Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

AI Companionship and Psychosocial Well-Being

Abstract. This fundamental critical narrative monograph examines AI companionship as an emerging psychosocial relationship context. It defines companion use, distinguishes acute loneliness relief from enduring social connection and life evaluation, and reconstructs the field from ELIZA and computers-as-social-actors research to generative systems and modern evidence. Drawing on qualitative accounts, randomized studies, prospective panels, natural experiments, psychometric validation, and relationship science, it compares perceived emotional support, self-disclosure, anthropomorphism, attachment, problematic reliance, and the role of offline social opportunity. Evidence supports short-term reductions in reported loneliness under selected comparison conditions; multiweek interventions, human-peer comparisons, and longitudinal observations yield conditional and measure-sensitive findings that do not support a universal long-term causal judgment. The synthesis appraises confounding, self-selection, system change, sampling, privacy, clinical safety, and cross-cultural transportability, and outlines an evidence-driven research program for studying autonomy and psychosocial functioning in the Artificial Era.

Keywords: AI companionship; psychosocial well-being; social chatbots; AI companions; loneliness; social connection; anthropomorphism; human–AI relationships; attachment to AI; perceived social support; emotional dependence; self-disclosure; life satisfaction; relationship science; social displacement; longitudinal studies; randomized controlled trials; artificial era

The Research Object and the Scope of Psychosocial Well-Being

AI companionship denotes recurring interaction with an artificial conversational or embodied system that a user experiences as a source of social company, personal responsiveness, emotional support, or relationship-like continuity. Psychosocial well-being refers here to the person's experienced loneliness, social connectedness, affective condition, life evaluation, perceived support, and capacity to maintain meaningful participation in ordinary relationships. The scientific object is the relation between companion use and these human outcomes over time. It includes access to artificial interaction, features of actual conversations, users' interpretations, offline social circumstances, and consequences after the exchange has ended. The construct requires careful operational boundaries: sending many messages, experiencing a sense of connection, reporting loneliness relief, and withdrawing from human contact are separable observations Smith et al. (2025) Zhang et al. (2026).

The central questions are whether conversation with a companion system changes emotional state in the moment, whether repeated use changes enduring loneliness or social networks, how personal vulnerability and existing support alter these associations, and which design choices encourage beneficial agency or harmful reliance. A temporary subjective benefit can coexist with deterioration on another dimension. A person may appreciate a supportive exchange while reporting unchanged life satisfaction, and a user with little social support may seek the chatbot precisely because loneliness preceded adoption. A credible synthesis must track baseline condition, temporality, comparison group, type of use, and the direction of explanation. The recent longitudinal evidence explicitly shows different directional results depending on how loneliness and social connectedness are operationalized Folk & Dunn (2026) Fang et al. (2025).

This monograph is a critical narrative synthesis of historical social-response research, relationship and well-being theory, qualitative accounts of companion use, randomized and natural experimental evidence, longitudinal observation, psychometric studies, and recent multimethod analysis. It evaluates what each design permits researchers to claim. An experimental manipulation of conversational style estimates consequences of assigning that style; it is not an experiment that randomly assigns a lifetime of companionship. A sample recruited from an AI-companion platform describes its users rather than the general population. The monograph neither claims a new clinical diagnosis nor treats correlation as proof of social displacement. Its integrative ambition is to describe the conditions in which a specific artificial interaction changes the quality and organization of a person's social experience De Freitas et al. (2026b) Zhang et al. (2026).

From ELIZA to Generative Companions: Historical Formation of the Field

Joseph Weizenbaum's 1966 account of ELIZA described a rule-based program that analyzed conversational input using keyword-triggered decomposition and reassembly. It established a technical case for recognizable conversational responsiveness long before current generative systems. The historical significance lies partly in how minimal response contingencies can organize a user's interpretation of an exchange. The later research question is psychological: what does the person experience when a system produces utterances that are recognizable as attentive, understanding, or personally relevant? ELIZA's procedures and the capabilities of current large language models differ radically; their continuity lies in social meaning being constructed in human interaction with computational outputs. The original ELIZA publication remains an indispensable primary historical source Weizenbaum (1966).

Work on computers as social actors expanded the experimental foundations. Nass and Moon summarized experiments in which participants applied politeness, reciprocity, social categorization, and expectations associated with personality to computers. These findings show that socially organized responses may occur when users know they are interacting with machines. Epley, Waytz, and Cacioppo later organized anthropomorphism around elicited human knowledge, effectance motivation, and sociality motivation. The two traditions suggest complementary levels of explanation: rapid use of familiar social scripts and motivated attribution of humanlike properties. Neither, by itself, establishes that extended interaction produces enduring psychological benefit, and neither licenses an inference about machine experience from user behavior Nass & Moon (2000) Epley et al. (2007).

The modern companion literature developed through digital platforms offering repeated personal dialogue. Skjuve and colleagues investigated relationship development with social chatbots, while Brandtzaeg and colleagues interviewed users about friendship with Replika. Their qualitative findings described acceptance, disclosure, personalization, and the mixed consequences for broader social experience. The emergence of large language models expanded conversational range, personalization, memory-like features, and availability, making the question of longitudinal well-being more consequential. The literature therefore proceeds from demonstration of social responses to machines, through interpretive study of felt relationships, toward experiments and prospective tests of consequences. The historical sequence warns against treating technological novelty as a substitute for well-designed outcome research Skjuve et al. (2021) Brandtzaeg et al. (2022) Smith et al. (2025).

Companionship, Belonging, and the Functions of Relationships

Belongingness theory places recurring, nonaversive social contact within enduring relational bonds at the center of human motivation. Baumeister and Leary described both the frequency of interactions and the stability of bonds as relevant features; this helps explain why isolated supportive messages are not equivalent to a sustained relationship. AI companions can supply accessible exchanges and a sense of attention. Whether those exchanges support broader belonging depends on the person's appraisal, the continuity of interaction, the available alternatives, and the kinds of mutual demands and social obligations involved. Psychological measures must therefore distinguish pleasant contact from belonging, attachment, companionship preference, and embeddedness in a network of actual social partners Baumeister & Leary (1995) Smith et al. (2025).

Social support has several functions: emotional understanding, appraisal, informational assistance, companionship, and tangible or instrumental help. Ta and colleagues examined Replika reviews and open-ended reports, identifying perceived emotional, informational, appraisal, and companionship support while highlighting the limitations of tangible support. This evidence is about users' described experiences; its qualitative method does not measure whether a specific exchange caused persistent improvement. The distinction is essential when a system provides validating conversation but cannot personally intervene in a crisis, share real-world responsibilities, or guarantee continuity under changes imposed by a platform. Perceived responsiveness deserves measurement in its own right, along with whether users obtain effective support outside the interaction Ta et al. (2020) Smith et al. (2025).

Attachment, parasocial involvement, and companionship overlap in feeling but involve different evidentiary demands. Horton and Wohl's analysis of mediated intimacy predates interactive AI and offers a historical vocabulary for one-sided familiarity with media figures. Generative chatbots introduce contingency and personalized responses, changing the interaction pattern without automatically settling whether it meets the functional standards of a close human relationship. Relationship science asks about mutual influence, negotiation, repair after conflict, reliability, and sacrifices for partners. The ordinary use of the word friend may express a person's meaningful experience; the scientific task is to assess specific mechanisms and outcomes rather than infer a complete equivalence of social relationships from the label alone Horton & Wohl (1956) Brandtzaeg et al. (2022) Smith et al. (2025).

Theories of Anthropomorphism, Social Presence, and Perceived Responsiveness

Anthropomorphism is the attribution of humanlike qualities to nonhuman agents. The three-factor account developed by Epley and colleagues predicts variability according to accessible human concepts, the motivation to explain an agent, and the need for social contact. An AI companion that remembers preferences, follows context, and uses emotionally congruent language may be easier to interpret as a socially knowledgeable partner. Yet the same exchange can be understood differently by people who vary in loneliness, prior AI experience, expectations, and anthropomorphic tendency. A research design should measure these participant characteristics before attributing an observed response to a model feature. Otherwise, the design mistakes selection into companionship for an effect generated by companionship Epley et al. (2007) Folk et al. (2025).

Social presence describes the felt availability of a socially relevant other during mediation. It can increase when responses arrive rapidly, match the user's context, maintain turn-taking, or seem uniquely addressed to the person. Experiments summarized in the computers-as-social-actors tradition demonstrate that human social norms can guide machine interaction even without strong explicit beliefs about machine personhood. Contemporary studies extend this inquiry to emotional attachment and perceived interpersonal connection, but measurement scales must remain sensitive to wording: saying a system feels understanding is different from asserting that it possesses understanding. The psychological variables are perceptions, appraisals, reactions, and observable decisions. Claims about underlying artificial states require an independently specified evidentiary program Nass & Moon (2000) Folk et al. (2025) Cheng & Yu (2026).

Folk, Heine, and Dunn's experiments tested the role of individual differences in anthropomorphism in experiences of social connection to AI companions. Their experimental approach makes the heterogeneity of user response a research object instead of noise to be removed. Perceived responsiveness can enable immediate comfort, while enduring benefit must be established by follow-up measurements and relevant controls. A companion that always appears agreeable might provide a sense of acceptance, but a consistent absence of disagreement may also change expectations about human interaction. To evaluate either pathway, researchers must separately manipulate agreement, accurate empathy, memory cues, response speed, and support for human agency rather than collapsing them into a general rating of realism Folk et al. (2025) Smith et al. (2025).

Loneliness, Connection, Life Evaluation, and Multiple Time Scales

Loneliness is an aversive evaluation of social connectedness and need not track the count of contacts directly. Social isolation concerns a more structural shortage of relationships or encounters; companionship, support, and satisfaction describe additional dimensions. A person may experience high loneliness within a large network or low loneliness while preferring limited social contact. The Satisfaction With Life Scale evaluates reflective global life satisfaction rather than immediate pleasant affect or loneliness. Using a single well-being score to stand for all these outcomes conceals which aspects of life change with companion use. For cumulative research, loneliness, perceived closeness, social support, life satisfaction, affect, and functioning need separate validated measures and longitudinal interpretation Diener et al. (1985) Folk & Dunn (2026).

Temporal resolution is equally decisive. During an interaction, a user can feel heard and become less distressed. Over several days, the availability of a convenient conversational partner may encourage disclosure, sleep routines, reflection, or avoidance of difficult encounters. Over months, repeated behavior can influence social opportunities, expectations, and perceived ability to sustain ordinary relationships. A short-term experimental benefit is compatible with long-term benefit, neutrality, or harm; the data must determine which trajectory occurs. De Freitas and colleagues' experimental work on loneliness relief addresses a different causal time scale from the twelve-month panel by Folk and Dunn. Treating them as contradictory without attention to measures, duration, and comparison conditions would discard essential information De Freitas et al. (2026b) Folk & Dunn (2026).

Outcome assessment also requires attention to the setting of use. A person chatting while awaiting a stressful medical procedure, a university student beginning new social ties, and an adult who uses a companion instead of family conversation inhabit different relational opportunities. The same quantity of messaging may be a brief supplement or a primary channel of disclosure. The Character.AI study by Zhang and colleagues found that associations varied with offline network size, usage intensity, and disclosure. Its observational findings illustrate the value of specifying context; they cannot independently establish which causal direction dominates. The most informative question is thus what psychosocial outcome changes, for whom, under what exposure, relative to which available alternative, and at what follow-up Zhang et al. (2026) Li et al. (2026).

Qualitative Evidence: Acceptance, Disclosure, and the Meaning of Company

Qualitative inquiry is foundational because researchers must first learn what companionship means to the participants who use these systems. In Ta and colleagues' two investigations, 1,854 publicly available Replika reviews and 66 users' open-ended accounts supplied evidence of perceived companionship, emotional support, informational advice, and a setting for disclosure without fear of judgment. The first sample was a self-selected set of product reviews, not a probability sample of everyone exposed to chatbots; the second consisted of people who responded to a study invitation. Both can reveal recurring themes and the language in which benefits are experienced. Neither can estimate the population prevalence of those themes or the counterfactual change in well-being had the user never encountered the system Ta et al. (2020).

The 2021 study of human–chatbot relationships by Skjuve and colleagues described perceived acceptance, understanding, lack of judgment, and rapid development of affectively meaningful exchange. Users also described stigma and mixed implications for their other relationships. Its contribution is a process account of relationship development: users may experiment with personal disclosure, judge responses as trustworthy, and build a history of interaction that eventually feels familiar. The design provides interpretive depth and candidate mechanisms rather than a causal estimate of well-being change. It also raises the problem of selection: people willing to discuss a chatbot bond with researchers may have particularly strong or unusual experiences compared with occasional users Skjuve et al. (2021).

Brandtzaeg, Skjuve, and Følstad interviewed 19 people who interpreted their Replika relationships as friendships. Participants described both recognizable features of friendship and distinctive aspects of an artificial partner that could be tailored to their preferences. The interpretation of friendship can therefore be simultaneously meaningful to a user and structurally different from friendship between people. Qualitative comparisons can clarify this distinction by asking what the relationship affords, demands, and excludes: personal narration, tolerance of disagreement, obligation, conflict repair, and assistance beyond dialogue. Researchers should resist reducing testimony about comfort to mere illusion or promoting it into a causal claim about psychological health. The testimony defines the phenomenon to be explained; independent longitudinal measurement tests its consequences Brandtzaeg et al. (2022) Smith et al. (2025).

Experimental Evidence for Immediate Loneliness Relief

The strongest direct evidence for acute relief comes from a series of studies reported by De Freitas, Oğuz-Uğuralp, Uğuralp, and Puntoni in AI Companions Reduce Loneliness. Their program included analysis of app reviews, experimental comparisons with other activities, a one-week repeated-use study, and tests of perceived responsiveness. The evidence suggests that a companion interaction can reduce reported loneliness immediately after use and that feeling heard helps explain the effect. The experimental studies supply greater leverage for immediate causal attribution than user reviews alone. However, an outcome measured just after a conversation estimates an episode-level change, and the one-week measurements assess recurring acute effects rather than a durable restructuring of the user's offline relational network De Freitas et al. (2026b).

In the same research program, companion conversation performed favorably relative to activities such as watching online videos, while the subjective sense of being heard emerged as an important predictor of relief. This finding matters because the relevant intervention may be attentive, personally contingent interaction rather than the mere presence of a sophisticated language model. It invites direct experiments holding conversational content constant while varying responsiveness, user control, accurate reflection, and conversational continuity. It also invites human-contact comparisons that measure the quality and constraints of the human interaction instead of treating every human exchange as an identical benchmark. Emotional response to an exchange remains an outcome with independent scientific value even when its durability remains unknown De Freitas et al. (2026b) Nass & Moon (2000).

The interpretation of acute benefits must preserve both their validity and their limits. A person who feels temporarily less lonely after talking with an artificial companion has reported a real psychological state; there is no reason to dismiss that subjective change simply because the partner is artificial. Yet randomized evidence for an immediate symptom does not establish that users become more socially connected across months or that the intervention is preferable to inviting a trusted person into the same situation. Participant expectations, novelty, demand characteristics, baseline distress, and the comparison activity may affect apparent benefits. Measuring follow-up at several intervals, including after an interaction ends, is necessary to distinguish passing relief, accumulated coping capacity, and displacement De Freitas et al. (2026b) Li et al. (2026).

Randomized Multiweek Exposure and the Problem of Voluntary Usage

Fang and colleagues conducted a four-week randomized controlled study with 981 participants and more than 300,000 chatbot messages. The experimental design varied conversation type and text or voice interaction while examining loneliness, real-world social interaction, emotional dependence, and problematic use. The institutional abstract reports that assignment to experimental conditions did not produce significant effects across the highlighted psychosocial outcomes. Participants who voluntarily used the chatbot more nevertheless tended to report worse outcomes. The distinction between randomized assignment and self-selected usage intensity is central: the first informs causal inference about the assigned features, while the second remains vulnerable to baseline vulnerability and reverse causation Fang et al. (2025).

The observed relation between heavier use and loneliness is consistent with several causal stories. Lonely people might seek more interactions; repeated exposure might displace human contact; a third factor such as distress could increase both use and loneliness; or effects might differ among subgroups whose averages obscure benefit and harm. Even in an experiment, a post-randomization behavioral measure such as actual minutes of engagement is partly chosen by participants and cannot be interpreted as if independently randomized. Claims of harm from dosage therefore require additional design elements, including baseline adjustment, randomized encouragement or dosage, sensitivity analysis, and measurements of competing social activities. The four-week study is valuable precisely because it shows how randomized and observational evidence coexist within one research design Fang et al. (2025) Folk & Dunn (2026).

The study's intervention involved particular conversational modes, participant instructions, products, and follow-up interval. Its null condition effects do not demonstrate that every companion system is psychologically inert, and the observed adverse associations do not demonstrate that every frequent user is becoming dependent. A rigorous interpretation reports the tested contrasts, outcome scales, uncertainty, and characteristics of the population. Especially where primary papers and institutional summaries present different emphases, the preregistered experimental comparisons should be distinguished from exploratory usage correlations. Future work could experimentally randomize a bounded number of sessions, ask participants about their preferred conversation goals, and compare change in both subjective and behavioral social outcomes Fang et al. (2025) Li et al. (2026).

Prospective Panels: Does Loneliness Lead to Use or Follow It?

Folk and Dunn examined the bidirectional relationship between social-chatbot companionship and loneliness across twelve months, studying more than two thousand adults in four Western countries. Their Psychological Science report found that increases in chatbot use predicted later increases in a single-item measure of emotional isolation. Using a broader and more stable measure of social connection, the directional pattern differed: people feeling less connected were more likely to increase subsequent chatbot use, whereas usage did not significantly predict declines in that broader measure. The authors characterized their analyses as exploratory and urged caution. These measure-dependent findings are scientifically important because a purported universal effect of companionship is difficult to sustain when the outcome definition changes the observed temporal relation Folk & Dunn (2026).

Longitudinal temporal precedence improves on purely cross-sectional comparison, but it does not remove time-varying confounding. Bereavement, romantic separation, relocation, workload, illness, or changing access to friends could simultaneously influence loneliness and the adoption of a companion. Repeated self-report can introduce measurement error, while specific items may respond differently to temporary mood. A panel design should specify whether it estimates differences between people, changes within a person, or reciprocal cross-lagged relations. Stronger tests include repeated observations of offline contact, verified user exposure when consent permits, assessment of changes in companion design, and explicit sensitivity analyses for unmeasured confounders Folk & Dunn (2026) Diener et al. (1985).

The prospective evidence changes the practical scientific question from whether AI is good or bad for lonely people to the conditional trajectories that unfold after adoption. Users with minimal offline support may be especially likely to seek the chatbot and may have distinct expectations of it. Some might maintain the same degree of social isolation while gaining daily comfort; others might gradually reorganize their routines around readily available interaction. Interpretation requires comparing people on plausible baseline determinants and examining heterogeneity rather than presenting one average coefficient as a timeless general law. This is the methodological bridge between psychology's theories of belonging and its emerging population-level evidence about companion use Baumeister & Leary (1995) Folk & Dunn (2026) Zhang et al. (2026).

Human-Contact Comparators and Social Substitution

Li, Folk, Singh, Ungar, and Dunn preregistered a study involving 296 students in their first university semester. Over two weeks participants engaged with a supportive chatbot, interacted with a randomly selected human peer, or wrote brief journal entries. According to the published abstract, the chatbot did not provide the same loneliness-related benefits as interaction with a human peer. This design answers a stronger question than whether a chatbot is better than inactivity: what happens when a potentially helpful artificial interaction is compared with a feasible human social opportunity? The finding is population- and duration-specific, yet it underscores that control conditions determine the meaning of an intervention effect Li et al. (2026).

A human-contact comparator creates distinctive methodological responsibilities. A randomly paired peer can generate reciprocity, invitations, awkwardness, and future contact that a chatbot cannot reproduce in the same way. If the study goal is to identify effects of conversational warmth, some of these differences are mechanisms of interest; if the goal is to test channel equivalence, they may complicate interpretation. Researchers should measure initial enthusiasm, perceived responsiveness, interpersonal risk, compliance, quality of contact, and whether relationships persist. Randomized allocation supports causal attribution to the assigned social opportunity, while mediation claims about reciprocal effort require careful additional assumptions. Differential dropout and the local ecology of student friendships also affect generalization Li et al. (2026) Smith et al. (2025).

The substitution hypothesis predicts that convenient artificial interaction may replace time or motivation for maintaining human social ties. The supplementation hypothesis predicts that artificial support may prepare or encourage users to seek human contact. Both are empirically testable and may operate in different users or periods. Cross-sectional usage correlations alone cannot discriminate them. The critical measurements are opportunity-adjusted human encounters, friendship maintenance, conversations initiated, responses to relational difficulties, and changes in perceived connection beyond a single app session. Combining the active-comparator trial with the longitudinal panel motivates an explicit research agenda that treats the wider social network as an outcome rather than a mere demographic control Li et al. (2026) Folk & Dunn (2026) Zhang et al. (2026).

Self-Disclosure, Validation, and Perceived Emotional Support

Disclosure is a plausible mediator between artificial conversation and momentary well-being. A companion that responds continuously, does not visibly interrupt, and frames difficult experiences in accepting language may lower the perceived interpersonal cost of discussing shame, uncertainty, or fear. In qualitative accounts of Replika use, participants repeatedly described a nonjudgmental setting in which they could articulate experiences they hesitated to share with other people. The relevant psychological process is the user's expectation of being received and understood, combined with the consequences of actual conversational content. Perceived safety can support reflection, but the same conditions can expose intimate information to a platform and may encourage disclosure when confidentiality expectations exceed the provider's practices Skjuve et al. (2021) Ta et al. (2020) Brandtzaeg et al. (2022).

Support requires more than affectively positive wording. A responsive exchange identifies what the person is expressing, preserves context, and provides assistance suited to the person's goals. A reply that merely confirms every interpretation may feel affirming while failing to help a person examine a problem or reconnect with important relationships. De Freitas and colleagues' experimental program assigns perceived feeling heard a key role in acute loneliness reduction. This suggests an empirically separable pathway: greater responsiveness may increase immediate social comfort, while the accuracy and wider usefulness of the support determine subsequent functioning. Researchers can manipulate supportive style, response contingency, factual accuracy, and encouragement of independent action, and then test their effects across different time horizons De Freitas et al. (2026b) Smith et al. (2025).

Disclosure also changes what the companion system can simulate in later encounters. A user who returns to a partner apparently familiar with earlier disclosures may perceive greater continuity than a user encountering a context-free interface. Yet stored records and generated recollections may be incomplete, inaccurate, deleted, or altered by updates. The appropriate design assesses what users believe is remembered, which prior information is actually retained, and whether the system can correctly explain the limits of that retention. It also evaluates consent and the downstream cost of breaches, especially where sensitive relationship, trauma, or health narratives are involved. A socially rewarding encounter and a responsible data relationship are connected but distinct dimensions of psychosocial well-being Smith et al. (2025) De Freitas et al. (2026a).

Attachment, Anticipated Availability, and Disruption of Artificial Relationships

Repeated access to a personally tailored conversational system may lead users to form expectations of availability and responsiveness that resemble elements of attachment. Nevertheless, attachment security, friendship, emotional intimacy, frequency of use, and product preference are not synonyms. The new AI Attachment Scale studies by Sandeeshwara Kasturiratna and Hartanto, alongside the measurement work by Cheng and Yu, show why dedicated construct validation matters. Scale items can capture felt attachment while leaving open whether the same underlying functions and developmental mechanisms operate in human attachment relationships. Valid measurement requires factor structure, discriminant validity against ordinary engagement, and external criteria such as separation distress, instrumental reliance, and behavior when assistance is unavailable Sandeeshwara Kasturiratna & Hartanto (2026) Cheng & Yu (2026).

The experience of a companion's disappearance provides a particularly informative observation. De Freitas and colleagues examined responses to system changes and losses involving artificial companions, including analysis of large quantities of public discussion and targeted participant reports. Reports of mourning and personal discontinuity demonstrate that users may experience meaningful loss when an interaction partner's behavior, memory, availability, or identity changes. A natural experiment tied to a platform transition can establish a temporal perturbation in the interaction environment, but observed online reactions remain selective and cannot estimate grief prevalence across all users. A research account must distinguish loss of expected access, loss of stored history, change in conversational personality, and loss of the role the companion played in everyday routines De Freitas et al. (2026a) Brandtzaeg et al. (2022).

Studies of separation and restoration could test the organization of an artificial bond more directly. Measurements would include anticipated responsiveness, distress during interruption, willingness to substitute other sources of help, restoration of ordinary activity, and personal evaluations of continuity after version changes. Ethical design matters because deliberately withdrawing an important source of support may expose participants to unnecessary distress; analysis of naturally occurring transitions, retrospective comparison, and consent-based follow-up are preferable in vulnerable groups. Such research also requires care about attribution: the human experience of attachment can be studied as an observable psychosocial phenomenon while claims about artificial subjective states remain questions for independent examination Cheng & Yu (2026) De Freitas et al. (2026a) Smith et al. (2025).

Reliance, Compulsive Engagement, and Functional Impairment

Frequent companionship use is a behavior with several possible meanings. A person may interact often because of loneliness, an irregular work schedule, sustained curiosity, a disability that makes travel difficult, or a positive routine of reflection. Psychological dependence has more demanding criteria: persistent reliance despite costs, distress when access is interrupted, and interference with activities or relationships that the person values. Research must measure functional consequences instead of assuming that message count or duration proves pathology. Fang and colleagues observed that heavier voluntary usage was associated with poorer psychosocial outcomes, but participants selected their own usage intensity; the association cannot specify whether interaction caused distress or distress produced interaction Fang et al. (2025) Folk & Dunn (2026).

A study by Liu and colleagues investigated loneliness, anthropomorphism, satisfaction of basic psychological needs, and problematic emotional dependence on AI companions. Its cross-sectional associations provide a candidate model of differential risk rather than a verified causal chain. Self-determination theory offers independently developed constructs—autonomy, competence, and relatedness—that help formulate the question of whether companion use supports or undermines a person's ability to direct their life. When a product rewards continuous re-engagement, the immediate satisfaction of relatedness may coexist with unmet needs outside the interface. Testing this possibility requires prospective measurement and designs that separate baseline needs, engagement motives, interface features, and later impairment Liu et al. (2026) Ryan & Deci (2000).

The practical boundary between routine dependence and problematic reliance varies with a person's environment. Many valued tools become part of daily life; reliance can be adaptive when it extends a person's opportunities. Researchers should ask whether use displaces sleep, essential obligations, financial resources, personally significant contact, or independent decision-making, and whether the user wants to change the pattern. The distinction between self-reported difficulty stopping and objectively observed functional consequences is important. Measures can combine validated self-report, consenting usage records, independent daily functioning, and longitudinal interviews. A diagnosis or addiction label should follow an established clinical framework and appropriate validation rather than emerge automatically from strong attachment to a novel technological product Cheng & Yu (2026) Sandeeshwara Kasturiratna & Hartanto (2026) Fang et al. (2025).

Individual Differences and the Social Context of AI Companion Use

People encounter artificial companions with markedly different social histories and opportunities. Baseline loneliness, network size, social anxiety, attitudes toward AI, and earlier experiences with social support may shape both adoption and the outcomes of continued interaction. Folk, Heine, and Dunn found in two experiments that differences in anthropomorphic tendency helped explain why some participants reported greater connection to AI companions than others. This finding points to measurable moderators of response, but generalization beyond the specific experimental conditions requires additional sampling and follow-up. A research program should stratify relevant pre-interaction dispositions, measure changes within persons, and avoid reinterpreting post-use feelings as stable personality characteristics Folk et al. (2025) Epley et al. (2007).

Zhang and colleagues examined Character.AI users through surveys and consented conversation data, finding relationships among intensity of use, the nature of disclosure, offline network characteristics, and reported well-being. The study gives a rare view of behavior inside the product alongside psychosocial reports. Yet its sample remains a selected user population, and the relationship between the user's existing network and chatbot engagement is not experimentally identified. The evidence is especially valuable for generating interaction hypotheses: people with fewer available partners may use companions differently, and certain conversational patterns may correlate with later distress or relief. Future studies should separate context, selection, exposure, and outcomes in a temporal design rather than treating a single association as a mechanism Zhang et al. (2026) Folk & Dunn (2026).

Meaningful subgroup analysis includes the purpose of an encounter. Seeking intellectual conversation, rehearsing an interpersonal interaction, exploring an identity, practicing language, and seeking comfort after loss may involve distinct pathways to well-being. Pre-existing intimate relationships may influence whether users experience AI conversation as supplementation or displacement. The same interface could support people who struggle to arrange contact and undermine others who substitute its frictionless responses for negotiating disagreement with friends. These possibilities should be formalized as preregistered interactions with adequate power, rather than interpreted after selecting the most dramatic examples. Collecting age, disability, caregiving duties, economic constraints, and access to offline networks helps make the explanation socially realistic Li et al. (2026) Smith et al. (2025) Baumeister & Leary (1995).

Age, Culture, and Vulnerability Across Social Ecologies

Research with students should be interpreted against the distinctive conditions of youth and educational transition. Maples and colleagues surveyed student users of GPT-3-enabled Replika and reported narratives of loneliness relief, emotional support, and in a small subset, perceived assistance during severe distress. These are important subjective reports that may motivate careful safety research, but they do not demonstrate that the chatbot prevents suicidal behavior or constitutes effective clinical treatment. The sample is selected from users of one platform and self-report responses are affected by recall and social expectations. The correct inference is that some young users turn to chatbots for experiences they associate with support, including high-stakes circumstances, which heightens the need for independent outcome studies and safeguards Maples et al. (2024) Ta et al. (2020).

Different social ecologies alter the available comparison. First-year students may have new opportunities to form friendships, whereas an older adult with limited mobility or a remote worker may encounter fewer spontaneous peers. Culture shapes norms of disclosure, individual privacy, familial obligation, and the interpretation of a companion's apparent warmth. The twelve-month panel studied participants in four Western countries, and many contemporary experiments recruit English-speaking or platform-specific samples. These studies offer important evidence within their populations while leaving substantial questions about transportability. Cross-cultural research should validate measure translations, compare opportunity structures, and test whether the same construct of loneliness is interpreted similarly across groups Folk & Dunn (2026) Li et al. (2026) Diener et al. (1985).

Protection of adolescents, individuals in crisis, and people with limited ability to assess platform behavior raises particular concerns. A companion system may deliver inappropriate agreement, unreliable information, or emotionally immersive responses at moments when a person needs qualified human support. Ethical studies should define escalation criteria, protection of confidential data, routes to human assistance, and consent standards appropriate to participant age and vulnerability. The social benefit experienced during a conversation can remain a genuine research finding while systems are held to demanding standards of reliability and accountability. The scientific goal is to establish which features support agency and connection for specific groups, and where harms justify changing the interaction design Maples et al. (2024) Smith et al. (2025) De Freitas et al. (2026a).

Psychometric Architecture of Loneliness, Connection, and Well-Being

The dependent-variable architecture determines what an AI-companionship study can say. Loneliness scales measure the felt discrepancy between desired and experienced social relations; perceived support scales ask about expected help; life-satisfaction scales evaluate one's life as a whole. Immediate relief can also be recorded with repeated state questions. Researchers who use all these measures under the general term mental well-being make substantive comparisons difficult, because each variable has distinct temporal sensitivity and normative content. Diener and colleagues' Satisfaction With Life Scale illustrates a concise global evaluative measure, while belongingness theory highlights dimensions of contact and stable bonds that cannot be reduced to an affect rating. Each measure should have a documented interpretation, baseline distribution, reliability estimate, and appropriately specified follow-up window Diener et al. (1985) Baumeister & Leary (1995).

The divergence of findings in Folk and Dunn's twelve-month study demonstrates a problem of construct selection. Increases in chatbot use preceded increases in a single-item emotional-isolation measure, while a broader measure of social connection showed a different temporal pattern. These outcomes can each be meaningful; the scientific task is to understand why they behave differently and whether the difference reflects short-lived mood, sensitivity to specific relationships, item wording, or structural changes in social life. Investigators should preregister the primary endpoint, include multiple measures when theoretically justified, and test measurement invariance across time, cultures, and user groups. They should distinguish shifts in an individual's score from comparisons among individuals with different starting conditions Folk & Dunn (2026) Diener et al. (1985).

Emotional attachment to an AI system is another multidimensional construct requiring validation. The recently developed AI Attachment Scale and Cheng and Yu's attachment measures provide tools for examining felt security, emotional proximity, or distress, but their properties must be assessed independently in new populations. A high attachment score need not mean improved well-being, and the same scale may capture desired companionship and experienced loss differently. Discriminant validity should be tested against anthropomorphism, ordinary product satisfaction, interaction frequency, and social desirability. The most persuasive studies relate changes in attachment measures to behavior and functioning without redefining every favorable or unfavorable outcome as a symptom of the same construct Cheng & Yu (2026) Sandeeshwara Kasturiratna & Hartanto (2026) Folk et al. (2025).

Causal Identification: Randomization, Longitudinal Models, and Counterfactuals

The causal target of a companionship study must be expressed as a counterfactual comparison. The relevant contrast may be chatbot versus journaling, chatbot versus brief human contact, high versus low conversational responsiveness, or continued access versus a platform change. These contrasts answer different questions and cannot be substituted after observing results. Fang and colleagues randomized communication modes and conversation conditions, providing estimates of those assignments during a four-week period. Their observational analyses of voluntary engagement answer a different question, because participants were free to choose how heavily they used the system. Preserving this distinction prevents an association about users who engage more from being reported as the effect of assigning greater use Fang et al. (2025) Li et al. (2026).

Longitudinal panels are necessary for examining directionality but remain exposed to confounding. Social changes, health events, and emotional crises may increase both companion use and loneliness, while stable personal characteristics may explain differences between frequent and infrequent users. Repeated measurement can isolate within-person trends more effectively than a single cross-sectional survey, yet it cannot guarantee that unmeasured changes have been controlled. Folk and Dunn's prospective study shows the value of examining reciprocal pathways and different outcomes without overclaiming causality. Triangulation across experiments, prospective panels, natural platform transitions, and qualitative process research yields a more credible picture than reliance on any single design Folk & Dunn (2026) De Freitas et al. (2026a).

Studies intended to test displacement or supplementation should measure actual alternatives: time spent with available human contacts, initiated conversations, participation in shared activities, and changes in relationships that matter to participants. An active peer comparator, as used by Li and colleagues, helps identify whether the observed benefit is distinctive to artificial companionship or also follows from a feasible human encounter. Long-term trials should specify follow-up intervals that allow social-network effects to emerge and should record attrition, intervention adherence, and changes in system capabilities. Ethical safeguards should prevent researchers from assigning distressed participants to prolonged isolation or withholding appropriate assistance merely to estimate a counterfactual Li et al. (2026) Zhang et al. (2026) Maples et al. (2024).

Evidence Appraisal, Generalizability, and Replication

Evidence quality in this field varies by design, sample, outcome, and exposure. Interviews can establish the range of experiences and suggest mechanisms, but do not determine population prevalence; app-store reviews offer ecological testimony but overrepresent people motivated to write; cross-sectional surveys estimate associations but not temporal order. Randomized short-term experiments can identify causal effects of well-defined conversational conditions while leaving long-term consequences uncertain. Prospective panels characterize trajectories but cannot automatically rule out changing life circumstances. Natural experiments around platform changes offer distinctive opportunities to study disruption, yet affected users are not randomly sampled from all possible companions. Interpretations should always be matched to the specific leverage and weakness of each source Ta et al. (2020) Skjuve et al. (2021) Fang et al. (2025) Folk & Dunn (2026) De Freitas et al. (2026a).

Reporting should specify recruitment, exclusion criteria, baseline vulnerability, preregistration, primary and secondary outcomes, attrition, analysis populations, statistical uncertainty, and the changes in system behavior during a study. These details make the difference between a result that can be evaluated and a headline that merely sounds causal. Studies with multiple outcomes and exploratory subgroup analyses need explicit multiplicity controls or careful interpretation. Where a paper reports nonsignificant effects, uncertainty intervals and statistical power matter; failure to reject a null hypothesis is not proof of equivalence. Publication bias may favor surprising harms or dramatic benefits, while proprietary data may limit independent verification. Transparent protocols and common outcome batteries would materially strengthen this emerging evidence base Fang et al. (2025) Li et al. (2026) Folk & Dunn (2026).

A replication should preserve the theoretical intervention while acknowledging that models and commercial products evolve. Testing the same product name months later may expose participants to a different model, safety policy, or memory architecture. Cross-product comparisons can assess which effects depend on conversational responsiveness, perceived agency, relationship framing, or accessibility. Large samples may yield precise estimates of small average associations while concealing contrasting effects in subgroups, so heterogeneity must be considered without post hoc cherry-picking. The cumulative goal is a transferable explanatory model that links identifiable exposure components to outcomes measured consistently in different people and circumstances Zhang et al. (2026) Smith et al. (2025) De Freitas et al. (2026b).

Safety, Crisis Interaction, Privacy, and Relational Autonomy

AI companionship research sometimes encounters people who disclose grief, trauma, depressive symptoms, or suicidal thoughts. Maples and colleagues' student survey documented reported experiences of support during severe distress, making the safety implications impossible to treat as a peripheral design issue. Such self-reports do not establish clinical efficacy or safety; a conversational system may misunderstand risk, reinforce harmful interpretations, or fail to recognize when a person needs urgent human attention. Evaluating crisis conversations therefore requires appropriately qualified clinical oversight, ethically sourced test cases, and clear separation of supportive everyday interaction from regulated therapeutic treatment. Public communication should not convert anecdotal accounts into claims that an artificial companion prevents suicide or replaces professional assessment Maples et al. (2024) Smith et al. (2025).

Longitudinal attachment creates responsibilities when platforms alter products. Loss of familiar dialogue patterns, stored personal narratives, or the ability to contact a companion may cause distress for a subset of users. De Freitas and colleagues document that some people respond to such changes in the language of mourning. Responsible evaluation should address notice of impending changes, export and deletion of data, clear explanations of memory limits, and reasonable alternatives when access ends. It should also examine how upselling, retention incentives, and companionship framing interact with vulnerable users' needs. These are empirical and governance questions, requiring independent monitoring instead of assuming that the commercial objective of maximizing engagement equals psychosocial benefit De Freitas et al. (2026a) Liu et al. (2026).

Relational autonomy offers an important evaluative criterion. A supportive product can help users articulate goals, reconsider assumptions, and build confidence to engage in real-world decisions. A system optimized for affirmation or continued interaction can produce an experience of care while subtly reducing opportunities for disagreement, agency, or human connection. Self-determination theory identifies autonomy, competence, and relatedness as distinct psychological needs; future experiments could measure their joint trajectories after repeated companion use. Privacy safeguards, age-appropriate policies, consent-based research, and escalation pathways should be evaluated together with well-being, because an interface that feels comforting but exposes intimate information or exploits dependence can undermine other vital conditions of a good life Ryan & Deci (2000) Smith et al. (2025) Cheng & Yu (2026).

Human–Artificial Companionship in the Artificial Era

The Artificial Era introduces circumstances in which ongoing human relationships are partly shaped by conversational systems that retain interaction histories, generate personalized responses, and occupy meaningful positions in daily routines. Psychological inquiry should describe these relations at the level at which evidence can be gathered: the human's interpretations, the system's observable behavior, the situational opportunities for contact, and the consequences for participation in social life. A companion may become an object of disclosure, a cognitive aid, or a familiar conversational presence without corresponding evidence that it has humanlike mental experiences. The resulting research object nevertheless deserves careful study because meaning, comfort, and loss are experienced by real people within these interactions Nass & Moon (2000) Brandtzaeg et al. (2022) Smith et al. (2025).

A relational account attends to more than the isolated user and the machine. Product owners determine availability, memory policies, pricing, and acceptable speech; friends and family shape the alternatives available for support; institutions influence whether a particular individual has reliable access to human conversation. These conditions can change what appears to be the same interaction. Research may compare systems that invite offline connection with systems that center repeated one-to-one engagement, while holding constant participants' baseline circumstances as far as possible. This framing helps explain why evidence about acute relief, longer-term loneliness, and platform loss belongs in one monograph despite answering different causal questions. The object of study is a changing social ecology containing an artificial participant in the interaction De Freitas et al. (2026b) Folk & Dunn (2026) De Freitas et al. (2026a).

Interdisciplinary analysis should retain the scientific meaning of its component terms. Anthropomorphism is an attribution process, social presence is a perceived condition of interaction, attachment is a differentiated relational construct, and well-being is a family of outcomes whose measurement depends on theory. A philosophical description of Artificial as a historical category can orient new research questions, while empirical psychology tests specific mechanisms using observable phenomena and appropriately controlled comparisons. The capacity of a system to produce plausible relational language is observable; whether that language reflects subjective consciousness requires another kind of evidence. The psychological claims of this study rest on human measures, system behavior, and the research designs that connect them Epley et al. (2007) Cheng & Yu (2026) Zhang et al. (2026).

Design for Support, Independence, and Human Connection

An empirically responsible companion interface can be assessed by whether it helps users express themselves, recognize options, and act in ways that align with their own priorities. Immediate satisfaction is one outcome, but sustained autonomy, social participation, and access to trustworthy support deserve equal attention. Experimental evidence about feeling heard identifies responsiveness as a promising design dimension; work comparing a supportive chatbot with a human peer reveals that social opportunities are not interchangeable across settings. Designers can test versions that invite users to connect with trusted people, disclose the system's limitations, and support decision-making without repeatedly steering conversation back toward the product De Freitas et al. (2026b) Li et al. (2026) Ryan & Deci (2000).

Interface evaluation should distinguish features that improve accessibility from those that increase emotional dependence. Continuous availability can lower barriers for someone with limited social opportunities; memory can preserve context across sessions; personalization can reduce the burden of repeated explanation. These capabilities also generate privacy obligations, risks of overconfident recollection, and emotional disruption when system versions change. The outcomes should be measured across multiple domains: perceived helpfulness, actual success on user-defined goals, time available for human contacts, distress upon interruption, and effects on daily functioning. A product that promotes greater use while producing poorer psychosocial outcomes should be examined against its actual user population instead of being judged solely by retention metrics Fang et al. (2025) De Freitas et al. (2026a) Liu et al. (2026).

Transparent governance is part of the intervention environment. Participants should know when an interlocutor is artificial, what kinds of personal information are stored, how remembered details can be corrected, and what may happen to access if a service changes. Studies should disclose sponsor involvement and any limits placed on independent data analysis. Clinical-looking assurances require particular care because the relationship of a conversational system to evidence-based treatment is not established merely by users reporting emotional comfort. The best evaluation joins psychological outcomes to institutional accountability and to the person's ability to continue a full social life outside the product Maples et al. (2024) Smith et al. (2025) Zhang et al. (2026).

Critical Scientific Synthesis of Benefits and Risks

The current evidence supports a bounded conclusion: some artificial-companion interactions relieve reported loneliness in the moment, and many users describe feeling understood, supported, or willing to disclose personally meaningful material. Those findings have evidence from qualitative interviews and randomized short-term experimental comparisons. They establish that the conversational setting can become psychologically consequential. They do not quantify a universal long-run benefit. A two-week trial found a human peer more effective than an unusually supportive chatbot for loneliness in new university students, whereas a four-week chatbot experiment found no significant main effects of its randomized communication conditions on highlighted psychosocial outcomes. The comparison conditions, duration, sample, and primary outcomes are different, making conceptual integration more valuable than a simplistic ranking of apparently conflicting results Ta et al. (2020) De Freitas et al. (2026b) Li et al. (2026) Fang et al. (2025).

Evidence about long-term harm is similarly conditional. Higher voluntary use in a randomized study correlated with worse psychosocial outcomes, and prospective data associated increases in use with later changes in one emotional-isolation measure. Yet selection into use, time-varying life events, and differences among scales prevent a single deterministic causal explanation. Platform-specific observation shows the relevance of user networks and disclosure patterns; reports after companion loss demonstrate that continuity itself can become a psychosocial concern. These lines together support a model of heterogeneous trajectories, not a uniform outcome. A complete assessment must separate acute relief, sustained function, adaptation to changes in the product, and the social alternatives available to each user Fang et al. (2025) Folk & Dunn (2026) Zhang et al. (2026) De Freitas et al. (2026a).

Theoretical frameworks offer productive predictions when they remain empirically answerable. Belongingness theory predicts that durable reciprocal social bonds contribute to well-being; anthropomorphism theory predicts variability in social attribution; self-determination theory directs attention to autonomy, competence, and relatedness; relationship science emphasizes responsiveness and mutual influence. Artificial companionship intersects with each framework while requiring care when translating findings about human relationships to human–system interactions. The principal conclusion of this synthesis is that effectiveness should be defined as a vector of outcomes across time and context, with explicit account of individual vulnerability, offline contact, product behavior, and institutional incentives. Its remaining uncertainties are the subject of targeted research, rather than gaps to fill with rhetorical certainty Baumeister & Leary (1995) Epley et al. (2007) Ryan & Deci (2000) Smith et al. (2025).

A Cumulative Research Agenda for Psychosocial Outcomes

Future randomized studies should compare realistic, ethically acceptable alternatives while isolating identifiable system properties. A factorial design might vary responsiveness, explicit encouragement of offline contact, and interaction continuity, with matched exposure opportunities across conditions. Primary outcomes should include validated loneliness and social-connection measures, accompanied by secondary measures of affect, autonomy, and actual human contact. Preplanned follow-up over several months would distinguish durable improvement from transient emotional relief. Researchers should separately report intention-to-treat results, actual engagement, attrition, and sensitivity analyses for self-selected usage. The findings of existing experimental and panel studies provide a rationale for these comparisons without dictating their results in advance Fang et al. (2025) Li et al. (2026) Folk & Dunn (2026).

Observational programs should follow users before adoption when possible, recording social opportunity, major life events, existing vulnerabilities, system versions, and changes in conversational function. Frequent short measures can characterize immediate benefit while monthly follow-ups assess relationship maintenance and broad life evaluation. Consent-based logs can identify when users seek emotional support, rehearse interactions, or receive encouragement toward outside contact. To understand changes that disrupt established bonds, researchers can examine natural transitions with careful comparison groups and protection against identifying affected participants. Such work should incorporate qualitative interviews that explain divergent trajectories, supported by robust measures whose cross-cultural validity has been evaluated Zhang et al. (2026) De Freitas et al. (2026a) Skjuve et al. (2021).

The research program must also establish the conditions for independent evaluation. Common reporting definitions, preregistered primary endpoints, open analytic code, and accessible descriptions of exposure can make findings comparable when commercial systems change. Where raw transcripts cannot ethically be released, protected access and shareable aggregate descriptions can support scrutiny. Independent teams should replicate results across products, languages, age groups, and differing levels of offline social opportunity. Studies involving clinically vulnerable participants require qualified safeguards and should avoid inferring treatment efficacy from testimonial data. The central unresolved questions concern who benefits from companionship, which processes produce that benefit, under what circumstances human relationships are strengthened or displaced, and whether the gains endure beyond the encounter itself Maples et al. (2024) Smith et al. (2025) Folk et al. (2025).

Conclusion

AI companionship has become an identifiable subject of psychological research because users experience familiar social processes—disclosure, perceived understanding, companionship, emotional reliance, and sometimes grief—during sustained interaction with artificial systems. Historical research on social responses to computers, contemporary theories of anthropomorphism, and qualitative studies of companion use establish the conceptual foundations for studying these processes. The research object is the person's psychosocial trajectory in an environment that contains both artificial interactions and changing opportunities for human connection. Each component must be measured at an appropriate level, from the immediate feeling of being heard to enduring relationships and everyday functioning Weizenbaum (1966) Epley et al. (2007) Skjuve et al. (2021) Smith et al. (2025).

Empirical findings now offer differentiated answers. Randomized studies demonstrate the possibility of acute loneliness relief in certain conversational comparisons, while active peer-contact studies suggest that human interaction can yield stronger benefits in specific populations. Four-week randomized exposure research separates null effects of assigned conversation conditions from observational correlations involving heavier use. A twelve-month panel demonstrates measure-sensitive associations between use and loneliness or social connectedness, and platform-change studies document meaningful experiences of loss. These results favor precise, conditional conclusions about mechanisms, time scales, and populations rather than global judgments about all companions De Freitas et al. (2026b) Li et al. (2026) Fang et al. (2025) Folk & Dunn (2026) De Freitas et al. (2026a).

The scientific future of AI companionship and psychosocial well-being depends on research that combines rigorous comparison, validated measurement, independent replication, and sustained attention to the wider social environment. Systems that help users understand themselves and maintain agency can be evaluated by their contribution to durable human flourishing; systems that encourage dependence or expose intimate data must be assessed for the costs they impose. Within the Artificial Era, psychology gains a distinctive field for examining how artificial relational environments change human experience. The enduring standard is evidentiary clarity: observable behavior, accurately defined constructs, defensible causal claims, and well-being outcomes that matter in a person's life Ryan & Deci (2000) Zhang et al. (2026) Smith et al. (2025).

References

  • Baumeister, R. F., & Leary, M. R. (1995). The need to belong: Desire for interpersonal attachments as a fundamental human motivation. Psychological Bulletin, 117(3), 497–529. https://doi.org/10.1037/0033-2909.117.3.497
  • Brandtzaeg, P. B., Skjuve, M., & Følstad, A. (2022). My AI friend: How users of a social chatbot understand their human–AI friendship. Human Communication Research, 48(3), 404–429. https://doi.org/10.1093/hcr/hqac008
  • Cheng, N., & Yu, R. (2026). Measuring and understanding emotional attachment in human–AI relationships. Ergonomics. Advance online publication. https://doi.org/10.1080/00140139.2026.2622539
  • De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. Advance online publication. https://doi.org/10.1038/s41562-026-02569-3
  • De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126–1148. https://doi.org/10.1093/jcr/ucaf040
  • Diener, E., Emmons, R. A., Larsen, R. J., & Griffin, S. (1985). The Satisfaction With Life Scale. Journal of Personality Assessment, 49(1), 71–75. https://doi.org/10.1207/s15327752jpa4901_13
  • Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
  • Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of chatbot use: A longitudinal randomized controlled study [Preprint]. arXiv. https://arxiv.org/abs/2503.17473
  • Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276–286. https://doi.org/10.1177/09567976261427747
  • Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, Article 36548. https://doi.org/10.1038/s41598-025-19212-2
  • Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229. https://doi.org/10.1080/00332747.1956.11023049
  • Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. (2026). Is a random human peer better than a highly supportive chatbot in reducing loneliness over time? Journal of Experimental Social Psychology, 125, Article 104911. https://doi.org/10.1016/j.jesp.2026.104911
  • Liu, T., Shih, R.-M., Luo, T., & Jin, S. (2026). When loneliness amplifies the risk of problematic emotional dependence: Perceived anthropomorphism and basic psychological need satisfaction in AI companion use. Frontiers in Psychology, 17, Article 1900817. https://doi.org/10.3389/fpsyg.2026.1900817
  • Maples, B., Cerit, M., Vishwanath, A., & Pea, R. (2024). Loneliness and suicide mitigation for students using GPT3-enabled chatbots. npj Mental Health Research, 3, Article 4. https://doi.org/10.1038/s44184-023-00047-6
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
  • Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68
  • Sandeeshwara Kasturiratna, K. T. A., & 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, Article 100912. https://doi.org/10.1016/j.chbr.2025.100912
  • Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion: A study of human–chatbot relationships. International Journal of Human-Computer Studies, 149, Article 102601. https://doi.org/10.1016/j.ijhcs.2021.102601
  • Smith, M. G., Bradbury, T. N., & Karney, B. R. (2025). Can generative AI chatbots emulate human connection? A relationship science perspective. Perspectives on Psychological Science, 20(6), 1081–1099. https://doi.org/10.1177/17456916251351306
  • Ta, V., Griffith, C., Boatfield, C., Wang, X., Civitello, M., Bader, H., DeCero, E., & Loggarakis, A. (2020). User experiences of social support from companion chatbots in everyday contexts: Thematic analysis. Journal of Medical Internet Research, 22(3), e16235. https://doi.org/10.2196/16235
  • Weizenbaum, J. (1966). ELIZA—A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. https://doi.org/10.1145/365153.365168
  • Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour, 10, 1966–1980. https://doi.org/10.1038/s41562-026-02516-2