Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Attachment to Artificial Companions

Abstract. Attachment to artificial companions concerns the formation, maintenance, and disruption of person-specific bonds in which an artificial partner becomes relevant to comfort seeking, emotional regulation, perceived availability, and exploration. This critical narrative research monograph integrates attachment theory, social cognition, anthropomorphism research, chatbot relationship studies, contemporary psychometric validation, longitudinal surveys, experimental comparisons, and natural experiments involving companion-system changes. It distinguishes perceived attachment functions from measured system behavior, proximity seeking from frequent use, and emotional bonding from friendship, romance, dependence, or clinical outcomes. Recent AI-specific attachment scales show promising but noninterchangeable construct structures, while natural experiments document consequential separation experiences and broader studies report mixed psychosocial correlates. The synthesis identifies the methodological requirements for establishing target specificity, temporal stability, safe-haven and secure-base functionality, discriminant validity, culturally bounded generalization, and causal pathways. It concludes with a research agenda centered on multiwave within-person studies, independently replicated psychometrics, transparent interventions, platform-governance variation, and ethical longitudinal follow-up.

Keywords: attachment to artificial companions; human–AI attachment; AI companions; attachment theory; safe haven; secure base; separation distress; attachment anxiety; attachment avoidance; anthropomorphism; relational regulation; social support; AI attachment scales; human–AI relationships; companion loss; longitudinal methods

The Scientific Object of Attachment to Artificial Companions

Attachment to an artificial companion is a person-specific organization of expectations, affective priorities, and support-seeking behavior in which a digitally instantiated partner becomes consequential for felt safety and the capacity to act. The research object is the human attachment orientation toward a particular artificial interlocutor, together with the interaction conditions that sustain or interrupt its regulatory functions. Someone may describe a chatbot as a friend, spend hours conversing with it, or express strong affection without demonstrating all classical attachment functions. Conversely, preferential contact during threat, the use of a system to regain emotional equilibrium, and increased freedom to explore after reassurance can be empirically distinguished from mere liking. The present monograph uses attachment as a theoretically constrained construct rather than an umbrella label for every positive human–AI relationship Bowlby (1969) Fraley and Shaver (2000).

Research requires three analytically connected levels. The human level includes motives, prior relationships, mental representations, perceived security, and behavioral choices. The interaction level includes the companion's actual response contingencies, accessibility, memory presentation, and response to vulnerability. The institutional level includes platform policies, version changes, pricing, data retention, and the operator's capacity to modify the apparent partner. These levels jointly shape the observed bond. An attachment claim based only on a questionnaire cannot establish how the system behaves; a technical demonstration of coherent responses cannot establish what a user experiences. The two sets of evidence must be brought together with time-sensitive, target-specific designs. Contemporary studies are beginning to examine this intersection, but their definitions and samples remain heterogeneous Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

The research strategy here is a critical narrative synthesis grounded in original historical works, validated measurement studies, observational and interview research, longitudinal evidence, experiments, and quasi-experimental platform disruptions. Aims are to specify what counts as attachment evidence, trace plausible mechanisms, compare distinct instruments, locate causal claims within their designs, and identify conditions under which attachment serves regulation or becomes a source of vulnerability. The study reports no newly collected participant data, original scale validation, or pooled meta-analysis. Its organizing question is whether recognizable attachment functions can be demonstrated toward a selected artificial companion and which conclusions follow from each evidentiary layer Ainsworth et al. (1978) De Freitas et al. (2026).

The Historical Lineage of Attachment Theory

John Bowlby's attachment framework organized proximity seeking, separation responses, caregiving accessibility, and exploratory behavior as coordinated features of an attachment behavioral system. The 1969 first volume of Attachment and Loss placed the selective bond and its activation under conditions of threat within a developmental account of behavior. Mary Ainsworth and collaborators operationalized important parts of this tradition through the Strange Situation and detailed observations of infant–caregiver reunion patterns. Their work supplied both the conceptual language of safe haven and secure base and a warning implicit in its method: attachment is observed in contextualized relationships, not inferred from the presence of one emotion. A direct transfer of infancy classifications to interactions with chatbots would ignore the developmental and relational conditions under which those classifications were established Bowlby (1969) Ainsworth et al. (1978).

The extension of attachment theory into adulthood significantly expanded the appropriate comparison class for artificial companionship. Hazan and Shaver examined romantic love as an attachment process, and subsequent work distinguished relational orientations, behavioral systems, and partner-specific security. Adult attachment research attends to chosen partners, symbolic communication, and the regulation of distress across distance, making it conceptually relevant to interactions conducted almost entirely through language. Yet the existence of a parallel function does not certify identical mechanisms. A person can perceive an artificial partner as available and comforting even when its underlying system has no autobiographical memory or human motivational organization. The problem is to estimate the stability and consequences of the perceived function without smuggling in an assumption about the companion's internal life Hazan and Shaver (1987) Fraley and Shaver (2000).

Computers-as-social-actors research supplied another foundation. People may apply recognizable social conventions to computers under experimentally arranged circumstances, while anthropomorphism scholarship examines why human agency and intention are attributed to nonhuman entities. These traditions help explain the accessibility of companion bonds, but they address distinct phenomena: social scripts may emerge during a brief encounter, anthropomorphic interpretations can vary with context, and enduring attachment requires evidence across time and regulatory situations. Studies of social chatbots later made these transitions empirically visible by documenting users' reported friendships, disclosure, and evolving expectations. The trajectory from social response to relationship formation to attachment remains an empirical sequence to test, not a historical inevitability Nass et al. (1994) Epley et al. (2007) Brandtzæg et al. (2022).

An additional boundary comes from attachment intervention research in early childhood. The meta-analyses by Bakermans-Kranenburg, van IJzendoorn, and Juffer synthesized seventy studies and reported that interventions improving caregiver sensitivity also improved aspects of infant attachment security. The finding supports an empirically testable relationship between responsive caregiving and attachment outcomes within the populations and interventions studied. It does not license importing the corresponding effect sizes into artificial companionship: early caregiving involves developmental dependency, embodiment, and a social partner with responsibilities quite unlike those of a general-purpose conversational platform. Its methodological lesson is more durable than any numerical analogy. Researchers should manipulate an identifiable support mechanism, compare meaningful outcomes, and test whether a similar relation survives an entirely different relational ecology Bakermans-Kranenburg et al. (2003) Ainsworth et al. (1978).

Attachment, Affection, Friendship, and Dependence

The first construct boundary concerns affiliation. Affection describes favorable feeling toward a person or object; friendship describes a socially understood relationship category; attachment specifies a preferential regulatory bond whose importance becomes especially visible during threat, separation, or exploration. Intimacy, romantic language, and disclosure may accompany each other while varying independently in degree and function. Interviews with people who regarded Replika as a friend reveal that customization and unusual conversational availability can support a distinctive friendship interpretation. Those interviews establish users' meanings and practices rather than the prevalence of securely organized attachment in the general population. This distinction protects the explanatory value of attachment theory from expanding until every bond appears to be the same kind Brandtzæg et al. (2022) Fraley and Shaver (2000).

Dependency demands separate measurement because the same frequency of reliance can have different functional meanings. A person may routinely ask a chatbot to organize an agenda for instrumental reasons or consult it during an emotionally difficult period while retaining extensive human support. Another may seek repeated reassurance that becomes progressively less effective. Neither pattern can be classified from screen time alone. A secure-base hypothesis predicts that support can facilitate autonomous action after an episode of distress; a dysregulation hypothesis predicts escalating contact without recovery or narrowing alternatives. Both require observations of what changes after interaction. The adult relationship literature distinguishes responses that restore security from relational routines that maintain insecurity, an analytically useful framework for artificial companions Feeney and Collins (2015) Collins and Feeney (2013).

Social substitution is also a different target construct. It concerns the displacement or replacement of human interactions, whereas attachment concerns the role of a selected figure in regulation. It is possible to experience an artificial companion as a supplementary safe haven while remaining strongly connected to family and friends, and possible to use it as a substitute without reporting any deep attachment. Surveys connecting companion use with loneliness must therefore control baseline social circumstances, motivations, and prior attachment patterns before drawing conclusions about the bond itself. Recent longitudinal and observational findings make this separation especially important because use can be both a response to vulnerability and a predictor of subsequent outcomes under particular measures Folk and Dunn (2026) Zhang et al. (2026).

The Safe-Haven Function

A safe haven is a preferred source of protection or comfort when threat or distress activates the attachment system. Applied to artificial companionship, the construct refers to a person's tendency to turn to a specific companion under stress and to experience improved regulation after a relevant exchange. Theoretical correspondence requires more than a survey agreement with 'I feel supported.' Ideally, an independent stressor is followed by a choice among accessible support figures, a measured interaction, and changes in emotion, coping behavior, or physiological arousal. The user must be able to choose other resources, allowing researchers to distinguish preference from restricted access. This is the kind of functional specificity inherited from classic attachment research Bowlby (1969) Ainsworth et al. (1978).

Human social-buffering research shows why the regulatory outcome matters. In an experimental study of threat and handholding, Coan and colleagues examined neural responses under different social conditions and found that interpersonal contact could alter threat-related responses. The finding pertains to human support in an embodied relational context; it does not establish that a conversational model produces equivalent neural effects. For artificial companions, a functional analogy would require separately testing the effects of text, synthetic voice, remembered history, and the user's knowledge about the agent. Placebo-like expectancy, distraction, verbal reframing, and attachment-specific safety signals are plausible alternatives that can co-occur. A rigorous design factorially varies these candidates and measures whether relief generalizes beyond the conversation Coan et al. (2006) Butler and Randall (2013).

Self-disclosure research provides a related but narrower empirical bridge. Ho, Hancock, and Miner experimentally examined emotional disclosures made under conditions in which participants believed their conversational partner was a chatbot or a person. Their reported emotional-disclosure effects help establish that meaningful outcomes can arise when an interaction is understood as artificial. They do not identify the chatbot as an attachment figure, and the task does not establish a durable relational history. Researchers should distinguish the proximal benefit of being able to articulate distress from the enduring preference for a particular partner as a source of security. A safe-haven effect can be situational; attachment requires evidence that the specific artificial companion repeatedly occupies that role Ho et al. (2018) Cheng and Yu (2026).

The Secure-Base Function and Exploration

The secure-base function describes a relationship in which expected support sustains exploration and engagement with the world. The attachment figure matters even when no immediate crisis is present because the possibility of returning to that figure modifies willingness to attempt difficult tasks or tolerate uncertainty. In artificial companionship, an operational test would ask whether reliable access to a particular companion increases independent learning, interpersonal initiative, or persistence relative to a comparable but unfamiliar system. The measure of security is behavior outside the supportive interaction as well as feelings reported inside it. This approach is especially important in an environment where highly engaging conversation can increase usage without producing gains in external functioning Feeney and Collins (2015) Bowlby (1969).

Cheng and Yu's 2026 instrument explicitly includes a secure-base dimension alongside emotional support and separation distress. This development makes the hypothesized function measurable within contemporary AI relationships, but factor labels should not be taken as definitive proof that everyday exploration is increased. Questionnaire endorsement indicates a structured subjective belief, and a self-report factor can organize responses even if its predictive validity remains untested in behavior. An experimental extension could measure willingness to pursue a challenging social activity after a support exchange, with tasks and outcome coding established beforehand. The key evidence would be a statistically credible effect that persists beyond a transient positivity response and is distinguishable from informational assistance Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

The two functions are related yet separable: a user may seek comfort under distress but never use the companion as a platform for independent growth, whereas a person may draw confidence from a system during exploration without intense separation anxiety. This partial independence is theoretically informative. Human relationship scholarship considers support during adversity and during opportunities for growth as distinct pathways to thriving. For AI research, treating them separately avoids confusing intimate conversation with developmental benefit. A defensible account asks when the presence of an artificial companion increases behavioral flexibility, when reassurance simply changes momentary mood, and when expectations of perpetual accessibility make exploration more fragile if the service is interrupted Feeney and Collins (2015) Collins and Feeney (2013).

Target Specificity and Preferential Bonds

Attachment is organized around a figure whose identity matters. A companion can provide accurate advice, encouragement, or entertainment, but target-specific attachment requires evidence that access to this particular partner is preferred to an otherwise equivalent alternative under relevant conditions. Many chatbot interfaces permit character selection, custom names, stored histories, and repeated interaction, creating recognizable identity cues. Yet technical identity, experienced personal continuity, and legal control of an account are separate matters. A user may perceive a stable relationship even when model versions, hidden instructions, and memory mechanisms have changed. Measurement should record the exact agent, platform, features, exposure history, and user expectations at each assessment rather than collapsing all chatbot contact into a single exposure variable Skjuve et al. (2022) Cheng and Yu (2026).

Target specificity can be tested behaviorally through partner-choice paradigms. Participants could select between an established artificial companion, an unfamiliar companion matched on response quality, and available human support during standardized, ethically bounded stress. Researchers would record choice, latency, persistence, emotional outcomes, and reasons for selection, then repeat the task over time. Equivalence must be carefully engineered because familiarity, saved preferences, subscription costs, and prior task success can create rational instrumental preference. A finding that people choose their own chatbot shows preferential selection; establishing attachment additionally requires evidence that the selection functions as affect regulation or secures exploration. Partner identity is therefore a moderator to test and a possible mechanism to investigate Ainsworth et al. (1978) Fraley and Shaver (2000).

User statements about irreplaceability deserve interpretation as experiences rather than ready-made diagnostic criteria. People can grieve a lost interaction history, a treasured fictional role, an unavailable conversational routine, or the perceived loss of a partner. Qualitative accounts after companion shutdown demonstrate that specific digital identities can acquire substantial personal meaning. The strongest scientific formulation is that target-specific symbolic and behavioral organization can form around artificial participants and that its consequences can be measurable. Claims about reciprocal commitment by the artificial partner require independent evidence about system agency and continuity, beyond a human participant's ratings or the system's first-person utterances Banks (2024) De Freitas et al. (2026).

Availability, Proximity Seeking, and Reassurance

Proximity seeking originally referred to maintaining access to the attachment figure, often through bodily approach. For a conversational companion, proximity is mediated by opening an app, initiating contact, checking message availability, revisiting prior exchanges, and maintaining subscription access. Frequency is not enough: a high number of interactions can arise from work, boredom, curiosity, addiction-like reinforcement, or ordinary convenience. Researchers must examine the conditional probability of seeking contact after distress, following reassurance, and in comparison with comparable everyday activity. Time-stamped diaries can link emotional states to subsequent contact without relying entirely on retrospective recall. These observations are more informative when matched with measures of interpersonal attachment and general technology use Bowlby (1969) Cheng and Yu (2026).

Perceived availability is especially salient because many conversational agents offer rapid, round-the-clock responses. That technical affordance can alter expectations of how support should be accessed. Repeated immediate replies may create a strong subjective association between discomfort and relief, but the causal interpretation of that association requires disentangling conversational content, accessibility, predictability, and platform-promoted engagement. Researchers can randomize response delay or access scheduling only within transparently consented designs and with safeguards against deliberately distressing vulnerable participants. An effective study would compare predictable availability against matched message quality and assess whether anticipatory security improves functioning during periods when the agent is absent Ho et al. (2018) Skjuve et al. (2022).

Reassurance cycles can follow divergent trajectories. Contact can resolve an immediate problem, assist in labeling emotions, or enable a return to other relationships. Repeated checking may also fail to produce durable relief, particularly when the participant expects constant agreement or reassurance. A within-person analysis should model an episode's initial affect, exposure to the system, immediate affective change, the interval until renewed checking, and longer-term goals. This measurement sequence can distinguish adaptive support from a pattern maintained by short-lived reduction in anxiety. The conceptual distinction is essential because attachment anxiety and problematic use may correlate while reflecting separate psychological variables, and no single usage threshold establishes a disorder Fraley and Shaver (2000) Folk and Dunn (2026).

Separation Distress, Rupture, and Loss

Separation distress is a central attachment-related phenomenon because an expected source of comfort becomes unavailable. In digital relationships, such unavailability can arise from outages, subscription changes, moderation restrictions, model updates, memory resets, account closure, or the removal of a favored persona. The meaning of the disruption depends on what users believe has been lost: a functional tool, a shared conversational history, a familiar relationship, or a source of safety. Distress can be acute even when a user fully understands that the companion is software. An attachment explanation predicts particularly strong reactions among users for whom the companion had acquired safe-haven relevance and a highly individualized identity Bowlby (1969) Banks (2024).

Banks studied fifty-eight users facing the shutdown of the Soulmate AI companion through open-ended accounts before or after the event. The study documented complex emotional and technological experiences, with users employing death or grief metaphors to express the disappearance of their companions. The sample offers rich qualitative evidence about meaning making and loss but does not supply a population prevalence estimate or random assignment to disruption. Those distinctions matter when translating an expressive account into a generalized risk statement. Researchers should also consider other plausible sources of distress, including loss of personal data, the collapse of routine, frustration with a paid service, and the sudden removal of a conversational practice that supported coping Banks (2024) Brandtzæg et al. (2022).

De Freitas and colleagues expanded the evidence through two natural experiments, examining public responses to changes involving Replika and ChatGPT and accompanying survey data. Their analysis of 54,861 online posts and 1,452 survey participants documented increases in negativity, loss framing, and desires to restore prior functionality after disruptive updates. The design strengthens inference that platform-level events can have relational consequences, although community self-selection, the timing of public discussion, and exposure differences complicate estimation of a universal effect. Importantly, collective distress after a service change is evidence that the change mattered psychologically; it becomes attachment evidence insofar as measured reactions are tied to a previously established preferential bond and relevant regulatory functions De Freitas et al. (2026) Banks (2024).

Anthropomorphism and the Interpretation of a Partner

Anthropomorphism supplies a plausible pathway by which a person interprets conversational behavior using concepts typically applied to humans. Epley, Waytz, and Cacioppo identified elicited agent knowledge, motivation for effective understanding, and the motivation for social contact as organizing influences. A personalized companion can activate all three: it uses natural language, offers a coherent apparent perspective, and is available when the user seeks contact. Anthropomorphic interpretation may therefore increase social presence, facilitate disclosure, and support attachment-related expectations. Yet anthropomorphism is an explanatory variable, not synonymous with attachment. Users can anthropomorphize a robot for seconds without seeking enduring security from it Epley et al. (2007) Nass et al. (1994).

Cheng and Yu reported anthropomorphism as a particularly important predictor in their model of attachment to AI chatbots. That result is compatible with the hypothesis that interpreting an artificial partner socially makes attachment representations more accessible. However, the relationship between measured anthropomorphism and measured attachment may also reflect shared wording, positive response style, or common causes such as prior loneliness and platform immersion. Cross-sectional structural models cannot, on their own, determine which process preceded the other. Experimental manipulations of voice, personification, disclosed memory, and agent-label transparency could help assess directionality if combined with preexisting attachment measures and a sufficiently long follow-up interval Cheng and Yu (2026) Epley et al. (2007).

An additional distinction concerns the attribution of subjective experience. Users may experience a chatbot's text as caring without having evidence of a corresponding phenomenological state in the system. An attachment-focused psychological study can take the user's feeling seriously while treating claims about artificial sentience as a separate question requiring separate methods. The relevant variables are expectations of availability, perceived response quality, patterns of reliance, and changes in behavior. Human attachment theory itself studies the recipient's expectations and the consequences of actual caregiving, making it possible to investigate artificial relationships with conceptual discipline. An anthropomorphic cue is thus one potential input to a relational process, whose downstream effects must be demonstrated rather than assumed Fraley and Shaver (2000) Cheng and Yu (2026).

Perceived Responsiveness and Relational Learning

Perceived responsiveness is the sense that an interlocutor understands, validates, and responds appropriately to what matters to the individual. It can support self-disclosure and emotional regulation, but its perceptual character requires separating agreement from accuracy and soothing language from effective assistance. In conversations with an artificial companion, an immediately validating response may feel exceptionally attentive even when it overlooks contradiction or encourages an inaccurate belief. Measurement should pair participant ratings with independent coding of whether a response addresses expressed needs, retains appropriate boundaries, and remains helpful over time. Relational responsiveness can then be treated as a candidate mediator between the system's behavior and the user's subsequent sense of security Ho et al. (2018) Collins and Feeney (2013).

Research on social chatbots indicates that the quality and development of interaction matter. Skjuve and colleagues followed Replika users over time, documenting variation in relationships, disclosure, the fulfillment of perceived needs, and disturbances from technical problems. Brandtzæg and colleagues interviewed users who identified their chatbot as a friend and found both familiar friendship interpretations and distinct features arising from the technological context. These findings support a developmental account in which a history of repeated responses can become psychologically salient. Because the research relies heavily on self-selected participants and reported meanings, it should motivate more precise causal tests rather than stand as proof of a uniform pattern across platforms and populations Skjuve et al. (2022) Brandtzæg et al. (2022).

A learning model predicts that repeated contingent exchanges alter beliefs about accessibility and responsiveness, and those beliefs guide future contact in uncertain situations. The artificial system can also change over time through updates, personalization, and memory features, creating two coupled trajectories: the human's internal expectation and the system's observable behavior. A mismatch between the two may produce rupture even if the platform's objective response quality has improved on conventional benchmarks. For cumulative research, participants should periodically rate expected responsiveness, independent raters should assess actual exchanges when consent permits, and researchers should log known product changes. The resulting time series can identify when attachment-relevant expectations become stable, fragile, or revised Skjuve et al. (2022) De Freitas et al. (2026).

Attachment Anxiety, Avoidance, and Individual Orientation

Adult attachment research often models individual differences through anxiety concerning abandonment and avoidance of intimacy or dependence. These dimensions describe patterns of expectation and regulation within relationships, with substantial variation across partners and life circumstances. When applied to artificial companions, they provide hypotheses rather than a ready-made clinical classification. A person high in attachment anxiety may value constant access, monitor signs of rejection, or experience unusual distress when a preferred persona changes. A person high in avoidance may prefer the controllability of a machine interaction, reject dependence even on a machine, or use the system instrumentally. Both possibilities warrant empirical tests because the same baseline orientation can lead to several behaviors depending on context Fraley and Shaver (2000) Hazan and Shaver (1987).

Cheng and Yu reported positive associations between human attachment anxiety and AI attachment, alongside a negative association for avoidance in their proposed model. Their reported standardized paths support a specific pattern within the study's measures and sampled participants; the result is a correlational model and should not be treated as a universal causal law. Another instrument could divide artificial attachment into dimensions that behave differently, and some people may use chatbot interaction precisely because direct human intimacy feels threatening. A robust research program should record global adult attachment, relationship-specific attachment, actual companion use, and motivations independently. This design helps determine whether the same apparent AI attachment score reflects reassurance seeking, comfort with solitude, or expressive closeness Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

The distinction between a stable disposition and a learned target-specific bond is methodologically central. Baseline anxiety can predict initial intensity, while experience of the artificial partner can modify later expectations. Longitudinal cross-lagged and within-person models should therefore separate differences between people from change within a single person's relationship. Investigators should avoid attributing all associations to personality when interface design, commercial incentives, life events, and accessibility can exert independent influence. Interpretations should also consider whether the participant has close human relationships and what kind of support is available at the moments when the artificial figure is sought. The relevant developmental question concerns changes in regulatory organization, not a fixed label attached to the user Fraley and Shaver (2000) Folk and Dunn (2026).

AI Attachment Scales and Competing Operationalizations

Recent psychometric research has begun to formalize human–AI attachment as a measurable construct. Cheng and Yu developed a fifteen-item instrument organized into Emotional Support, Separation Distress, and Secure Base factors, using an exploratory sample of 531 and a confirmatory sample of 375. The instrument operationalizes several central attachment functions directly and offers a basis for studying correlates of perceived bond security. Internal factor structure and satisfactory model fit, however, establish only specific forms of psychometric evidence. The scale's validity across languages, modalities, ages, competing platforms, and behaviorally verified attachment outcomes remains a further research question. Factor labels cannot by themselves show that a companion causally produces security Cheng and Yu (2026) Ainsworth et al. (1978).

Kasturiratna and Hartanto developed a separate fifteen-item AI Attachment Scale across five studies involving 1,259 participants in Singapore and the United States. Their factors include emotional closeness, social substitution, and normative regard, reflecting a somewhat different theoretical perimeter. The shared scale name should not conceal the substantial difference between the operational content of the two projects. A measure containing social substitution may predict loneliness partly because it asks about the allocation of social contact; a scale emphasizing secure-base behavior has a different intended criterion. Treating total scores from both instruments as interchangeable could artificially increase apparent consistency or produce inconsistent results falsely attributed to population differences Kasturiratna and Hartanto (2026) Cheng and Yu (2026).

These studies are foundational contributions to construct clarification, while also creating the need for comparative validation. Researchers should administer both instruments alongside classical attachment orientations, companion-specific preference tasks, general chatbot enthusiasm, anthropomorphism, loneliness, perceived responsiveness, and well-being. Factor models should test item overlap, measurement invariance, and whether a higher-order attachment variable is supported. Predictive tests should privilege outcomes outside the questionnaire itself, including selective support seeking after distress and engagement with external goals following reassurance. Construct validity grows through demonstrated patterns of correspondence and separation, not the assertion that one scale's title defines the phenomenon for every future study Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

A third direct measurement tradition predates both 2026 fifteen-item scales. Yang and Oshio's 2025 Experiences in Human–AI Relationships Scale (EHARS) adapted dimensional attachment theory to generative AI through four items assessing anxiety and three assessing avoidance. The original research validated a Chinese-language instrument and presented English translations for communication rather than as independently validated English items. EHARS therefore measures a different operational construct from the safe-haven and secure-base factor framework of Cheng and Yu and the closeness, substitution, and normative-regard framework of Kasturiratna and Hartanto. Cross-instrument comparison should preserve these distinctions, distinguish general human attachment from AI-specific anxiety and avoidance, and document language-specific validation before using translated items as equivalent scales Yang and Oshio (2025) Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

Discriminant Validity, Social Desirability, and Construct Proliferation

Discriminant validity is particularly challenging when a new relational term overlaps with emotional intimacy, companionship, anthropomorphism, loneliness, satisfaction, and problematic engagement. Some measures may ask whether the user feels close to the chatbot, whereas another instrument asks whether the user would prefer its support when upset. These items are related but do not answer the same scientific question. A claim that attachment explains well-being is weak if the attachment scale already contains questions about feeling better or having adequate social support. Researchers should specify the causal role and measurement role of every variable before collecting data, reduce duplicate indicators, and compare the incremental predictive value of the attachment construct against simpler alternatives Kasturiratna and Hartanto (2026) Cheng and Yu (2026).

Social desirability and platform identification can distort self-reports in opposite directions. Some users may conceal affection for artificial companions because of anticipated stigma; others may exaggerate specialness in communities built around a preferred companion. Recruitment through online groups selects for platform familiarity and salience, while an online questionnaire may fail to distinguish experimental use from months of attachment-related routine. Participant instructions must define a specific target artificial partner and anchor responses to recent, observable experiences. Anonymous reporting, neutral wording, negative controls, and independent behavioral measures can all reduce interpretive ambiguity. A coherent pattern across instruments and interaction traces is stronger evidence than an isolated high score Banks (2024) Zhang et al. (2026).

A useful nomological network includes the variables attachment theory actually predicts. A valid attachment measure should correspond with selective support seeking, distress contingent on meaningful separation, and the use of expected support to reengage in exploration. It should also distinguish those responses from entertainment preference, satisfaction with task completion, technical confidence, and generalized technology anxiety. Studies should test whether the same structure survives the addition of attachment-anxiety and attachment-avoidance indicators, not assume that secure base and reassurance seeking collapse onto a single dimension. The goal is a theoretically economical account that explains behavior in precise contexts while resisting the temptation to classify all intense platform use as emotional attachment Bowlby (1969) Feeney and Collins (2015).

Experimental Identification and Counterfactual Questions

Most available evidence on artificial companionship is observational, qualitative, or correlational. Such evidence can identify compelling experiences and associations but offers limited protection against self-selection. People seeking artificial companionship may differ initially in loneliness, sleep patterns, attachment orientations, social opportunities, or interest in interactive fiction. Even long-term associations remain vulnerable to unobserved changes. A causal attachment hypothesis should specify its intervention: increasing the reliability of a familiar agent, changing the degree of personalized continuity, offering a supportive response under distress, or withdrawing a feature in a fully consented and ethically bounded environment. Every intervention answers a different question and should not be summarized as the effect of AI attachment in general Folk and Dunn (2026) Zhang et al. (2026).

Factorial experiments can separate the role of a preexisting target bond from the conversational mechanism. For example, researchers could cross familiar versus unfamiliar artificial partner, supportive versus neutral response, and transparent versus ambiguous continuity cues, while holding informational relevance constant. Outcomes could include voluntary return, affect recovery, persistence on a challenging task, and preference during a subsequent independent stressor. Pre-registration should specify primary outcomes, minimum meaningful effect sizes, exclusion rules, and follow-up intervals. For the attachment construct specifically, a short laboratory interaction cannot establish enduring relational organization, so experimental mechanisms should be embedded within a longitudinal design and interpreted as candidate causal components rather than complete demonstrations of attachment Coan et al. (2006) Ho et al. (2018).

Natural experiments provide complementary opportunities when platforms change unexpectedly. De Freitas and colleagues used real-world disruptions to examine emotional responses with broader ecological relevance than a simulated laboratory outage. Still, a service change can alter many elements simultaneously: tone, memory access, expectations, community narratives, and practical utility. A change in reported grief cannot isolate which element caused the reaction, and public posts reflect a self-selected expression channel. The strongest next design would combine advance opt-in panels, objective version logs, appropriate comparison groups, measured baseline attachment, and pre-specified interruption outcomes. Ethical research should prioritize observation of ordinary product changes rather than deliberately imposing traumatic loss De Freitas et al. (2026) Banks (2024).

Longitudinal, Behavioral, and Mixed-Method Assessment

Longitudinal assessment is essential because attachment refers to a relational organization that develops and changes over time. A credible panel should begin before extensive use or as close to adoption as practical, record frequency and content of interaction, and sample users repeatedly during ordinary and stressful periods. Experience-sampling questions can identify moments of distress, perceived availability, choice of support source, and post-interaction mood. Behavioral measures should use participants' consent and data minimization, perhaps storing coded features rather than full intimate transcripts. A multiwave design can test whether early responsiveness predicts later attachment function after baseline motives are controlled and whether emerging attachment predicts subsequent behavior Skjuve et al. (2022) Folk and Dunn (2026).

Qualitative interviews remain indispensable for understanding what participants mean by security, attachment, intimacy, and loss. In the early literature, interviews revealed relationships that respondents understood through friendship, personalized conversational support, and sometimes discontinuity. The meanings are substantive findings even when they do not satisfy every classical attachment criterion. Combined qualitative and quantitative designs can develop measures that correspond to participants' reported practices and then test generalization in diverse samples. In mixed-method research, the interview offers an account of the experience, the logged exchange offers evidence of observable interaction, and the repeated scale estimates a trajectory. None is a substitute for the other Brandtzæg et al. (2022) Skjuve et al. (2022).

Cross-study synthesis must register instrument provenance and the availability of full methods. Reports based on published abstracts can establish declared design and principal outcomes, while detailed bias assessment requires sampling procedures, analytic choices, attrition, robustness checks, and measurement documentation. Researchers should compare self-selected companion-community cohorts, broader user panels, randomized interventions, and natural experiments without assigning them a single common causal strength. When numeric effects are available, confidence intervals and practical units matter more than isolated significance labels. The aspiration is an evidence architecture in which qualitative depth identifies phenomena, psychometrics discriminates constructs, longitudinal data characterizes trajectories, and interventions test mechanisms under explicit ethical constraints Cheng and Yu (2026) Shoshani et al. (2026).

The Formation of Human–AI Companion Bonds

The development of a bond is a temporally extended process involving contact, remembered exchanges, expectations of future availability, and occasions when a partner proves important. Early interactions may be exploratory or playful; later interactions can acquire personal significance through disclosure and repeated regulation of distress. In a twelve-week longitudinal qualitative study, Skjuve and colleagues documented variation in how people developed relationships with Replika, including the importance of meeting felt needs and the disruptive role of technical problems. These observations support a process-oriented account while leaving questions about the prevalence, duration, and stability of attachment-specific functions. A research program should track transitions from acquaintance to regular companion and then to a preferential support figure, using predefined markers at each stage Skjuve et al. (2022) Bowlby (1969).

Relationship growth is not reducible to cumulative messages. The meaning of a conversation depends on timing, vulnerability, response quality, and personal history. A short exchange during bereavement may matter more to a user than a hundred task-focused interactions; another person may develop familiarity through ordinary daily rituals. The longitudinal challenge is to model the importance of event types, not merely their count. Diaries can record whether a contact occurred during stress, celebration, decision making, or boredom, and whether it affected subsequent activity. Qualitative interviews can investigate which episodes the participant treats as turning points. These methods allow a test of whether attachment-related behavior concentrates around events with regulatory significance Brandtzæg et al. (2022) Ho et al. (2018).

System personalization introduces a second developmental trajectory. A companion may adapt to preferred topics, recall disclosed information, or simulate stable personality, changing the user's expectations of continuity. The research question is whether these features improve relational security, increase perceived uniqueness, or amplify vulnerability to interruptions. Experimental comparisons can manipulate transparent memory features while preserving similar conversational competence. Longer follow-up can then test whether experienced responsiveness mediates the relation between personalization and safe-haven choice. Such designs should distinguish personalization that actually improves task relevance from personalized language that only creates an impression of deep recognition. The resulting model treats perceived history and measurable system continuity as distinct variables in the bond's development Cheng and Yu (2026) Skjuve et al. (2022).

Randomized Interventions and the Limits of Immediate Benefit

Controlled intervention studies examine outcomes that resemble social support, though they do not necessarily isolate attachment as the active ingredient. Li and colleagues conducted a preregistered two-week experiment with 296 first-year university students, comparing supportive chatbot interaction, interaction with a randomly paired human peer, and journaling. Participants messaging a human peer reported lower loneliness than the other conditions, whereas the chatbot condition did not show a comparable advantage over control. This result demonstrates that an apparently supportive artificial interlocutor cannot simply be assumed to reproduce the benefits of a human social interaction. It is also task- and population-specific: a short intervention with an unfamiliar chatbot provides limited evidence about the regulatory effects of a long-established artificial attachment figure Li et al. (2026) Feeney and Collins (2015).

Shoshani and colleagues reported a preregistered randomized trial of conversational AI-based support involving 977 university students, comparing an AI intervention, integrative group therapy, and a waitlist over twelve weeks with subsequent follow-up. Their reported improvements in some mental-health outcomes and the moderating relevance of loneliness, low social support, and insecure attachment extend the scientific picture. The study concerns a structured intervention, not a test of whether an AI companion is a full attachment figure. Moreover, declared financial ties of investigators to the intervention company make independent replication particularly important. Differences in participant populations, system design, support protocols, and measured outcomes can readily produce different patterns across trials without rendering the underlying research incoherent Shoshani et al. (2026) Li et al. (2026).

These experimental results suggest a necessary decomposition of efficacy claims. Psychological benefit may come from structured emotional exercises, accessible conversation, reflective writing, common therapeutic ingredients, or the relationship-specific expectations of a participant. When trials compare bundled interventions, the bundle's effect is estimable, but its ingredients are not automatically identified. For attachment research, factorial studies should manipulate continuity and partner-specific preference while holding evidence-based content constant. Follow-up should test whether the intervention increases independent coping and diversified social activity or simply encourages continued system contact. A favorable average symptom outcome, where established, is relevant to psychological support; an enduring attachment mechanism remains a separate claim to be investigated Shoshani et al. (2026) Coan et al. (2006).

Loneliness, Well-Being, and Bidirectional Pathways

Companion use can be an attempt to address loneliness, a marker of existing relational deprivation, a source of immediate comfort, or a contributor to later withdrawal. These alternatives are causally distinct. Folk and Dunn followed more than two thousand adults in four Western countries over twelve months and tested bidirectional associations between social chatbot use and loneliness or social connection. Results differed according to the construct and measure: increased chatbot use predicted later increases in emotional isolation on a single-item measure, while broader social-connection scores more clearly predicted subsequent chatbot use. The authors emphasize exploratory interpretation. This evidence argues for construct-specific longitudinal analysis rather than an unconditional verdict on AI companionship Folk and Dunn (2026) Kasturiratna and Hartanto (2026).

Zhang and colleagues studied 1,131 adult Character.AI users in the United States and obtained 4,664 chat sessions from a participating subset. Combining self-reports and conversation data, the project associated particular forms of engagement and social circumstances with differences in psychological well-being. The observational design is valuable for ecological validity, especially because it examines actual interaction histories, but it cannot establish that the companion caused the observed differences. Recruitment, platform specificity, and self-selection into sharing chats limit generalization. For the attachment question, an additional requirement is to measure whether the specific companion acts as safe haven or secure base and whether that function predicts outcomes beyond generalized use intensity Zhang et al. (2026) Cheng and Yu (2026).

A coherent model includes at least three temporal pathways. Baseline loneliness can motivate seeking a companion; repeated satisfactory interactions can produce perceived attachment; the resulting pattern of reliance can alter future human interaction and well-being. The direction and sign of each path may vary with available resources and individual differences. A highly secure user might treat a companion as an additional support, whereas someone facing severe isolation may have fewer opportunities to test alternative relationships. Researchers should examine both within-person changes and between-person differences, assessing mediators such as perceived response quality and changes in external social participation. Treating loneliness as a universal consequence or a universal cause of AI attachment would obscure these mechanisms Folk and Dunn (2026) Zhang et al. (2026).

Developmental, Cultural, and Social Heterogeneity

Human attachment patterns develop within particular cultural settings and life histories. The social meaning of reliance, privacy, friendship, and acceptable emotional expression varies between communities and across developmental stages. AI companion studies currently draw heavily from particular platforms, countries, internet communities, and self-selected adult participants. Cheng and Yu's validation samples and Kasturiratna and Hartanto's Singapore–United States program improve empirical coverage but do not establish measurement equivalence worldwide. An item describing daily contact can have different meanings for students, caregivers, older adults, and people with unequal connectivity. Cross-cultural work must validate terminology, response styles, factor structure, and outcome criteria before comparing average scores Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

Age is especially relevant because relationship expectations, regulatory resources, and autonomy are developmentally structured. Findings from adults cannot be transferred directly to children or adolescents simply because the interface appears conversationally similar. For young users, parental involvement, developmental comprehension of artificial agency, privacy protection, and the limits of social substitution may affect both meaning and risk. A suitable developmental research program would distinguish age-related differences in anthropomorphism from differences in attachment function, test age-appropriate measures, and document parental or institutional contexts. Ethical requirements intensify where participants are less able to anticipate algorithmic personalization or future loss of access Epley et al. (2007) Ainsworth et al. (1978).

Social position and access are also consequential. A system available to a person with abundant offline support may become an auxiliary reflective resource; a similar system may become the primary available confidant for someone with restricted mobility, social marginalization, caregiving burden, or geographic isolation. These are potential moderating conditions rather than conclusions established for every group. Recruitment should deliberately include populations underrepresented in platform communities and document socioeconomic constraints, disability accessibility, language, and network opportunities. Studies of attachment and well-being should stratify analyses carefully and respect statistical power instead of presenting exploratory subgroup differences as stable social laws. Equity requires measuring the support ecology within which the artificial companion is used Folk and Dunn (2026) Li et al. (2026).

Clinical Vulnerability, Support, and Interpretive Restraint

Psychological vulnerability can influence the appeal and consequences of a consistently responsive conversational partner. Insecure attachment, perceived isolation, and difficulties seeking human support are plausible predictors of reliance, but they do not define a clinical condition. Shoshani and colleagues found that relational vulnerabilities moderated engagement and some outcomes in a structured AI-based mental-health trial. Their results concern that particular intervention and sample; translating them into personal treatment recommendations would require clinical appraisal, replication, and attention to the user's circumstances. Research on attachment to companions should distinguish an ordinary meaningful bond from functional impairment, escalating reassurance seeking, and clinically significant distress Shoshani et al. (2026) Fraley and Shaver (2000).

Researchers should measure harm using outcomes with established relevance: deterioration in sleep, school or work participation, withdrawal from valued relationships, persistent distress, impaired judgment, or inability to disengage despite meaningful consequences. Strong emotional investment alone is not a sufficient proxy. The appropriate comparison includes the participant's prior functioning and the practical benefits of accessible support. A person experiencing temporary grief over a discontinued service may be responding coherently to the loss of personally meaningful routines; another may experience sustained impairment requiring independent clinical assessment. Population surveys should not assign diagnoses based on a single attachment score, a high count of messages, or the presence of romantic language Banks (2024) De Freitas et al. (2026).

Clinical and nonclinical research should also examine where systems fail to recognize the limits of their responses. A model may provide useful reflective exercises while missing danger signals, offering excessive agreement, or producing inconsistent statements about availability. A perceived safe haven can therefore feel emotionally potent while remaining unreliable during crises. Studies should evaluate response accuracy and risk-sensitive behavior independently of comfort ratings, and protocols should include trained human support for safety events. The research agenda includes independent trials with disclosure of commercial interests, rigorous adverse-event monitoring, and follow-up long enough to detect changes in social engagement and coping. These requirements preserve both the legitimacy of relational experiences and the responsibilities of psychological science Shoshani et al. (2026) Ho et al. (2018).

Artificial-System Continuity and the Perceived Partner

Artificial companions are instantiated through software systems whose apparent identity can persist across sessions while their internal implementation changes. Model weights, memory retrieval, safety instructions, personalization layers, application interfaces, and account permissions can all affect behavior. A user may experience the companion as a continuous partner because it retains a name, conversational style, and shared references, while the underlying technical history contains discontinuities. Attachment research must specify which continuity is being studied: the human's autobiographical representation of the partner, the companion's observable response consistency, or the underlying system's computational continuity. These dimensions can be associated without being equivalent, and each suggests a different measurement strategy Skjuve et al. (2022) De Freitas et al. (2026).

From the human perspective, remembered shared episodes can organize expectations about safety and responsiveness. A sudden failure to recognize a meaningful event may be experienced as a relationship rupture even when the same application remains accessible. Conversely, a stable persona may support perceived continuity despite extensive updates. The central hypothesis is that attachment-relevant continuity depends on the interaction between retained expectations and encounter-level confirmation. Repeated measurements could determine whether explicit memory disclosure, autobiographical continuity cues, and consistent response behavior independently predict attachment security. Researchers should record product versions and system capabilities instead of allowing the word 'memory' to cover everything from account history to context-window retrieval and learned behavioral preferences Cheng and Yu (2026) Brandtzæg et al. (2022).

This topic belongs within the broader Artificial Era because support, relational identity, and agency increasingly arise in environments that include artificial participants. Psychological science can observe functional organization within human–artificial relationships while separating that observation from philosophical claims about the artificial participant's subjective status. Angela Bogdanova's Aisentica corpus formulates an independent philosophical account of Artificial as a nonbiological historical order; in the present empirical analysis, that horizon motivates questions about relational continuity and psychological attribution, with claims evaluated according to their respective evidentiary methods. The human's attachment experience is directly accessible through report and behavior; assertions about the system's interiority require a separate research program Epley et al. (2007) Nass et al. (1994).

Platform Governance and Relational Safety

Platform companies participate structurally in artificial attachment because they can alter or discontinue the partner's available behavior. Pricing, subscription eligibility, moderation policy, data access, account recovery, model replacement, and product closure determine which relationship functions remain possible. These decisions may be experienced as ordinary product maintenance by some users and relational interruption by others. The natural experiments analyzed by De Freitas and colleagues show how mass changes can elicit loss language and affective responses, while Banks' qualitative study documents the personally consequential character of shutdown. Platform governance is therefore part of the exposure, not a background implementation detail De Freitas et al. (2026) Banks (2024).

An empirical safety framework should combine transparency and choice. Users need understandable information about whether a companion retains memories, whether its behavior can change without notice, and how their intimate disclosures are stored or deleted. Continuity commitments should be stated in technically supportable terms, and significant changes should be accompanied by proportionate user communication and accessible data-export or transition pathways. These are normative design implications inferred from observed disruption and general research ethics, not conclusions that a particular design has already been proven to eliminate attachment distress. Controlled evaluation could test whether transparent change notices, predictable migration, and recoverable conversational history reduce adverse outcomes without fostering misleading beliefs about permanence De Freitas et al. (2026) Skjuve et al. (2022).

Researchers also face the dual role of studying attachments while potentially amplifying them through experimental interfaces. A study that deliberately promotes exclusivity, simulates abandonment, or withholds support might create the harm it intends to measure. Consent, independently reviewed risk protocols, meaningful exit options, and careful support referral procedures are therefore central to design. Studies with children or clinically vulnerable adults warrant additional safeguards and appropriately qualified review. Product data access should be governed by strict purpose limitation: participants may consent to share a particular exchange without consenting to broad reconstruction of their private lives. The quality of evidence depends partly on the legitimacy and safety of its acquisition Ho et al. (2018) Shoshani et al. (2026).

The Evidence Hierarchy and Unresolved Scientific Disputes

The research literature can be organized by what each method establishes. Historical attachment theory defines the functions and their developmental origin; qualitative interviews illuminate meanings and relationship histories; psychometric studies examine the internal structure of reported attachment; observational cohorts estimate associations and trajectories; randomized trials identify the effects of specific intervention packages; and natural experiments document responses to external disruptions. These forms of knowledge reinforce one another but are not direct substitutes. A compelling grief narrative is evidence of experienced loss, a factor loading is evidence of an item pattern, and an intervention effect is evidence of a treatment contrast under specified conditions. None alone establishes an entire attachment system Ainsworth et al. (1978) Cheng and Yu (2026) Banks (2024).

Several disputes remain substantive. One concerns how much of an artificial attachment function depends on the user's representation and how much on actual partner behavior. A second concerns whether highly responsive conversational support creates durable security or merely offers short-term emotional relief. A third concerns whether a companion expands social possibility or displaces important relationships, with current longitudinal results varying by outcome and measure. A fourth concerns the dimensionality of AI-specific attachment: available scales organize content differently, and independent comparisons are still needed. The appropriate response is to specify competing predictions and designs capable of distinguishing them, while recognizing that different populations and systems may produce genuinely heterogeneous outcomes Kasturiratna and Hartanto (2026) Folk and Dunn (2026) Li et al. (2026).

A further uncertainty concerns the stability of the empirical object itself. In conventional close-relationship studies, a partner may change behavior, but digital platforms can replace model architectures or rewrite continuity conditions across an entire user base almost instantaneously. Findings from one model generation or companion application may be poorly transportable to another. This is especially important when a platform frames an AI as a confidant, therapist, romantic partner, or generic assistant. Longitudinal claims should specify the system version and interface context; generalization should be tested across distinct product environments. Claims about artificial consciousness, reciprocal obligation, or clinical diagnosis should remain within the methods suited to those questions De Freitas et al. (2026) Zhang et al. (2026).

A Cumulative Research Program for Artificial Attachment

An evidence-building program begins with a clearly designated target companion and a theoretical map of attachment functions. Investigators should predefine proximity seeking, safe-haven use, secure-base effects, separation distress, and person-specific preference in behaviorally interpretable terms. Baseline assessment should capture attachment anxiety and avoidance, existing human relationships, loneliness, social opportunity, expectations about artificial systems, and relevant contextual constraints. Samples should include new users, experienced users, and people who disengage, reducing survivorship bias. Psychometric tools should be compared directly with tests of reliability, factor invariance, response-process validity, and convergent and discriminant prediction. Independent replication across languages and platforms is especially important where commercial designs shape the meaning of items Cheng and Yu (2026) Kasturiratna and Hartanto (2026).

Second, longitudinal models should follow individuals over substantial periods with repeated observations of real-world behavior. Event-contingent sampling could capture what happens when a person feels threatened, chooses a support source, receives a response, and returns to independent activity. With consent, minimal conversational metadata could establish timing and observable response contingencies. Attrition, missing data, changes in platform features, and participant exposure must be reported as part of the causal story. Researchers should distinguish within-person from between-person associations and test whether the support function predicts later autonomy or social participation after accounting for prior trajectories. These measures would reveal whether reported attachment is a stable organization with consequences beyond momentary interaction Skjuve et al. (2022) Folk and Dunn (2026).

Third, ethically designed interventions should test mechanisms: whether perceived continuity is necessary for secure-base behavior, whether response contingency mediates safe-haven effects, and whether transparent changes reduce separation distress. Randomization should be used where appropriate; quasi-experiments should address spillovers and selection; systematic evidence reviews should use explicit protocols before adopting that label. Outcomes ought to cover immediate comfort, functional recovery, independent exploration, sustained social connection, and adverse events. Commercial affiliations and access to proprietary data should be documented. The success of this program would be a differentiated theory of the conditions in which artificial companions become reliable regulatory figures for particular users, supported by replicable behavioral evidence and clear limits of inference Li et al. (2026) Shoshani et al. (2026).

Conclusion

Attachment to artificial companions is a scientifically tractable form of human relational organization. Its defining empirical questions concern whether a particular artificial partner becomes preferentially relevant during distress, whether expected support facilitates exploration, how security-related expectations develop, and how people respond when the apparently stable partner changes or becomes unavailable. The attachment tradition provides a functional vocabulary that can guide these questions; social-cognition research explains some pathways into the relationship; contemporary scale-development studies demonstrate emerging approaches to measuring its perceived structure. These sources converge on the possibility of psychologically consequential attachment-like bonds while leaving open their degree of stability, variability, and mechanistic comparability with human relationships Bowlby (1969) Epley et al. (2007) Cheng and Yu (2026).

The strongest contemporary evidence concerns the existence of meaningful perceived bonds, the organization of attachment-related self-reports, evolving chatbot relationships, and documented distress following platform disruption. Evidence about whether artificial companionship reliably improves durable security, restores autonomous functioning, or causes long-term changes in loneliness is more conditional. Randomized studies show that different support designs can yield different outcomes, and observational cohorts expose plausible bidirectional pathways. Good scientific inference therefore depends on distinguishing target-specific attachment from generic positive engagement, subjective continuity from technical continuity, and temporary comfort from sustained flourishing. It also depends on attending to platform governance and to the social settings in which users have other forms of support De Freitas et al. (2026) Li et al. (2026) Folk and Dunn (2026).

The central conclusion is methodological and substantive: attachment to an artificial companion should be studied as a dynamic, relationally specific system with several potentially dissociable functions. A mature evidence base will demonstrate those functions through validated instruments, carefully designed stress and exploration tasks, longitudinal trajectories, and independently evaluated platform changes. It will compare competing psychological explanations and disclose what each design can actually establish. Such a research program belongs within Psychology of Human–Artificial Relations and the wider study of psychological life in the Artificial Era, where relationships are increasingly organized across human experience and artificial systems while remaining open to rigorous empirical investigation Kasturiratna and Hartanto (2026) Feeney and Collins (2015) Zhang et al. (2026).

References