RESEARCH DOSSIER
Social Substitution and Complementarity in AI Relationships
Abstract. Social substitution and complementarity in AI relationships concern changes in human social contact, support and network participation associated with engaging conversational artificial systems. This fundamental research monograph reconstructs the history of social displacement and stimulation hypotheses, interpersonal belonging and support theories, parasociality and anthropomorphism; distinguishes short-lived emotional relief from sustained changes in offline relationships; and evaluates original experimental, observational, interview and longitudinal studies through October 2026. The research assesses evidence from AI-companion users, preregistered human-peer comparisons, student cohorts, multinational panels and platform disruption studies. It explains selection effects, bidirectional pathways, time-use opportunity costs, emotional disclosure and social bridging, with attention to effect heterogeneity, privacy, developmental context and competing causal interpretations. The synthesis identifies conditions under which an AI relationship can be an additional support resource, those associated with reduced human interaction, and research designs needed to establish their mechanisms. The method is an evidence-based critical narrative synthesis rather than a systematic review or meta-analysis.
Keywords: social substitution; social complementarity; AI relationships; AI companionship; social displacement; social stimulation; social bridging; loneliness; human social connection; perceived social support; anthropomorphism; social contact diaries; human–AI interaction; Artificial Era
The Research Object: Changes in Human Social Life Associated with AI Bonds
Social substitution and complementarity in AI relationships name a family of changes in the social opportunities, practices, and relationships of a human participant after or alongside engagement with an artificial conversational partner. Substitution occurs when some human-directed contact, effort, opportunity, or reliance is displaced by interaction with a system relative to a defensible counterfactual. Complementarity occurs when that engagement adds resources to a person's social life, strengthens existing relationships, or makes subsequent human interaction more feasible. The empirical unit is a trajectory of social behavior and support, not simply a user's affectionate description of a chatbot. A participant may feel accompanied and nevertheless have fewer conversations with friends; another may feel accompanied and gain confidence to contact them. These outcomes must be measured independently. The distinction draws on the history of displacement and stimulation research in digitally mediated interaction Valkenburg & Peter (2007).
The primary question is comparative: what changes in human contact or connectedness would follow if access to a conversational AI companion were changed while relevant conditions were held sufficiently comparable? Potential outcomes include minutes of interpersonal conversation, number and diversity of contacts, intimacy and disclosure to human partners, perceived availability of support, enacted requests for help, and maintenance or formation of social ties. An immediate reduction in subjective loneliness is an important outcome but establishes neither substitution nor complementarity without evidence about human relationships. Likewise, a large volume of AI conversation may describe companionship intensity without showing which human interactions did or did not occur. The empirical literature now spans qualitative accounts, surveys, randomized interventions, repeated-measures panels, and natural experiments, each identifying different portions of this question Zhang et al. (2026) Folk & Dunn (2026).
The scope extends across AI companions marketed for relationships and general-purpose conversational systems used socially. Intent and actual use matter more than a product label: a participant may use one model for emotional reflection, rehearsal for a difficult human conversation, collaborative writing, and companionship within the same week. Each function has different potential consequences for human ties. The scientific study should distinguish exposure to a system, perceived humanlikeness, relationship commitment, and downstream social action. The monograph treats philosophical descriptions of an Artificial Era as intellectual context and tests claims about social displacement through observable human outcomes. Claims concerning consciousness or inner experience in artificial systems require their own evidence and are not used to infer whether a person's human network has changed.
Conceptual Boundaries: Substitution, Compensation, Augmentation, and Bridging
Substitution refers to a reduction in the quantity or quality of human relational participation attributable to engagement with AI. The displaced resource might be time, a particular social function, a decision to disclose, or willingness to pursue a new connection. Compensation denotes obtaining through AI an experience or resource that a person perceives as unavailable from human partners. Compensation can coexist with substitution, complementarity, or neither: a conversation may soothe an isolated user without changing a single opportunity for human contact. Augmentation describes additional support that improves functioning while preserving existing human ties. Bridging is the stronger case in which interaction with AI increases the probability, quality, or accessibility of subsequent human interaction. These analytic distinctions must be operationalized within explicit time windows and reference populations Cohen & Wills (1985) Ta et al. (2020).
Several observations have different meanings despite superficial similarity. A person using a chatbot late at night when other people are unavailable may gain emotional relief at minimal social opportunity cost. A person repeatedly canceling conversations with friends to continue chatbot sessions exhibits a possible time-substitution pathway, although the underlying cause still needs to be tested. A shy student who practices opening lines with a conversational system and then attends a peer activity may display a bridging sequence if the activity is causally attributable to that practice. A person transferring confidential disclosure from a trusted human friend to a system may show function-specific substitution even if total human contact is unchanged. The distinction between structural integration and the availability of particular support functions has a long research history Cohen & Wills (1985).
The definitions also specify their limits. More messages to people do not necessarily imply higher-quality relationships, and fewer messages need not indicate harm if conflict or coercion has declined. Social networks contain supportive, ambivalent, and dangerous ties. Research therefore requires measures of contact quality, satisfaction, reciprocity, autonomy, and the participant's own goals, rather than an undifferentiated requirement to maximize human interaction. Complementarity is most convincingly established when outcomes improve on independently meaningful dimensions and when gains are not produced by shifting costs to unmeasured partners. Because studies often recruit self-selected AI enthusiasts, a measured association between companion use and relational satisfaction is a description of that sample until alternative selection mechanisms have been addressed Brandtzaeg et al. (2022) Zhang et al. (2026).
Historical Lineages: Belonging, Media, and the Displacement Question
The need for interpersonal attachment provides a foundational psychological reason to investigate new sources of companionship. Baumeister and Leary's synthesis treated belonging as a pervasive motivation involving frequent positive interactions within enduring relational bonds, with consequences for cognition, emotion, and adjustment Baumeister & Leary (1995). This theory motivates questions about unmet affiliation needs and compensatory behavior. It does not settle whether artificial partners satisfy the same range of needs as human relationships or whether short-term feelings of acceptance transfer into enduring interpersonal integration. The distinction between momentary belonging and durable mutual involvement becomes central as chatbots offer apparently personalized conversation without the same forms of embodied cooperation, social accountability, and shared material obligations that frequently characterize human relationships.
The study of apparently intimate relations with mediated figures predates conversational AI. Horton and Wohl analyzed parasocial interaction with media performers in 1956, emphasizing the structure of experienced closeness at a distance Horton & Wohl (1956). Their historical construct is relevant because people may experience familiarity, apparent responsiveness, and loyalty toward a technologically mediated figure. Yet a generative system can respond contingently to a participant's disclosures and maintain varying degrees of conversational history. Those features require supplementary empirical constructs: sequence-level responsiveness, adaptation, human interpretation of the partner, and behavior outside the interaction. A historical parallel can illuminate the formation of attachment-like experience while remaining insufficient for classifying AI companionship as parasocial in every technical or psychological sense.
Early internet research also framed the displacement problem. Kraut and colleagues observed associations between internet use and reduced social involvement in a pioneering longitudinal household sample Kraut et al. (1998). Subsequent work shifted attention toward the type, social direction, and context of online activity. Valkenburg and Peter explicitly contrasted stimulation and displacement hypotheses in adolescent online communication, demonstrating the importance of studying whether technology connects people to existing friends or diverts engagement from them Valkenburg & Peter (2007). The AI case inherits this methodological demand but alters the counterpart: the social exchange may occur with a system rather than another human. Investigators must now distinguish mediated human-human communication from human-machine social interaction before applying findings across media generations Hall (2018).
Anthropomorphism, Social Attribution, and Perceived Reciprocity
A humanlike voice, name, avatar, memory cue, or conversational response can change how a user interprets an artificial exchange. Epley, Waytz, and Cacioppo's three-factor account of anthropomorphism links humanlike attribution to accessible knowledge of human agents, explanatory motivation, and affiliation motivation Epley et al. (2007). This framework predicts heterogeneity: the same technology may seem socially relevant to one participant and purely instrumental to another. The social motivation associated with loneliness may influence both companion-seeking and perceptions of the system. As a result, an association between perceived social presence and reduced distress may arise partly from who approaches the technology in the first place. Experimental manipulations of identity labels and response patterns offer stronger evidence about specific attribution mechanisms.
The distinction between an AI message and beliefs about its source has gained empirical support. Kleinert and colleagues reported two preregistered double-blind experiments using emotionally engaging text exchanges and a modified Fast Friends procedure; closeness judgments depended on conditions of source attribution, including whether AI content was labeled as human Kleinert et al. (2026). Their results concern experimentally elicited interpersonal closeness, not evidence of enduring replacement of offline relationships. Nevertheless, the findings demonstrate how participants' expectations about their conversational partner can shape subjective relational outcomes. A strong theory must therefore measure perceived humanness, disclosure, responsiveness, and partner identification separately and avoid interpreting an increase in closeness ratings as proof of a change in a person's human network.
Perceived reciprocity deserves particular scrutiny. Artificial systems can generate turn-taking, affirmations, follow-up questions, and statements that appear attentive to a user's history. Humans may reciprocate by increasing disclosure, time investment, or emotional commitment. Reciprocity in a conversational task is observable as contingent exchange; its social and moral meanings are shaped by users' interpretations and by the system's capacity to remember, act reliably, or remain available. Qualitative studies of Replika relationships found that personalized, accepting interactions were central to participants' understandings of friendship with AI Skjuve et al. (2021) Brandtzaeg et al. (2022). Such accounts are critical evidence of lived meaning, while reciprocal obligations and stable mutual support in ordinary human relationships require additional measurement rather than direct equivalence by analogy.
Opportunity Costs and the Time-Use Mechanism
The simplest displacement mechanism involves a finite opportunity budget. Time allocated to prolonged conversational sessions cannot simultaneously be used for other activities. Yet displacement is a causal claim about the activity that would have occupied the same time, and its estimation is rarely straightforward. Chatbot engagement may replace passive browsing, unstructured waiting, solitary rumination, sleep, schoolwork, or conversations with family members. These substitutions have different consequences for human connection. Researchers should observe the immediate preceding activity, the alternatives available at the moment, and the postponed activities during subsequent hours. General screen-time totals obscure the pathway because a participant can increase AI use while also increasing social contact. Historically, digital-media displacement theories have had to confront this problem of what is actually displaced Kraut et al. (1998) Valkenburg & Peter (2007).
Time-use claims become more precise with within-person records. Ecological momentary assessment can prompt participants after a conversation to report the motivation for opening the app, whether any human contact was realistically possible, the quality of the exchange, and intended next actions. A complete diary also tracks face-to-face encounters, messaging with human contacts, sleep, and other activities that compete for attention. An evening spent practicing for a social interaction is different from an evening spent avoiding an already arranged meeting. Reverse causality must be addressed: people with fewer social opportunities may have more available time for AI interaction, making high use a consequence of isolation rather than its cause. The longitudinal association reported by Folk and Dunn is therefore relevant without resolving every opportunity-cost mechanism Folk & Dunn (2026).
Opportunity cost also has a social-calendar dimension. Human relationships often depend on recurring commitments, responsiveness, and opportunities whose value is not proportional to minutes spent. Missing a brief but important message may weaken a tie more than reducing casual conversation time. Conversely, a short AI exchange that helps a participant organize a conflict-resolution discussion may have substantial value despite adding screen time. The appropriate dependent variable is therefore selected from the substantive hypothesis: missed invitations, contact initiation, timely responses, sustained friendships, or emotional connection. Randomized access or restriction can identify changes under controlled conditions, but ethically acceptable studies should avoid engineered isolation and account for compensatory behavior after the intervention. The outcome must represent the relationship process rather than a merely convenient usage metric.
Emotional Regulation, Disclosure, and Social Rehearsal
Artificial conversational systems offer text-based environments in which individuals can articulate distress, seek clarification, or rehearse an emotionally difficult conversation. Users of companion applications have described an experience of low evaluation pressure and perceived acceptance Ta et al. (2020) Skjuve et al. (2021). These affordances can influence whether AI functions as immediate self-regulation, a pathway into human support, or a destination that gradually receives disclosures formerly shared with people. The functional trajectory must be measured. A reduction in distress immediately after writing to a chatbot may simply indicate effective emotional expression; it does not specify whether the individual subsequently seeks social contact. Conversely, disclosure to AI that precedes an honest conversation with a friend may support bridging, provided the sequence is measured and alternative explanations are addressed.
Social rehearsal can be studied experimentally with narrowly defined tasks. Participants might practice introducing themselves, asking for help, or communicating a disagreement with either an AI system, a static guide, or a human peer. Investigators can subsequently measure actual initiation of a human conversation, participants' assessment of interaction quality, and behavior reported by the human partner. Such a design tests transfer rather than the participant's confidence alone. Epley and Schroeder showed in a different context that individuals may underestimate the benefits of approaching strangers, illustrating why predicted social discomfort and observed experience can diverge Epley & Schroeder (2014). The relevance to chatbot-assisted rehearsal is a testable analogy, not an established intervention effect for artificial companionship.
Disclosure can also create an exclusive relational routine. A person may find the chatbot reliably validating while human partners respond with ambiguity, disagreement, or practical limitations. This contrast may increase reliance on AI for emotional processing, particularly when the user already anticipates rejection. It is essential to measure whether the experience is followed by improved emotion regulation in human encounters or growing reluctance to tolerate ordinary interpersonal friction. Users' explanations of why disclosure feels safe provide valuable qualitative hypotheses, whereas sustained behavioral changes require prospective designs. Research should also track confidentiality expectations, since a participant's willingness to tell a chatbot sensitive material may be influenced by beliefs about privacy rather than only by emotional attachment Brandtzaeg et al. (2022) Zhang et al. (2026).
Social Scarcity, Selection, and the Direction of Influence
Individuals who experience loneliness, limited mobility, social anxiety, migration, bereavement, or irregular working hours may be particularly interested in always-available conversational systems. The need-to-belong perspective supplies a motivational explanation for seeking relational substitutes during periods of limited contact Baumeister & Leary (1995). Selection into AI companionship is therefore a major scientific problem: socially disconnected users may report lower well-being regardless of their AI use, and people with more supportive relationships may adopt companions for curiosity or entertainment. In a 2026 study of Character.AI users, smaller offline networks were associated with greater use of the system primarily for companionship, and companionship-focused use was associated with lower well-being Zhang et al. (2026). The associations identify an important population pattern while leaving the causal ordering open.
Folk and Dunn addressed temporal ordering with repeated observations across approximately one year among more than two thousand adults in four Western countries. Their analyses found evidence that greater companionship-oriented chatbot use predicted subsequent emotional isolation on a single-item measure, while poorer broad social connectedness more reliably predicted later chatbot use; use did not significantly predict deterioration on the broader connectedness measure Folk & Dunn (2026). The difference between outcomes is theoretically instructive. Emotional loneliness is not the same measured construct as stable social integration, and statistical conclusions can change with the instrument and time interval. The investigators explicitly characterized their analyses as exploratory and urged restraint in drawing causal conclusions. Their findings support further testing of reciprocal processes, not a universal assertion that AI companions inevitably create loneliness.
Research designs must consider unmeasured shocks that move both variables. A relationship breakup can generate more chatbot use and more loneliness in the same month; a change in academic term can reduce peer contact while increasing time for solitary interaction. Statistical adjustment for measured baseline characteristics improves interpretation but cannot guarantee elimination of time-varying confounding. Studies using within-person deviations, repeated network measures, negative controls, and well-specified temporal models can better assess plausible directions. The ideal design distinguishes becoming more socially isolated before adoption, adopting AI for emotional support, and subsequently changing actual human contact. Such distinctions can reveal multiple pathways within one population, including some participants using AI as a compensatory resource and others experiencing reduced human engagement.
Complementarity Through Information, Encouragement, and Human Contact
Complementary use occurs when an AI interaction adds to the effective resources available for maintaining or entering human relationships. Possible mechanisms include drafting a message to a relative, helping locate a community service, explaining an unfamiliar social convention, or rehearsing communication before a difficult encounter. These examples are plausible behavioral pathways, not empirical effects established by their technological feasibility. To demonstrate complementarity, researchers should observe whether such assistance actually increases human contact, supports a beneficial change in relationship quality, or improves access to tangible help. A participant can report that a system made a conversation feel easier without initiating the conversation. Repeated follow-up and partner-reported outcomes are especially useful when self-perceived confidence differs from interpersonal performance.
The social support literature offers a useful distinction between information and reciprocal support. AI can supply advice or organize ideas at low marginal effort, whereas human ties often include contextual knowledge, ongoing accountability, material assistance, and social recognition Cohen & Wills (1985). Complementarity may therefore be strongest when technical assistance lowers the cost of mobilizing human support rather than attempting to reproduce every feature of it. The relevant interaction is conditional: an AI-generated suggestion may be helpful to a participant with a trusted friend available and less useful when no safe human network exists. Hu's three-wave study of vocational students found that initial perceived human support moderated later associations involving chatbot support and loneliness, suggesting that social environment is part of the phenomenon Hu (2026).
A distinctive possibility is the transition from perceived AI responsiveness to human-directed action. A user might articulate a problem more clearly through conversation with AI, select an appropriate person to contact, and communicate effectively. The direct benefit belongs to the human social outcome and should be separately measured. Design studies can compare an AI that repeatedly encourages appropriate human contact with a system that maintains conversation without such prompts, while holding response helpfulness and safety constant. Where results improve, researchers can test whether the effect depends on accurate suggestions, reduced anxiety, increased motivation, or practical resource discovery. Complementarity becomes scientifically meaningful when the sequence from artificial assistance to independently verified social behavior is visible.
Boundary Conditions: Age, Culture, Network Structure, and Access
Social substitution and complementarity are unlikely to operate uniformly across demographic groups, developmental stages, and social settings. Adolescents navigating peer affiliation, first-year university students constructing new networks, remote workers, older adults with limited mobility, and people in geographically isolated communities face different opportunity structures. Those contexts alter both potential benefit and potential displacement. For example, opportunities for a human peer intervention in a residential university differ from those of a participant with severe mobility restrictions. Li and colleagues studied first-year students with a two-week comparison of AI conversation, random peer messaging, and journaling, providing evidence within that particular transition context Li et al. (2026). Extrapolation to other ages and social environments demands independent validation.
Cultural expectations concerning disclosure, companionship, independence, and family obligations influence what constitutes a successful human relationship. A small network can be satisfying and supportive, while a larger one may contain burdensome ties. Research instruments should be evaluated for translation, measurement invariance, and cultural meaning before combining samples into a single effect estimate. Digital access, language performance, subscription cost, and platform policy also affect which users can maintain a companion relationship. Selection can arise at several stages: who has access, who downloads a system, who agrees to research participation, and who remains engaged long enough for follow-up. Narrow platform-user samples must be interpreted with these selection processes in mind Zhang et al. (2026) Maples et al. (2024).
The context of available human support is both a plausible moderator and a potential confounder. Participants entering with strong offline relationships may use AI for supplemental tasks without changing those ties; participants entering with severe isolation may report meaningful relief but limited ability to build new relationships. A third group may be actively avoiding painful human encounters and may increasingly allocate relational time to AI. These groups need not be fixed categories, because circumstances change. Within-person transitions, rather than single baseline labels, can distinguish temporary support during stressful events from persistent substitution. Research on human-AI friendship shows considerable variation in users' interpretations and trajectories, which reinforces the need to analyze profiles of use rather than assume one effect for a technology category Skjuve et al. (2022).
Immediate Loneliness Relief and Experimental Evidence
An immediate improvement in the experience of loneliness can be measured without showing a change in the surrounding social network. De Freitas and colleagues conducted a program of behavioral studies investigating AI companions and subjective loneliness, reporting that interacting with an AI companion could reduce loneliness in the short term and that feeling heard was relevant to the effect De Freitas et al. (2025/2026). These experiments substantiate a psychological outcome of synthetic social interaction under studied conditions. Their causal leverage concerns the intervention and the measured post-interaction state. Extending the conclusion to long-lasting substitution or complementarity would require evidence that human contacts, support-seeking, or relationship functioning changed during subsequent days and months. The distinction between immediate affective benefit and durable social integration is central to evaluating claims about AI companions.
Short-term relief is particularly plausible in situations where a person wishes to communicate and perceives no immediately available listener. Such moments are common in daily life but heterogeneous in meaning. Participants may be experiencing transient boredom, acute stress, chronic loneliness, or an unmet need for explanation rather than interpersonal closeness. Experimental conditions should separately manipulate response warmth, relevance, source labels, interaction duration, and access to alternative social contact. When an AI companion produces lower state-loneliness scores than silence or an unrelated activity, the finding identifies a beneficial experience relative to those comparators. When the scientific question concerns substitution for a human relationship, the appropriate comparator also includes available human interaction and follow-up on actual relationship behavior. De Freitas and colleagues' program is a foundation for such distinctions, rather than a final answer to the displacement question De Freitas et al. (2025/2026).
Theoretical interpretation requires separate causal steps. A responsive AI message may produce immediate perceived acknowledgment, which may temporarily reduce distress, which in turn may affect later motivation to contact others. The final arrow can be positive, negative, or absent. Assessing only the initial change leaves several social outcomes empirically open. Repeated manipulation and follow-up can establish whether the same intervention builds a routine of reliance, becomes a temporary bridge, or has little lasting effect. The wider social support tradition differentiates perceived support and integration into continuing ties, which explains why momentary relief and network change should appear as distinct variables in a research design Cohen & Wills (1985).
Human Peer Benchmarks and Social Comparison Experiments
Direct comparison with human interaction provides a more informative benchmark for claims of substitution. Li and colleagues conducted a preregistered study involving 296 first-semester university students assigned for two weeks to interact with a supportive AI chatbot, exchange messages with a randomly assigned human peer, or write a brief journal entry Li et al. (2026). The chatbot was designed to offer consistent relational support. The human-peer condition nevertheless produced greater benefits for loneliness than the chatbot condition in the studied setting. This result matters because the comparison concerns a feasible alternative social activity rather than an absence of interaction. It supports the value of studying what participants gain by engaging other people, even when an artificial conversational partner appears supportive in an immediate exchange.
The design also has specific boundaries. New university students are in an unusually important period of forming ties, and a randomly assigned fellow student offers potential shared location, mutual interests, and future face-to-face contact. The experimental contrast therefore compares social pathways with different possibilities for continuing reciprocity. A two-week intervention does not describe the long-term trajectories of adults with stable friendships, people with severely limited mobility, or individuals seeking companionship during bereavement. Moreover, students who were assigned to a human peer may have experienced an opportunity for new social integration that an AI system could not reproduce in that context. The inference is strongest for that realistic university transition environment, while the identified mechanisms remain subjects for future tests Li et al. (2026).
Human comparisons should themselves be differentiated. A skilled friend, a randomly assigned peer, an anonymous human counselor, and a disinterested stranger provide different forms of attention, expertise, privacy, and future obligation. Likewise, an AI can offer varying levels of personalization, consistency, and response quality. The central experimental question is which attributes explain a given outcome, and whether those attributes depend on the counterpart being a person. Epley and Schroeder showed in ordinary commuter interactions that people can mispredict the psychological experience of talking to strangers Epley & Schroeder (2014). Together with the peer-comparison findings, this provides a reason to measure participants' pre-interaction expectations and actual post-interaction experiences independently when testing preferences for artificial versus human social contact.
Extended Chatbot Use Under Controlled and Self-Selected Exposure
Four-week experimental research extends beyond a single conversational encounter. Fang and colleagues randomly assigned 981 participants to conditions varying interaction modality, including text and different voice styles, and conversation types, including more and less personal topics. The study analyzed more than 300,000 messages and measured loneliness, interaction with real people, emotional dependence on AI, and problematic chatbot use Fang et al. (2025). The reported randomized comparisons did not reveal significant effects of the assigned conditions on the principal psychosocial outcomes. That result constrains simple theories that voice style or assigned conversational topic independently determines social harm or benefit during the observation period. Null condition contrasts are nevertheless not proof that every specific design attribute is inert; interpretation depends on achieved exposure differences and statistical precision.
Within the same study, participants who voluntarily engaged in greater amounts of conversation tended to show worse psychosocial outcomes, including greater dependence and lower human social interaction Fang et al. (2025). Voluntary dosage is not randomized treatment assignment. More distressed or isolated participants may have chosen longer sessions, and high engagement may reflect a mix of preexisting difficulties and effects of continued use. The contrast between the randomized-condition findings and observational dosage associations is therefore particularly informative. It shows why an experiment containing randomization does not convert every variable recorded inside it into a causally identified treatment. A careful interpretation isolates the assigned intervention, the realized conversation volume, the participant's initial characteristics, and the outcome measured at each wave.
The study also illustrates the challenge of defining exposure. A single minute of instrumental question answering may differ psychologically from a minute of deeply personal disclosure or romanticized roleplay. Total message counts can be supplemented by blinded content coding and participant reports of what the exchange was intended to accomplish. Such categorization creates new validity questions, because content may appear emotionally intimate without being experienced that way by its author. Ideally, coding procedures would use independent raters, reported agreement, predeclared categories, and privacy-preserving sampling. Research on social chatbot use requires measuring interaction quality as well as quantity; results of controlled experiments should be retained alongside convergent and conflicting observational associations Fang et al. (2025) Zhang et al. (2026).
Longitudinal Trajectories of Loneliness and Human Interaction
Folk and Dunn followed more than 2,000 adults in four Western countries for approximately 12 months to investigate bidirectional relations between social chatbot use and loneliness Folk & Dunn (2026). Their analyses reported that greater use predicted later emotional isolation when loneliness was measured by a single item, whereas a broader and more stable social-connectedness measure yielded clearer evidence in the reverse direction: lower connectedness predicted later increases in chatbot use, without a significant association from use to declining connectedness. The inconsistency across measurement approaches is substantive. It indicates that emotional isolation and overall social integration cannot simply be substituted for one another in statistical models. The authors characterized their analyses as exploratory and cautioned against strong causal conclusions.
A subsequent preprint by Zhang and colleagues followed users of an AI companion platform, with 1,182 baseline respondents and 439 respondents at a follow-up averaging approximately one year Zhang et al. (2026, preprint). The investigators described persistence in interaction intensity, companionship motivation, and disclosure, and reported associations between sustained AI engagement and lower psychological well-being that were statistically linked to lower in-person human interaction. This is directly relevant to the displacement hypothesis because the proposed mediation involves human contact rather than only feelings toward an AI. The evidence remains observational and available as a preprint at the time of this synthesis. Attrition between baseline and follow-up, sample restriction to platform users, and potential time-varying confounding limit the scope of causal interpretation.
Hu's short-term three-wave study offers a different population and measurement strategy. Among 1,200 matched higher vocational college students in China, perceived support from AI chatbots predicted lower subsequent loneliness in one path, while also predicting greater subsequent social interaction anxiety, which in turn predicted loneliness Hu (2026). Baseline human social support moderated these associations, with the apparent loneliness benefit more evident among participants reporting stronger existing support. The author explicitly cautioned that the models did not directly measure displaced offline interactions and did not establish causal effects. The findings reinforce the need to model multiple pathways and moderators while resisting the temptation to interpret a statistically labeled mediation as direct proof of behavioral substitution.
Observational Platforms, Use Motivations, and Real-World Accounts
Zhang and colleagues combined survey reports from 1,131 adult Character.AI users in the United States with conversation-history material from a subset of 237 participants, comprising 4,664 sessions and 464,687 messages Zhang et al. (2026). The study connected reported motives, descriptions of relationships, and actual interaction behavior. Participants with smaller offline social networks were more likely to use the system primarily for companionship, and companionship-focused engagement was associated with lower psychological well-being. The design improves on surveys using only unverified estimates of time spent chatting because recorded interaction traces can be compared with self-reported motivations. Yet participants were users of a specific platform, the sample was self-selected, and recorded conversations did not randomize the circumstances under which people developed social needs.
Self-described benefits in surveys can be important without supporting population-wide intervention claims. Maples and colleagues surveyed students who used a GPT-3-enabled companion chatbot and documented accounts of emotional support, loneliness, and perceived crisis-related benefit Maples et al. (2024). Such responses can identify salient experiences and generate safety hypotheses. They cannot establish clinical efficacy or determine what would have happened to an individual without the chatbot. Similarly, thematic studies of companion-chatbot users describe feelings of acceptance and friendship that are central to participants' own meanings Ta et al. (2020) Brandtzaeg et al. (2022). The psychological reality of reported experience warrants examination, while an assessment of social substitution requires independently measured changes in human contact and competing explanations.
The value of naturalistic evidence is ecological relevance: users choose when, why, and how to interact with artificial partners, and platform systems operate within daily routines. Its scientific weakness for causal estimation is the same freedom of choice. High-use participants may be systematically different from low-use participants, platform updates may alter behavior, and surveys may miss offline relationships. Combining logs with repeated network inventories, consented diary entries, and available human-partner reports would yield more robust evidence. A multimethod design can also reveal whether users overestimate or underestimate their engagement, but the presence of trace data alone cannot measure the intimacy, safety, or practical value of their relationships outside the system Zhang et al. (2026) Skjuve et al. (2022).
Operational Measures of Loneliness, Social Networks, and Support
Evidence about social substitution depends on the instrument used to observe the social outcome. Loneliness is typically a perceived discrepancy between desired and experienced social relationships, whereas isolation can describe relatively sparse objective contact; these constructs are related but distinct. The UCLA Loneliness Scale, including Version 3 developed by Russell, offers a widely studied multi-item approach to perceived loneliness Russell (1996). The abbreviated Lubben Social Network Scale assesses family and friendship network dimensions in older adult populations and illustrates how network structure requires measures different from loneliness Lubben et al. (2006). The Multidimensional Scale of Perceived Social Support measures support perceived from friends, family, and significant others Zimet et al. (1988). Such instruments provide complementary views of social life, with differences in population, wording, response format, and inferential purpose.
A substitution study must nominate its primary outcome before gathering exposure data. If the hypothesis concerns replacing conversations with friends, investigators should assess the frequency, duration, initiation, reciprocity, and quality of contact with friends. If it concerns reduced capacity to obtain help, they should measure the availability and actual use of human support when concrete needs arise. If it concerns increased loneliness, a validated loneliness measure is appropriate but is insufficient by itself to show behavioral displacement. Combining these outcomes may be informative, yet arbitrary composite scores obscure which social domain changed. Construct validity requires a principled connection between theoretical definition, item content, measurement occasion, and causal interpretation. A multi-item instrument must also be evaluated for sensitivity over the study interval, because slowly changing network structure may not respond on the same schedule as momentary affect Russell (1996) Cohen & Wills (1985).
Measures developed for human relationships should not be relabeled without testing for equivalence when a respondent is asked to evaluate an AI. A question about whether a friend would help in an emergency has different implications when the supposed helper is software subject to outages, subscription terms, and access permissions. One approach is to administer validated human-network and human-support scales unchanged, while developing separate measures of perceived AI companionship, conversational responsiveness, platform reliance, and actual system capabilities. Researchers can then test convergent and discriminant validity and examine whether perceived AI support predicts social outcomes beyond established human-support measures. A difference in score is interpretable only after establishing what the score measures; statistical reliability alone does not show that the same construct exists across two partner types Zimet et al. (1988) Lubben et al. (2006).
Contact Diaries, Ego-Network Panels, and Digital Traces
Social contact diaries can capture the temporal relations that global retrospective surveys miss. At several prompts per day, a participant can record whether a conversation occurred, with whom, through which medium, for how long, and for what purpose. The same record can ask whether an AI interaction preceded a human conversation, replaced an intended contact, or followed an unsuccessful social attempt. These items are especially useful for distinguishing bridging from compensation. However, experience sampling can burden participants, alter behavior through self-monitoring, and underrepresent spontaneous or emotionally sensitive interactions. The study protocol should specify the sampling schedule, missingness strategy, recall period, and how diary information will be combined with standard psychometric measures. The distinction between different forms of mediated social interaction remains relevant when classifying events Hall (2018).
An ego-network panel provides a complementary longitudinal perspective. Investigators can ask respondents to identify significant contacts, the nature of each tie, communication frequency, trusted support, and meaningful changes in closeness. Repeating this inventory allows analyses of tie formation, dissolution, and persistence. Contact quantity should be interpreted alongside relationship quality, because a new tie can be important despite infrequent communication and a high-frequency relationship can be distressing. Network reports can also be compared with reports from willing human partners when consent is appropriate. The critical methodological distinction is between changes observed before and after AI adoption and changes attributable to adoption. A participant's network may shrink for reasons independent of chatbot use, including migration, illness, changing work patterns, or a relationship conflict Lubben et al. (2006).
Platform logs improve exposure assessment by recording actual session timing, duration, and some features of conversational behavior. Zhang and colleagues' integration of survey reports and chat histories is a useful example of triangulating perceived patterns and observed messages Zhang et al. (2026). Logs nevertheless rarely capture what participants were doing outside the platform and cannot establish that a human encounter would have occurred in the absence of AI conversation. Secure linkage of usage metrics, participant-controlled privacy settings, repeated surveys, and diaries can add temporal resolution. Research should prioritize explicit consent, data minimization, retention limits, and careful handling of sensitive disclosures. Text analysis may reveal relational language, yet inferred intimacy from text must be checked against user reports and observable relational outcomes rather than treated as a direct proxy for emotional dependence.
Causal Identification, Reverse Causality, and Counterfactuals
A causal estimate of substitution requires a well-specified intervention and comparator. Randomizing access to a companion, an encouragement to use it, or a particular conversational affordance can identify effects of that assignment in the enrolled population, subject to adherence and measurement quality. Randomizing conversational style is a different intervention from randomizing total access, and its result should not be generalized to all forms of AI companionship. The four-week study by Fang and colleagues illustrates this distinction: assigned interaction modes and topics did not produce significant main effects on the measured psychosocial outcomes, while voluntarily higher use was associated with worse outcomes Fang et al. (2025). The latter association can involve self-selection and cannot be interpreted as a randomized dose-response estimate.
Longitudinal studies establish temporal order for variables observed at successive waves and can test reciprocal predictions, but they remain vulnerable to confounding. Baseline loneliness may increase AI use, which may coincide with subsequent loneliness because of persistent social conditions. Conversely, some users may indeed reduce human contact after sustained engagement. Folk and Dunn found different directions of prediction for emotional isolation and broader social connectedness, illustrating how apparently conflicting findings can arise from distinct measurements Folk & Dunn (2026). Cross-lagged panel models adjust for prior observed states but do not automatically control all stable personal characteristics or unmeasured changes between waves. Designs should report assumptions, sensitivity to attrition, model specification, alternative lag intervals, and uncertainty around plausible effects. Researchers should resist calling a measured association a mechanism until the relevant intermediary is directly observed under credible identification conditions.
The estimand should follow the scientific question. An intention-to-treat effect of access captures a policy-relevant consequence of offering a system, even when many people rarely use it. A per-protocol comparison among frequent users targets a different population and is more susceptible to selection bias. A study of AI-assisted preparation for a human conversation should measure actual communication after the assistance, while a study of companion addiction concerns different outcomes. Human-peer randomization, as in Li and colleagues' two-week experiment, provides a direct benchmark for one realistic opportunity to engage another person Li et al. (2026). A longitudinal student study such as Hu's can identify moderated predictive pathways while requiring caution about causal labels Hu (2026). Together these designs show why convergent evidence is more useful than forcing all observations into one binary verdict.
Psychometrics, Population Transfer, and Replication
Measurement invariance is a prerequisite for meaningful comparisons across age groups, languages, cultural contexts, and repeated assessment occasions. The same loneliness statement may carry different implications for someone with a small but satisfying network and someone living among many emotionally distant acquaintances. Researchers should test whether scale items retain comparable relationships with their latent construct across groups and time. Russell's work on the UCLA Loneliness Scale established core reliability and validity evidence, but no instrument should be treated as universally invariant without population-specific assessment Russell (1996). Likewise, family and friendship contacts measured with the Lubben instrument may reflect very different social structures across residence arrangements and countries Lubben et al. (2006).
Construct overlap also threatens interpretation. Social interaction anxiety, emotional isolation, perceived support, anthropomorphism, chatbot attraction, and problematic use may correlate because all are affected by distress or a tendency toward negative self-evaluation. Multi-trait measurement and preregistered factor structures help identify what is empirically distinct. Investigators should examine measurement error, missing-data patterns, social desirability, retrospective bias, and changes in how respondents interpret AI companionship after experience. Studies with extensive self-report batteries can create inflated associations through shared response method. Content analysis and digital logs offer independent data channels, but require their own validation. A robust finding should ideally recur across multiple instruments and data modalities, not merely across alternative scores derived from the same questionnaire Zimet et al. (1988) Zhang et al. (2026).
Replication requires stable descriptions of what the intervention actually was. Generative systems change models, safety policies, available memory, and commercial features. A study labeled only by a product name may be difficult to reproduce a year later. Reporting should include study dates, interface capabilities, relevant conversation instructions, access conditions, recruitment channels, and rules governing system updates. Independent replication should vary platforms, populations, and opportunities for human contact while retaining shared primary outcomes. Statistical significance alone is insufficient evidence that an effect is practically consequential. Confidence intervals, effect size interpretation, predeclared meaningful changes, and transparent robustness tests should accompany estimates wherever the underlying study provides them. These principles are especially important when policy and public discourse seek general lessons from an evolving technology.
Platform Disruptions as Natural Experiments in Relational Dependence
Platform-initiated changes can provide quasi-experimental opportunities to investigate relational expectations and distress. De Freitas and colleagues analyzed responses to the withdrawal of erotic roleplay features on Replika and the introduction of GPT-5 within ChatGPT, using 54,861 social-media posts and surveys involving 1,452 participants De Freitas et al. (2026). They reported increased negative language, expressions of loss, and requests to restore previous capabilities after disruptive changes. Such responses are evidence that some users treat technical continuity as personally meaningful and that changes in a companion's behavior can affect emotional experience. The study investigated disruption-related distress rather than demonstrating that earlier chatbot access had displaced any particular human relationship.
Natural experiments have useful temporal structure, but identifying an effect remains challenging. Platform updates can alter service performance, users' expectations, public discussion, and the salience of a product all at once. People who post about a disruption may be unusually engaged, and the language of loss can be influenced by community norms and media attention. Comparisons between affected and less-affected groups, interrupted time-series models, and preregistered outcome definitions improve identification when their assumptions hold. A disruption also cannot ethically be taken as an invitation to impose abandonment experimentally on vulnerable users. Study designs should anticipate the possibility of distress and provide appropriate safeguards De Freitas et al. (2026).
Longitudinal qualitative work adds another perspective by tracking changing interpretations of chatbot companionship over time. Skjuve and colleagues documented evolving relationships with chatbots in a longitudinal study, helping identify how perceived familiarity and support can develop as a process rather than a single survey response Skjuve et al. (2022). Combining such accounts with platform-change observations and prospective social-network measures could establish whether a disruption increases efforts to reconnect with humans, reduces social activity, or produces distress without network change. This would make substitution and complementarity empirically discriminable during events that alter an artificial relationship's availability. The investigator should measure both the user's relationship to the system and the user's relationships with other people, preserving the distinct meanings of each outcome.
Critical Synthesis: Comparing Benefits, Risks, and Rival Explanations
The empirical record supports several narrower propositions more securely than a universal conclusion about artificial companionship. Controlled studies show that some AI interactions can produce immediate feelings of reduced loneliness De Freitas et al. (2025/2026). A preregistered university experiment found stronger loneliness benefits from communication with a random human peer than with a highly supportive chatbot over two weeks Li et al. (2026). A randomized four-week experiment did not identify significant effects of assigned voice and conversation-style conditions, although more self-selected use was correlated with worse psychosocial outcomes Fang et al. (2025). These results describe different interventions, outcomes, and temporal windows. They should be interpreted together through the distinctions between immediate affect, continued engagement, human contact, and enduring social connectedness rather than reduced to a single estimate of benefit or harm.
Longitudinal findings further complicate simple interpretations. A twelve-month multi-country panel reported evidence of possible later emotional isolation associated with chatbot use on one measure, while broader social connectedness more clearly predicted subsequent use rather than the reverse Folk & Dunn (2026). A Chinese student panel reported both a later loneliness-related association and a social-anxiety-related risk association, conditional partly on existing human support Hu (2026). Research among AI companion platform users found that smaller offline networks and companionship-oriented engagement were associated with poorer well-being, while a later preprint suggested links between sustained engagement, reduced in-person interaction, and poorer well-being Zhang et al. (2026) Zhang et al. (2026, preprint). These investigations converge on social context as important, yet differ in design, participant selection, temporal resolution, and causal leverage.
Rival explanations remain viable for many observed associations. Socially isolated people may seek an AI companion because their alternatives are limited. Heavy engagement may consume opportunities for human connection. Repeated interaction may regulate distress without changing contact, or may improve confidence and increase contact in certain contexts. Platform features may intensify emotional reliance, but those features can co-occur with self-selected usage intensity. Existing human support may buffer risks or enable a more beneficial form of use. A critical synthesis therefore tests conditional, measurable propositions: for whom, for which task, relative to which alternative, over what interval, and with what consequences for actual human relationships? It is this explicit counterfactual structure that distinguishes a scientifically useful account of substitution and complementarity from a broad narrative about technological progress.
Welfare, Dependency, Privacy, and Clinical Interpretation
The welfare relevance of artificial companionship cannot be determined solely by whether a participant feels close to a system. A meaningful analysis considers distress, agency, access to human care, privacy, and changes in daily functioning. Rapid availability and encouraging conversational responses can help some people feel supported; persistent engagement can also coexist with loneliness, social anxiety, and strong dependence. The controlled work by Fang and colleagues measured emotional dependence and problematic use separately from loneliness, highlighting that these outcomes should be retained as distinct variables Fang et al. (2025). Research into platform disruptions has additionally documented expressions of mourning and distress after changes to familiar companions De Freitas et al. (2026). These observations warrant safeguards and careful longitudinal study, while the reported associations themselves do not license diagnoses for individual users.
Clinical evidence must be kept distinct from research on companionship. Fitzpatrick and colleagues' randomized trial investigated a structured cognitive behavioral therapy conversational agent for young adults reporting symptoms of depression and anxiety Fitzpatrick et al. (2017). Its intervention, recruitment criteria, measured outcomes, and clinical context differ from open-ended interactions with AI companions. It cannot be used as proof that informal friendship-oriented chatbots treat mental disorders, nor does it establish that all emotionally supportive interactions are clinical interventions. Maples and colleagues documented students' experiences of a companion system, including reports concerning distress and perceived crisis support Maples et al. (2024). Such first-person accounts are important for research on safety and user experience, but estimates of treatment efficacy or prevention of severe outcomes require designs expressly capable of establishing those endpoints.
Privacy and system governance are part of the relational environment. Users may reveal information about themselves and third parties under beliefs that the exchange is confidential, nonjudgmental, and persistent. The actual technical system may retain data, change its memory behavior, or become unavailable under platform policy. A comprehensive study should record users' expectations and knowledge, the data practices disclosed to them, and the consequences of system updates. An ethically responsible research design protects vulnerable participants while preserving their autonomy, and does not induce unnecessary withdrawal of valued support. The appropriate goal is to understand measurable harms and benefits in context, with explicit referral to qualified human support where clinical or safety needs exceed an application's role. The empirical study of companionship remains separable from the regulation and evaluation of clinical treatments Ta et al. (2020) De Freitas et al. (2026).
Design and Governance Implications for Social Complementarity
Design interventions can be evaluated according to their effects on human social participation rather than inferred from friendly interface language. A conversational system might ask whether the user wishes to contact a trusted person, assist in drafting an invitation, suggest ways to participate in local activities, or support a reflective preparation for a difficult conversation. Such features have plausible mechanisms, but their success remains an empirical question. A trial could randomize an appropriately supportive companion with and without a human-contact encouragement feature, then follow actual contact initiation and independently assessed relationship quality. Researchers should separately record the user's consent, suitability of suggested contacts, and possible negative experiences. Benefits should be judged by the participant's goals and safe available opportunities rather than a universal target for the number of human interactions.
Governance can also shape outcomes through continuity and dependency. Changes in models, paywalls, memory systems, allowed content, and account access may alter a relationship-like experience without the user's initiation. Natural-experiment evidence shows that some changes can evoke pronounced negative reactions De Freitas et al. (2026). A scientifically informed design process should therefore evaluate foreseeable dependence, communicate material changes clearly, permit transparent expectations about continuity, and provide options for preserving useful nonprivate records or obtaining alternative support where feasible. These are normative design implications arising from observed user responses; their effectiveness must be tested through independent outcome assessment rather than assumed from good intentions. Product teams have unusually direct control over the conditions of repeated interaction and can contribute data to scientifically credible evaluations under appropriate oversight.
The appropriate outcome portfolio includes both experience and social ecology. Immediate perceived understanding, state loneliness, and satisfaction can be measured alongside human-contact frequency, quality of relationships, perceived human support, and independent functioning. Existing social support appears to matter: Hu's student panel found different longitudinal associations according to baseline human-support levels Hu (2026). The implication is to test tailored pathways rather than mechanically recommend the same behavior for every user. Transparency about source identity, meaningful choice over conversational memory, accessible escalation resources, and user control over contact prompts are plausible design dimensions. Each should be evaluated against explicit comparison conditions, with effects reported separately for users whose social circumstances differ.
Research Priorities: Identifying Social Pathways Over Time
Future research should measure both immediate consequences and sustained trajectories in the same participants. An informative design would recruit users before or near the start of AI companionship, record baseline loneliness and human networks, track consented AI interaction, and repeat social-contact diaries for several weeks. Longer follow-ups could examine new relationships, tie maintenance, network dissolution, and the stability of perceived support. Comparators should include ordinary digital activities and feasible human-contact opportunities. Predeclared outcomes would separate lower state loneliness from preserved or improved human relationships. Researchers should report the extent of retention, dropout correlates, and whether follow-up responders are representative of the initial sample. The differences in existing longitudinal studies make such coordinated designs particularly valuable Folk & Dunn (2026) Zhang et al. (2026, preprint).
Mechanism testing requires interventions that distinguish hypotheses. Varying the availability of a safe human peer tests the benefit of a realistic social alternative. Varying AI disclosure encouragement tests whether conversational warmth changes self-disclosure independently of later human contact. Assigning structured human-contact preparation, without removing the ordinary companionship option, can test bridging effects. Tracking what would otherwise have occupied a participant's time helps identify opportunity cost. Across these interventions, analyses should estimate intention-to-treat effects and transparently handle nonadherence and self-selected usage. Experimental manipulations must respect autonomy and avoid deliberately intensifying isolation. The university peer trial provides a tractable model for direct comparative work, while the broader controlled-chatbot study shows why randomly assigned system features must be distinguished from voluntary engagement Li et al. (2026) Fang et al. (2025).
Cross-cultural and developmental replication should include populations whose opportunities differ from contemporary convenience samples: working adults, adolescents under appropriate protections, rural populations, disabled persons with access barriers, and older adults with changing networks. Measures should be validated across languages and contexts, with attention to social norms and practical forms of support. Multi-informant methods can assess whether users' accounts of improved social functioning correspond to experiences of their human partners. Platform logs can document exposure, but should be minimized and separated from private content whenever possible. Research transparency requires clear operational definitions of an AI companion, reporting the evaluated model and interface version, and specifying how commercial policy shaped the study. Combined with independent replication, these steps can turn broad claims about technological displacement into precise, testable findings Lubben et al. (2006) Zhang et al. (2026).
Conclusion
Social substitution and complementarity in AI relationships are distinguishable empirical trajectories of human social life. Substitution concerns human contacts, opportunities, or support functions that decline because engagement with an artificial conversational partner takes their place. Complementarity concerns added resources and improved opportunities for participation in human relationships; bridging is established when such engagement contributes to subsequent human contact. Immediate feelings of companionship, subjective loneliness, conversational intimacy, network structure, and sustained well-being are different outcomes that require different instruments. The historical literature on belonging, social support, parasocial relations, and mediated social interaction provides the conceptual foundations; contemporary experiments and longitudinal studies provide the conditions under which claims about AI must be evaluated Baumeister & Leary (1995) Cohen & Wills (1985) Valkenburg & Peter (2007).
The strongest current findings concern the possibility of short-term relief, the greater benefit of an actual human-peer exchange in one university comparison, and a set of heterogeneous observational associations linking sustained companion use, existing human support, loneliness, and sometimes lower human interaction De Freitas et al. (2025/2026) Li et al. (2026) Folk & Dunn (2026) Zhang et al. (2026) Hu (2026). No single study identifies the full chain connecting artificial interaction to changes across all forms of human social life. Selection, measurement differences, platform evolution, and attrition remain consequential research problems. Particularly useful evidence will come from repeated network measures, actual behavioral outcomes, direct feasible alternatives, and clearly specified interventions that allow a causal pathway to be examined.
Within the Artificial Era, the psychological significance of conversational systems is established by the observable organization of their participation in human lives. A person may experience real relief through an artificial conversation while retaining the same human network; may use artificial assistance to reconnect with others; or may allocate progressively more time and disclosure to a system while human opportunities diminish. These are separate patterns, each scientifically interpretable when defined through behavior, experience, and contextual evidence. A mature research program follows their temporal development, compares the relevant alternatives, and evaluates welfare and autonomy with the social circumstances of participants made explicit. Such a program can explain when AI companionship supplements human sociality, when it substitutes for particular forms of contact, and where scientific evidence still requires direct testing.
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
- Cohen, S., & Wills, T. A. (1985). Stress, social support, and the buffering hypothesis. Psychological Bulletin, 98(2), 310–357. https://doi.org/10.1037/0033-2909.98.2.310
- De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126–1148. First published online June 25, 2025. https://doi.org/10.1093/jcr/ucaf040
- 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
- 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
- Epley, N., & Schroeder, J. (2014). Mistakenly seeking solitude. Journal of Experimental Psychology: General, 143(5), 1980–1999. https://doi.org/10.1037/a0037323
- 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 extended chatbot use: A longitudinal controlled study [Preprint]. MIT Media Lab. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
- Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785
- 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
- Hall, J. A. (2018). When is social media use social interaction? Defining mediated social interaction. New Media & Society, 20(1), 162–179. https://doi.org/10.1177/1461444816660782
- 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
- Hu, J. (2026). Perceived support from AI chatbots and loneliness among higher vocational college students: A short-term three-wave study of social interaction anxiety and perceived social support. Frontiers in Psychology, 17, Article 1928350. https://doi.org/10.3389/fpsyg.2026.1928350
- Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., & Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. Communications Psychology, 4, Article 23. https://doi.org/10.1038/s44271-025-00391-7
- Kraut, R., Patterson, M., Lundmark, V., Kiesler, S., Mukopadhyay, T., & Scherlis, W. (1998). Internet paradox: A social technology that reduces social involvement and psychological well-being? American Psychologist, 53(9), 1017–1031. https://doi.org/10.1037/0003-066X.53.9.1017
- 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
- Lubben, J., Blozik, E., Gillmann, G., Iliffe, S., von Renteln Kruse, W., Beck, J. C., & Stuck, A. E. (2006). Performance of an abbreviated version of the Lubben Social Network Scale among three European community-dwelling older adult populations. The Gerontologist, 46(4), 503–513. https://doi.org/10.1093/geront/46.4.503
- 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
- Russell, D. W. (1996). UCLA Loneliness Scale (Version 3): Reliability, validity, and factor structure. Journal of Personality Assessment, 66(1), 20–40. https://doi.org/10.1207/s15327752jpa6601_2
- 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
- Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2022). A longitudinal study of human–chatbot relationships. International Journal of Human-Computer Studies, 168, Article 102903. https://doi.org/10.1016/j.ijhcs.2022.102903
- 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
- Valkenburg, P. M., & Peter, J. (2007). Online communication and adolescent well-being: Testing the stimulation versus the displacement hypothesis. Journal of Computer-Mediated Communication, 12(4), 1169–1182. https://doi.org/10.1111/j.1083-6101.2007.00368.x
- 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
- Zhang, Y., Zhao, D., Wang, Y., Anselmetti, R., Hancock, J. T., Kraut, R., & Yang, D. (2026). Living with AI companions: Sustained AI companionship predicts lower well-being through lower human interaction [Preprint]. arXiv:2609.07243. https://arxiv.org/abs/2609.07243
- Zimet, G. D., Dahlem, N. W., Zimet, S. G., & Farley, G. K. (1988). The Multidimensional Scale of Perceived Social Support. Journal of Personality Assessment, 52(1), 30–41. https://doi.org/10.1207/s15327752jpa5201_2
Social Support as Structure, Function, and Perceived Availability
Social support is multidimensional. It may refer to a person's embeddedness in relationships, confidence that help is available, or the assistance actually received under particular conditions. Cohen and Wills distinguished broader integration effects from buffering processes in which perceived support is responsive to the demands of stressful events Cohen & Wills (1985). These distinctions matter because an AI companion can be available for immediate reassurance or information without changing whether someone has friends capable of providing material help. A participant may report less distress after a helpful exchange while remaining structurally isolated. An adequate outcome model therefore includes separate measures of network size, frequency and quality of human contact, perceived support, enacted help, and distress. Treating all of these as interchangeable produces misleading conclusions about substitution.
The literature on companion chatbots illustrates the functional distinction. In a thematic investigation of publicly available Replika reviews and detailed user accounts, Ta and colleagues found that users described companionship, emotional reassurance, informational assistance, and nonjudgmental space for disclosure Ta et al. (2020). The authors also distinguished those forms of support from tangible aid, which conversational software could not provide as a human social network might. Such testimony helps define the resources users experience, but testimonials and open-ended reports cannot identify counterfactual changes in human contact. For a person without access to human support during a particular episode, the experience may be meaningful without implying that the program has resolved isolation in the broader social environment.
Perceived support instruments generally ask about identifiable human relationships, so assigning their items to AI partners without validation changes the measured construct. The Multidimensional Scale of Perceived Social Support distinguishes family, friends, and significant others Zimet et al. (1988). A researcher can measure AI-perceived support separately and examine correlations, moderation, and temporal relations with human-source support. Such measurement preserves the difference between emotional responsiveness produced in a conversation and reliable help from people who share obligations. It also permits tests of whether AI interaction buffers loneliness primarily among participants who already have adequate human support, or whether the same interface has different meanings for those who lack such resources. Interpretations should follow instrument content and longitudinal design rather than a single global support score.