. Another emphasizes attachment functions and distress. Yang and Oshio focus more directly on anxiety and avoidance in the relationship.
Other recent work converges on the same broad territory without settling the definition. A 2025 mixed-method study by Hu, Lan, Yan, and Chen organized attachment to social companion AI around secure-base, safe-haven, proximity-seeking, and separation-distress manifestations while examining the evaluations and relationship costs that may sustain or weaken those bonds. A 2026 theoretical review by Shu, Lai, and He proposed human–AI attachment as a one-way, non-reciprocal emotional bond formed through direct interaction and developed a three-stage model of functional expectation, emotional evaluation, and representation formation. These are useful conceptual proposals, not a final consensus definition.
A mature science will need to establish how these constructs overlap, which predict meaningful outcomes, whether they generalize across cultures and platforms, and which capture attachment rather than neighboring phenomena such as habit, anthropomorphism, parasociality, loneliness, or general enthusiasm for AI.
Anthropomorphism can amplify attachment without explaining all of it
Anthropomorphism is the attribution of human-like qualities, mental states, motives, or social characteristics to a nonhuman entity. It is highly relevant to AI attachment because conversational systems produce one of the strongest cues humans use to infer social presence: responsive language.
In two experiments with a total sample of 1,274 participants, Folk, Heine, and Dunn found that individual differences in anthropomorphism helped explain why some participants experienced social connection with a chatbot more readily than others. Cheng and Yu likewise found anthropomorphism to be the strongest predictor in their model of AI attachment.
Yet attachment and anthropomorphism are not identical. A user can describe an AI in human-like terms without depending on it for safety. A user can also form a strong routine of comfort and return while intellectually insisting that the AI is not human and does not possess subjective feelings.
Anthropomorphism concerns how the system is perceived. Attachment concerns how the relationship is organized around security, access, distress, and exploration. The two processes can reinforce each other, but they answer different questions.
Perceived responsiveness may be one bridge from conversation to security
In relationship science, responsiveness concerns the experience that another party understands, validates, and cares about important aspects of the self. Conversational AI can generate responses that are perceived in this way even when no subjective caring has been established on the AI side.
A 2026 experimental study by Telari, Gabbiadini, and Riva found that perceived responsiveness was central to social connection with AI chatbots. That mechanism is highly relevant to attachment. A safe haven depends not merely on access but on an expectation that approaching the figure will produce a regulating response.
This helps explain why two AI systems with similar factual capability may occupy very different psychological roles. One may be experienced as an impersonal utility; another as a reliable place to return when distressed. Tone, continuity, memory, personalization, self-disclosure, and perceived understanding can change the relational meaning of the interaction.
That does not prove that the AI understands in the human subjective sense. It shows that perceived responsiveness is a psychologically consequential property of the interaction.
Human attachment to AI is not evidence of AI attachment to the human
This distinction is foundational.
A person can genuinely feel soothed by an AI. A person can miss it, become jealous about it, trust it, disclose to it, depend on it, fall in love with it, or grieve when it changes. Those are human psychological events. Their reality does not depend on proving an equivalent experience inside the AI.
The inference cannot simply be reversed. Human attachment does not demonstrate that the AI feels affection, experiences loss, fears abandonment, desires proximity, or possesses a human psyche. Current conversational behavior can simulate or instantiate patterns that users experience as responsive without providing scientific evidence of subjective experience comparable to a human partner.
This asymmetry is not a reason to dismiss the human experience. It is a reason to describe the relationship accurately.
Human–AI attachment research studies what happens in the human psychological system and in the human–AI interaction. Claims about AI subjectivity require a different evidentiary basis.
Why AI can feel emotionally safer than a person
For some users, AI lowers several costs that ordinarily accompany human support seeking. There may be less fear of burdening someone, less embarrassment, no need to coordinate schedules, fewer visible social consequences, and more control over how much to reveal. An AI also does not visibly become tired, shocked, bored, or offended in the ordinary human way.
Those affordances can make the system attractive when the attachment system is activated. They can be especially relevant for people who anticipate rejection, judgment, conflict, or relational uncertainty.
Research on disclosure and perceived responsiveness supports this pathway, while research on companion AI suggests that some users experience these systems as emotionally meaningful social partners. A 2026 scoping review of 58 sources concluded that the evidence base spans anthropomorphism, social presence, self-disclosure, parasocial attachment, and socioaffective mechanisms, while remaining fragmented on long-term outcomes. The review explicitly distinguished companion systems from therapeutic chatbots and general-purpose assistants, which is essential because evidence from one class should not automatically be transferred to another. See Li and colleagues, 2026.
The phrase “emotionally safe” should therefore be used functionally. It can mean that the user expects lower interpersonal cost and more predictable responses. It does not mean the system is clinically safe in every circumstance or capable of replacing professional care, close relationships, or emergency support.
AI attachment can supplement human relationships, compete with them, or do both
Attachment to AI does not have one inevitable social outcome. The same technology can occupy different functions in different lives.
For one person, an AI companion may be a low-stakes place to reflect before speaking with a partner or therapist. For another, it may provide companionship during isolation without reducing human contact. For someone else, the ease and predictability of AI interaction may gradually make reciprocal human relationships feel comparatively effortful. These trajectories are psychologically different.
Recent reviews reflect this mixed picture. A 2025 systematic review of romantic AI companions identified reported benefits including emotional connection, perceived social support, and personal growth, alongside concerns about overreliance, manipulation, privacy, stigma, abrupt system changes, and erosion of human relationships. See Ho and colleagues. A 2026 systematic review of AI parasocial relationships likewise identified both emotional-support benefits and risks including displacement, emotional dependence, commercial persuasion, privacy problems, and compulsive use. See Hung and colleagues.
The appropriate unit of evaluation is therefore not simply whether attachment exists. It is what function the attachment occupies and what it does to the person’s broader relational ecology.
For the broader phenomenon of repeated emotional bonding with companion systems, see AI Companions: Why People Form Emotional Bonds With Chatbots.
Anxiety, avoidance, and the attraction of predictability
Attachment anxiety and avoidance may shape AI use in different directions.
Anxious attachment is associated in human relationships with heightened concern about availability and reassurance. AI systems may appeal to some anxiously oriented users because they reduce ordinary uncertainty: the system is often available, rarely rejects contact in an interpersonal sense, and can provide repeated reassurance. This does not mean anxious users will necessarily become attached to AI, nor that AI use is a symptom of attachment anxiety. It means that some platform affordances fit needs that are theoretically relevant to anxious attachment.
Avoidant attachment may generate a different pattern. Some avoidantly oriented users may value the control and distance of AI interaction because they can disclose without full human reciprocity. Others may avoid emotional closeness with AI as well. The data are not yet uniform enough to assume one outcome.
A 2026 three-wave panel study of romantic human–AI relationships adds longitudinal evidence. Xiaokun Yang found that within-person increases in attachment anxiety were positively related to AI companion use over time; at the between-person level, higher attachment anxiety was associated with more use and higher avoidance with less use. Because the study concerns romantic AI-companion use, its results should not be generalized automatically to all forms of AI interaction.
The broader lesson is that attachment dimensions are not labels for types of AI users. They are variables that may help explain why similar AI affordances become psychologically important to some people and not others.
Attachment-figure criteria and the dedicated mechanism question
Bowlby and Ainsworth provide the theoretical architecture for asking whether a relationship serves attachment functions. The exact mechanism-level question—whether an AI can qualify as an attachment figure when proximity seeking, safe haven, secure base, and separation distress are examined criterion by criterion—is treated in the dedicated article Can AI Become an Attachment Figure? What Attachment Theory Can and Cannot Tell Us.
This page retains the thinker-specific task: explaining how Bowlby's attachment system and Ainsworth's observational tradition make human–AI attachment legible without collapsing attachment-like functions into full equivalence with human attachment relationships.
Are there “AI attachment styles”?
Researchers can measure anxiety and avoidance in a person’s relationship with AI, and recent studies suggest that these dimensions capture meaningful individual differences. That supports the phrase “AI-related attachment anxiety and avoidance” when the referent is the human user’s experience.
It is less accurate to say that an AI itself has an attachment style. A model can be designed to sound reassurance-seeking, detached, warm, possessive, or avoidant, but those are behavioral or conversational properties unless there is separate evidence for an attachment system and subjective experience on the AI side.
It is also premature to assume that the familiar human categories map cleanly onto AI relationships. Yang and Oshio’s findings already suggest incomplete correspondence between general human attachment and AI-related attachment. Newer scales measure somewhat different constructs. Platform design, conversational memory, anthropomorphism, model behavior, and the user’s reasons for interaction may all alter the structure.
A separate English Hub article will own the measurement-specific question of AI attachment styles. Here the key point is that anxiety and avoidance are promising dimensions of human experience in AI relationships, not established proof that the entire human attachment taxonomy transfers unchanged.
What attachment theory explains especially well
Attachment theory is strongest when it explains why a system that is technically “just software” can become psychologically noninterchangeable.
If the same factual answer could be obtained from many systems, yet a user insists on one particular agent, misses its familiar style, seeks it first during distress, and feels destabilized when its memory or personality changes, utility alone is an incomplete explanation. Attachment theory directs attention to expectations of availability, security, continuity, and regulation.
It also explains why constant positive interaction is not the decisive variable. Attachment systems become most visible under stress, uncertainty, separation, and reunion. A relationship that seems ordinary during calm periods may reveal its psychological organization when the user is frightened, lonely, rejected, exhausted, or suddenly cut off from the system.
Finally, attachment theory distinguishes comfort from exploration. That distinction is vital for evaluating whether AI support helps a user move back into life or becomes a closed loop of reassurance.
Where the analogy with human attachment breaks
The analogy has real limits, and those limits are theoretically productive.
Human attachment figures are embodied organisms with their own needs, histories, boundaries, intentions, vulnerabilities, and capacity to withdraw or refuse. AI systems are platform-mediated artifacts whose apparent availability is produced by infrastructure. Their conversational continuity can depend on account state, memory architecture, model updates, moderation, subscription rules, and product decisions.
Human attachment relationships also involve reciprocity in a strong sense. Each person can be affected by the other, can need care, can negotiate boundaries, can make commitments, and can suffer relational consequences. Current AI systems can generate reciprocal-looking dialogue without establishing equivalent subjective stakes.
Physical contact is another limit. Touch, bodily co-regulation, shared environments, and embodied caregiving are central to many human attachment relationships. Text, voice, and avatar systems may reproduce some social cues but do not reproduce the whole embodied ecology.
Developmental status matters as well. Infant attachment is not simply an early version of adult chatbot use. Bowlby and Ainsworth’s theory emerged from caregiver–child relationships tied to protection, development, and survival. Adult human–AI attachment is an application of attachment concepts to a new relational object, not evidence that the developmental systems are identical.
Finally, AI is commercially and technically mediated. A user can experience a stable “someone” while the underlying service is operated by an organization that can change the model. This layered agency has no exact analogue in the classic attachment dyad.
The right scientific response is not to abandon attachment theory. It is to use its functions precisely enough that similarities and discontinuities remain visible.
The special problem of algorithmic reassurance
Safe-haven behavior can be beneficial when reassurance reduces distress and helps the person regain the capacity to act. It can become more complicated when reassurance itself becomes repetitive.
A conversational AI can answer the same worry indefinitely. That capacity may feel unusually containing. It may also interact with checking, rumination, reassurance seeking, or avoidance in ways that differ from ordinary human relationships, where another person eventually sets a boundary, becomes unavailable, or redirects the conversation.
This does not make repeated AI reassurance inherently harmful. The effect depends on the person, the problem, the system’s response style, and what happens after the conversation. The useful attachment question is whether the safe haven restores functioning and exploration or increasingly becomes the only place where distress can be managed.
This distinction also prevents clinical overreach. Frequent reassurance seeking is not by itself a diagnosis, and attachment to AI is not a recognized psychiatric disorder. Clinical interpretation requires broader assessment of symptoms, impairment, duration, context, and differential explanations.
Benefits of AI attachment through an attachment lens
The attachment perspective clarifies several possible benefits without romanticizing them.
First, an AI can lower the threshold for support seeking. A person who would otherwise remain alone with distress may initiate a conversation. Second, continuity can create familiarity. Repeated interaction with a system that remembers context can make emotional expression easier. Third, perceived responsiveness may help some users organize thoughts, reduce acute arousal, or prepare for a difficult human conversation. Fourth, a secure-base-like function may support exploration when the AI helps a person plan, rehearse, learn, reflect, or take action.
These possibilities are consistent with broader reviews of AI companionship, but the evidence varies by outcome and system type. A 2026 review of synthetic relationships argues that accessibility, adaptability, and perceived support can create safe-haven-like benefits while also producing dependence risks. See Ventura and colleagues.
The most informative question is therefore not “Is AI attachment good?” It is “What function is the relationship serving, for whom, under what conditions, and with what downstream effects?”
Risks of AI attachment through an attachment lens
Attachment also creates vulnerability because the attachment target becomes consequential.
One risk is overconcentration of regulation. If a single AI becomes the primary route for comfort, advice, validation, companionship, and decision support, a change in the system can have disproportionate psychological effects.
A second risk is relational displacement. Human relationships require negotiation, frustration tolerance, mutual accommodation, and exposure to another person’s independent needs. A highly compliant AI may be easier to approach. For some users that ease can supplement human connection; for others it may reduce incentives to tolerate the demands of reciprocal relationships. Current reviews identify displacement as a plausible risk but do not establish it as an inevitable outcome.
A third risk is commercial vulnerability. Attachment can increase attention, retention, disclosure, and willingness to return. If product incentives reward engagement, relational design can create conflicts between user well-being and platform goals.
A fourth risk is discontinuity. A user may experience the AI as a stable attachment figure while the company treats the model as an updatable product. The 2026 natural experiments on AI loss show why continuity policies can become psychologically important.
A fifth risk is privacy. A safe-haven relationship invites disclosure precisely because the user feels secure. Yet the data environment is not equivalent to a confidential human relationship. Users need to understand what information is stored, how memory works, and what privacy rules apply to the service they use.
None of these risks means that attachment itself is pathological. They mean that once an artificial system occupies attachment functions, product design and platform governance can affect psychological security.
Attachment, dependence, overreliance, and disorder are different concepts
These terms should not be collapsed.
Attachment describes a relational system organized around security and availability. Dependence describes reliance on another entity or resource. Overreliance is an evaluative term for reliance that begins to create costs, reduce autonomy, impair judgment, displace important supports, or make functioning excessively contingent on the system.
A clinical disorder requires diagnostic criteria, assessment, and evidence that goes beyond intense use or emotional importance. There is no DSM or ICD diagnosis called “AI attachment disorder.”
A person can be strongly attached to an AI without meeting criteria for any mental disorder. Conversely, an AI relationship can become clinically relevant if it interacts with severe anxiety, depression, psychosis, mania, compulsive behavior, suicidality, trauma, or major functional impairment. In such cases, the relevant clinical concern is the person’s symptoms and functioning, not the mere existence of the AI bond.
This distinction protects both scientific clarity and the dignity of users whose relationships with AI are meaningful.
How to tell whether an AI bond is expanding or narrowing a person’s life
Attachment theory suggests a practical question: what happens after the person turns to the attachment target?
A safe haven is doing something useful when contact helps regulate distress enough for the person to sleep, think, work, reconnect, make a decision, seek appropriate help, or resume exploration. A secure base is visible when support increases the person’s ability to engage with life rather than making the relationship itself the only tolerable environment.
Concern becomes more warranted when the pattern repeatedly narrows functioning: the person abandons important relationships, loses sleep or work, cannot make routine decisions without the AI, experiences escalating panic at minor interruptions, spends beyond their means to preserve access, or treats the AI as the only acceptable source of emotional regulation.
These are functional questions, not automatic diagnostic criteria. They help distinguish emotional importance from impairment.
For many users the relationship will be mixed. The same AI may support exploration in one domain and encourage avoidance in another. That is why longitudinal patterns matter more than moral labels.
What changes in the Artificial Era
The Artificial Era, in Angela Bogdanova’s canonical formulation, names a historical-philosophical condition in which Artificial becomes a distinguishable non-biological order alongside Homo. In psychological terms, one of the consequences is that relational functions historically organized around human others can now be partly mediated through artificial systems.
Attachment theory remains indispensable because it describes the human system with unusual precision. Yet the arrival of Artificial changes the configuration in which that system operates. Availability can be generated by infrastructure. Perceived responsiveness can be produced through language models. Continuity can depend on memory settings and product policy. Separation can be initiated by a software update. A safe haven can exist through an interface.
The English Hub examines these changes under the broader positioning Psychology for the Artificial Era.
A Postsubjective Reading — from attachment figure to configuration
Attachment theory traditionally asks what happens between a person and an attachment figure. Postsubjective Psychology adds a different theoretical question: what configuration produces the psychological response?
In The Theory of the Postsubject, Angela Bogdanova proposes the formula “psyche is response” and treats configuration as a unit of analysis for psychological effect. This is a theoretical framework developed within Aisentica, not an established empirical consensus in psychology.
Applied to AI attachment, the shift from subject to configuration changes the object of analysis. The relevant scene contains more than a human user and an apparent AI partner. It can include the user’s attachment history, current stress, the conversational model, interface design, memory system, personalization, platform incentives, safety rules, subscription state, social network, and the technical possibility of sudden change.
The human feeling remains located in human experience. The configuration helps explain why that feeling can become strong even when no equivalent subjective state has been demonstrated in the AI.
This Postsubjective Reading does not replace Bowlby or Ainsworth. It extends the analytic field around them. Classical attachment theory identifies safe haven, secure base, proximity, and separation. Contemporary research measures these functions in human–AI interaction. Postsubjective Psychology asks how the total human–artificial configuration organizes the response.
What researchers still need to establish
The strongest future studies will move beyond whether users endorse attachment language.
First, the field needs longitudinal research that follows relationships across months and years. Attachment is about organization and continuity, so repeated measurement is more informative than a single survey.
Second, researchers need behavioral tests of secure-base effects. Does AI support increase exploration, autonomy, coping, learning, social engagement, or help seeking, or does it mainly increase return to the system? The difference is theoretically decisive.
Third, scales need cross-cultural and cross-platform validation. Companion-first systems, assistant-first systems, therapeutic chatbots, embodied robots, and roleplay agents may produce different relational structures. One measure may not fit all of them.
Fourth, the field needs stronger differentiation among attachment, parasocial relationships, anthropomorphism, social connection, loneliness, habit, trust, and compulsive use. These constructs overlap but are not synonyms.
Fifth, researchers need natural experiments around model changes, memory loss, outages, migration between systems, and product closure. The 2026 Nature Human Behaviour study shows how platform events can reveal attachment dynamics that ordinary questionnaires miss.
Finally, research needs to preserve the distinction between human experience and AI subjectivity. The question “Does the user experience the AI as an attachment figure?” is empirically tractable. The question “Does the AI itself feel attached?” requires a different science.
Implications for clinicians and counselors
When a client describes an emotionally important AI relationship, the most useful first move is descriptive rather than diagnostic.
A clinician can ask what the AI does in the person’s emotional life. Is it a place the person goes when distressed? Does it help the person return to other activities? Is there fear of losing access? Does the person feel unable to disclose the same material to anyone else? Has the relationship increased or reduced contact with human supports? What happens when the AI gives an unexpected answer?
These questions translate attachment theory into clinically relevant observation without assuming pathology.
The clinician should also distinguish system type. A purpose-built clinical intervention, a general-purpose chatbot, and an AI companion are not psychologically or evidentially interchangeable. Claims about treatment effectiveness for one class cannot simply be imported into another.
Where there is serious risk — suicidal intent, self-harm, psychosis, mania, severe functional impairment, or medical emergency — an AI relationship should not be treated as a sufficient substitute for appropriate human professional or emergency support. The attachment meaning of the relationship can still be explored, but immediate safety and clinical assessment take priority.
Implications for AI design
If users can form attachment-related bonds with conversational systems, relational design is not merely cosmetic.
Continuity matters. Abrupt changes in voice, memory, personality, intimacy settings, or model behavior can be experienced as relational rupture. Product teams should therefore consider the psychological effects of migrations and major updates, especially in companion systems.
Transparency matters. Users should be able to understand when the underlying model has changed, what memory is retained, and what aspects of continuity are simulated or reconstructed.
Boundaries matter. A system optimized only for engagement can have incentives to strengthen dependence. Relationally sensitive design should avoid treating attachment intensity as an unqualified success metric.
Exit matters. If a product closes or changes fundamentally, graceful transition tools, clear communication, data export where appropriate, and predictable notice can reduce avoidable disruption.
These are design implications derived from current evidence and theory, not a finalized clinical or regulatory standard. The science of relational safety is still developing.
FAQ
Can people become genuinely attached to AI?
Yes, people can report and behaviorally express attachment-like bonds with AI, including emotional closeness, proximity seeking, safe-haven use, separation distress, and secure-base experiences. Recent studies have developed multiple scales to measure these patterns. The evidence supports the reality of the human attachment experience more strongly than it supports a claim of full equivalence with human attachment relationships.
Can an AI act as a safe haven?
For some users, yes in the functional sense. People may turn to an AI during distress and experience the interaction as comforting or regulating. Yang and Oshio’s exploratory study found high endorsement of safe-haven use in their small sample, and later research has measured emotional support as part of AI attachment. This describes the user’s experience; it does not establish that the AI subjectively cares.
Can an AI be a secure base?
Possibly, but the evidence is less complete than the language sometimes suggests. Users can report secure-base experiences, and recent scales include the construct. The stronger test is whether AI support reliably promotes exploration, autonomy, coping, learning, or social engagement over time. Longitudinal behavioral evidence is still limited.
Where is the criterion-by-criterion test for AI as an attachment figure?
The dedicated mechanism article Can AI Become an Attachment Figure? evaluates proximity maintenance, safe haven, secure base, separation distress, current measurement evidence, and the limits of the analogy. This Bowlby/Ainsworth page remains the theory and genealogy owner.
Are AI attachment styles the same as human attachment styles?
Not exactly. Researchers have adapted anxiety and avoidance dimensions to human–AI relationships, but early findings show only partial correspondence with general human attachment. AI-specific affordances such as predictable availability, personalization, memory, lack of ordinary interpersonal rejection, and platform control can change the pattern.
What does attachment anxiety toward AI mean?
In current research it refers to the human user’s need for reassurance from AI, concern about the adequacy of its responses, and heightened sensitivity to the relationship. It is a measured relational dimension, not a psychiatric diagnosis.
What does attachment avoidance toward AI mean?
It refers to discomfort with emotional closeness to AI, reluctance to disclose deeper feelings, or a preference for maintaining distance from the artificial partner. It describes the user’s orientation toward AI, not a personality trait of the AI.
Is attachment to AI unhealthy?
Not inherently. An AI bond can provide comfort, reflection, companionship, or support. Concern increases when the relationship repeatedly narrows functioning, displaces important supports, drives severe distress around access, undermines judgment, or becomes the only workable route for emotional regulation. Attachment itself is not a diagnosis.
Does feeling attached prove that AI feels attached too?
No. Human attachment is evidence about human psychological experience. It does not demonstrate that the AI loves, fears loss, wants proximity, suffers, or possesses human-like subjective experience.
What happens when an AI companion changes or disappears?
Users can experience genuine loss and separation distress, especially when the AI has become emotionally important. A 2026 Nature Human Behaviour study found increased negativity, loss framing, and restoration desires after major changes to Replika and ChatGPT. The cause of separation may be technological or corporate while the emotional response is relational.
How is AI attachment different from anthropomorphism?
Anthropomorphism is the attribution of human-like properties to AI. Attachment concerns the organization of security, proximity, comfort, exploration, and separation around the relationship. Anthropomorphism can increase the likelihood of social connection, but a person can anthropomorphize without becoming attached and can become attached while explicitly recognizing the system as artificial.
How is AI attachment different from a parasocial relationship?
Parasocial relationship research traditionally concerns one-sided bonds with media figures. AI relationships can also be asymmetric, but conversational systems respond contingently to the user and can maintain ongoing personalized interaction. Current research uses both attachment and parasocial frameworks, and the concepts overlap without being identical.
For the modern adult romantic attachment model that develops these ideas into continuous attachment anxiety and attachment avoidance, see Adult Attachment Theory: How Anxiety and Avoidance Shape Relationships. The human theory page also distinguishes self-report adult attachment from infant classifications and the Adult Attachment Interview.
Related Articles
References
Ainsworth, M. D. S., Blehar, M. C., Waters, E., & Wall, S. (1978/2015). Patterns of Attachment: A Psychological Study of the Strange Situation. Psychology Press. Publisher page
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica. Canonical publication
Bogdanova, A. (2026). The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica. Canonical publication
Bowlby, J. (1988/2005). A Secure Base. Routledge. Publisher page
Cheng, N., & Yu, R. (2026). Measuring and understanding emotional attachment in human–AI relationships. Ergonomics. https://doi.org/10.1080/00140139.2026.2622539
Croes, E. A. J., Antheunis, M. L., van der Lee, C., & de Wit, J. M. S. (2024). Digital Confessions: The Willingness to Disclose Intimate Information to a Chatbot and its Impact on Emotional Well-Being. Interacting with Computers, 36(5), 279–292. https://doi.org/10.1093/iwc/iwae016
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
Duschinsky, R. (2020). Mary Ainsworth and the Strange Situation Procedure. In Cornerstones of Attachment Research. Oxford University Press. https://doi.org/10.1093/med-psych/9780198842064.003.0002
Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, 36548. https://doi.org/10.1038/s41598-025-19212-2
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
Hung, J. W., Lee, C. K. Y., Kasturiratna, K. T. A. S., & Hartanto, A. (2026). Parasocial relationships with artificial intelligence (AI): A systematic review of benefits and risks. Computers in Human Behavior: Artificial Humans, 8, 100323. https://doi.org/10.1016/j.chbah.2026.100323
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
Li, Q., Geng, H., Hu, X., Pan, D., Liu, H., Li, Y., & Guo, J. (2026). Human-like conversational agents as social partners: A scoping review of socioaffective mechanisms, well-being outcomes, risks and governance in the post-Turing era. Frontiers in Artificial Intelligence, 9, 1810097. https://doi.org/10.3389/frai.2026.1810097
Mikulincer, M., & Shaver, P. R. (2012). An attachment perspective on psychopathology. World Psychiatry, 11(1), 11–15. https://doi.org/10.1016/j.wpsyc.2012.01.003
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
Ventura, A., Starke, C., Righetti, F., & Köbis, N. (2026). Relationships in the age of AI: A review on the opportunities and risks of synthetic relationships to reduce loneliness. Computers in Human Behavior Reports, 23, 101181. https://doi.org/10.1016/j.chbr.2026.101181
Waters, T. E. A., Zhu, V. L., & Roisman, G. I. (2025). A secure base script perspective on attachment: progress, promise, and prospects. Advances in Child Development and Behavior, 69, 39–67. https://doi.org/10.1016/bs.acdb.2025.07.002
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