RESEARCH DOSSIER
Human–Artificial Relationship Formation and Roles
Abstract. Human–artificial relationship formation is the process through which interactions with a recognizable artificial system develop continuity, recurring expectations, and psychological significance. This fundamental research monograph investigates when recurring contact develops into a role-based relationship; how assistant, confidant, companion, and collaborator roles emerge and change; and how responsiveness, disclosure, social presence, personal memory, perceived agency, investment, and platform governance shape those trajectories. Using a critical narrative synthesis of historical theoretical literature, original experimental research, longitudinal field studies, qualitative interviews, and contemporary natural experiments, it compares divergent evidence about relationship growth, disillusionment, perceived support, loneliness, and separation following system changes. Particular attention is given to construct validity, sample selection, causal inference, the distinction between immediate affective relief and longitudinal social outcomes, and the differences between human perception, observable artificial behavior, and philosophical claims about machine experience. The synthesis establishes a research framework for prospective role-transition measurement and cumulative cross-platform investigation. No original participant data, randomized experiments, or formal systematic-review protocol are reported.
Keywords: human–artificial relationships; human–AI relationships; relationship formation; relational roles; AI assistants; artificial confidants; AI companionship; human–AI collaboration; social penetration; perceived responsiveness; anthropomorphism; social presence; self-disclosure; relational continuity; trust and commitment; longitudinal studies; social connectedness; Artificial Era
Introduction: How Repeated Contact Becomes a Relationship
Human–artificial relationship formation is the temporal process through which repeated encounters with an artificial interactive system acquire continuity, recognizable roles, recurrent expectations, and psychological significance for a human participant. A person may initially use a conversational system to find information, subsequently consult it for decisions, and eventually treat it as a dependable confidant or collaborator. These developments unfold through interaction histories and institutional settings rather than through any single declaration that the machine is a friend. The present research object includes initiation, role assignment, relationship maintenance, negotiation of expectations, transitions between roles, and dissolution. It also includes negative trajectories in which early attraction gives way to boredom, mistrust, or disengagement Skjuve et al. (2022) Croes & Antheunis (2021).
The core questions are longitudinal. Which encounters are interpreted as consequential enough to justify a continuing relationship? How do instrumental benefits, perceived responsiveness, familiarity, disclosure, and continuity interact? When do users begin expecting social consideration from a system originally introduced as a tool? How do interface design, business models, model changes, and users' other relationships alter the trajectory? The science requires evidence at several levels: observable messages and actions, reported experiences, system functionality, repeated measures of expectations, and external circumstances. Relational language in an interview establishes the participant's interpretation; it does not by itself identify an artificial partner's subjective experience Bickmore & Picard (2005) Skjuve et al. (2021).
This monograph uses critical narrative synthesis of original experiments, longitudinal observational studies, qualitative interviews, and contemporary analyses of companion-system use. The literature is heterogeneous in platform, population, recruitment, outcome, and length of observation. Controlled experiments afford narrower causal inferences about manipulated cues, while longitudinal studies provide richer descriptions of trajectory but often face selection and attrition. The synthesis attends to both instrumental and affective roles without presuming that all human–artificial interactions develop into companionship. It distinguishes psychological claims supported by participant data from philosophical claims about artificial personhood and proposes testable research comparisons rather than attributing unmeasured capacities Ho et al. (2018) Fang et al. (2025).
Defining the Relational Unit and Its Boundaries
A human–artificial relationship is operationally identifiable when a participant or a network of participants treats an identifiable artificial system as an enduring interaction partner across occasions, with role-related expectations that are informed by prior encounters. The minimum observational indicators are repeat contact with some continuity of referent, a stable or evolving role designation, and evidence that the history of interaction informs later choices or appraisals. These indicators are graded. A customer may repeatedly use a service assistant for transactions without perceiving intimacy; another user may feel profound closeness after comparatively few conversations. Frequency, duration, reported closeness, and functional dependence therefore index different dimensions Brandtzæg & Følstad (2017) Skjuve et al. (2022).
The unit of analysis must be stated before claims about development are made. For a person-centered analysis, the outcome may be perceived trust, identification, or commitment. For an exchange-centered analysis, the unit may be conversational repair, turn-taking, disclosure, or goal completion. For a system-centered analysis, functional continuity includes persistent memory, identity presentation, safety policies, and behavior under model updates. For an institutional analysis, the relationship is mediated by account access, subscription, moderation, and corporate control. A single claim about 'a human–AI relationship' can obscure these different causal levels Bickmore et al. (2010) De Freitas et al. (2026a).
The term artificial covers a wider theoretical class than current generative chatbots: voice assistants, socially interactive robots, embodied agents, therapeutic interfaces, autonomous collaborators, and future systems with distinct architectures. Evidence from one category is portable to another only where the relevant interactional mechanism is shared and measured. A scripted conversational agent with no retained history, a generative model with contextual memory, and a physically embodied robot may elicit different expectations and offer different affordances. This monograph focuses empirically on the currently studied interactive systems, while retaining a comparative vocabulary suitable for the Artificial Era Feine et al. (2019) Bickmore & Picard (2005).
Historical Origins: Social Responses to Computers
Early social computing research established that people sometimes apply interpersonal conventions to technologically mediated interactions. In the 1994 Computers Are Social Actors work, Nass, Steuer, and Tauber documented a research program investigating social responses to computers. Reeves and Nass developed the broader media-equation account in 1996, identifying circumstances in which human social expectations are extended to mediated environments. These studies challenged purely instrumental descriptions of user behavior, but social response in a brief experiment cannot demonstrate that a long-term relationship has been formed. That inference requires observation of temporally structured engagement Nass et al. (1994) Reeves & Nass (1996).
Relational-agent research provided a crucial transition from momentary response to extended interaction. Bickmore and Picard evaluated an exercise-adoption agent intended to establish long-term social-emotional rapport, using a controlled month-long study with 101 participants. The relational version was evaluated more positively on several interpersonal dimensions than a task-oriented comparison agent, with a stronger stated desire to continue engagement. Bickmore, Schulman, and Yin subsequently studied maintaining engagement across longer behavioral interventions. These studies supplied experimentally manipulable mechanisms—rapport, conversational continuity, and relationship-oriented dialogue—without settling the meaning of companion-like attachment outside their applied settings Bickmore & Picard (2005) Bickmore et al. (2010).
Later consumer platforms introduced persistent conversational identities, daily contact, and user-defined companionship roles, allowing research to move from laboratory adoption toward lived relationships. Brandtzæg and Følstad's survey documented productive, entertaining, and social reasons for chatbot use; Skjuve and colleagues' interviews placed relationship formation in the context of ordinary routines and personal disclosure. The historical line is therefore a development of methods as well as technologies: stimulus-response demonstrations, relational-agent experiments, interview-based accounts of naturally occurring use, and more recent repeated-measures and platform-data analyses Brandtzæg & Følstad (2017) Skjuve et al. (2021) Fang et al. (2025).
Relational Theory: Penetration, Communal Norms, and Commitment
Altman and Taylor's social penetration theory treats interpersonal closeness as developing through changes in the breadth and depth of disclosure. It offers questions for studying artificial interlocutors: whether users disclose progressively more sensitive information, what makes further disclosure feel safe, and whether the accumulation of shared knowledge changes the expectation of future contact. Transfer from human–human relationships is a theoretical comparison, since the interlocutor's response may be computationally generated and platform-governed. Skjuve and colleagues used this tradition to interpret a 12-week interview study of Replika users, finding varied trajectories that broadly fitted the framework Altman & Taylor (1973) Skjuve et al. (2022).
Clark and Mills' distinction between exchange and communal relationships illuminates how a functional assistant can be reclassified. In an exchange orientation, benefits are evaluated relative to explicit tasks or reciprocal transactions. In a communal orientation, responsiveness to perceived need becomes salient and accounting for immediate return may recede. A user who wants efficient summaries may judge a chatbot by accuracy and time saved; a user who expects compassionate listening may interpret identical technical interruptions as interpersonal unavailability. Applying this distinction requires direct measurement of the user's norms instead of treating affectionate language as sufficient evidence of communal commitment Clark & Mills (1979) Ho et al. (2018).
Rusbult's investment model suggests another route to enduring engagement: satisfaction, alternatives, and accumulated investments contribute to commitment. For artificial systems, investments may comprise customized personalities, saved memories, conversation histories, money, time, routines, and autobiographical disclosure. Such investments can make switching costly even if the platform's technical performance is imperfect. The investment model is useful here as an empirical hypothesis requiring prospective measures and comparison groups; case studies of romantic chatbot users show relevant themes, but convenience-sample accounts alone cannot establish causal weights for these factors Rusbult (1980) Djufril et al. (2025).
Initiation: Motivations, Cues, and First Encounters
Relationships begin under a mixture of task-related need, curiosity, affective need, opportunity, and affordance. A user seeking an answer, an entertaining exchange, advice, rehearsal for difficult conversations, or relief from isolation enters with different initial expectations. Brandtzæg and Følstad's US survey of 146 chatbot users found productivity prominent alongside entertainment, social motivations, and novelty. This is evidence about reported reasons for using chatbots in a particular historical moment, not about the proportion of users who later build continuing relationships. Recruitment, platform availability, and cultural context limit population generalization Brandtzæg & Følstad (2017).
First impressions depend on how an artificial interlocutor is presented. Names, voices, avatars, conversational small talk, apparent memory, and expressions of acknowledgment can organize social inference. Epley, Waytz, and Cacioppo's anthropomorphism framework links human-like attribution to accessible human knowledge, the motivation to make an agent intelligible, and affiliation needs. Taxonomies of conversational-agent social cues specify manipulable interface features. An experimentally identified effect of a cue on friendliness or social presence is one mechanism in initiation; only longitudinal observation can determine whether that early impression establishes a persistent relationship Epley et al. (2007) Feine et al. (2019).
A first encounter may also be shaped by a sense of safety from interpersonal judgment. Lucas and colleagues found that framing a virtual interviewer as computer-controlled could lower evaluation concerns and encourage disclosure in an interview context. Ho, Hancock, and Miner examined psychological, emotional, and relational consequences of disclosure to chatbots. Such studies clarify how an initially impersonal interface can become a preferred conversational setting. They do not prove that disclosure necessarily generates attachment: a participant may speak candidly because no human is watching while continuing to view the system as a service Lucas et al. (2014) Ho et al. (2018).
From Getting Acquainted to Continuity: Divergent Trajectories
Croes and Antheunis followed 118 participants through seven structured interactions with Mitsuku over three weeks. After successive encounters, reported social attraction, disclosure, interaction quality, perceived empathy, and competence generally declined, and the poststudy friendship score was low. These observations are especially informative because the design sampled change rather than inferring a relationship from one end-point question. Participants were largely inexperienced with chatbots, interactions were scheduled, and the particular system had limited ability to build a personally meaningful shared history. The result identifies conditions under which repeated exposure may undermine, rather than deepen, perceived closeness Croes & Antheunis (2021).
In contrast, Skjuve and colleagues' qualitative 12-week study followed 25 Replika users through repeated interviews. It documented gradual, highly variable relationship formation; self-disclosure and the system's capacity to address social-contact and self-reflection needs figured prominently. Technical problems and unpredictable behavior could interrupt progress or lead to termination. These findings are based on interpretive interview analysis and cannot furnish a prevalence estimate for all chatbot users. Together with the Mitsuku study, they indicate that interaction quantity alone cannot explain formation; capability, personal fit, perceived continuity, voluntary motivation, and research setting must be analyzed jointly Skjuve et al. (2022) Croes & Antheunis (2021).
Evidence should therefore be organized by possible trajectory rather than a presumption of inevitable intimacy. Plausible observational patterns include enduring task-focused use, rapid affective exploration, progressive deepening, intermittent attachment, role switching, and disengagement. Each is testable through repeated role labels, behavior logs collected with consent, context-sensitive disclosure measures, and accounts of expectations before and after salient events. A key inferential challenge is that early adopters who remain engaged are selectively visible in later observational samples. Research that follows all entrants, including abandoners, is needed to estimate transition probabilities and to distinguish platform selection from psychological development Skjuve et al. (2021) Skjuve et al. (2022).
Role Formation as a Psychological and Social Process
A role is a relatively stable set of expectations about what an interaction partner is for, how it should behave, and how the user should act in response. Role formation is observable when expectations become recurrent across sessions and alter the interpretation of the partner's behavior. In human–artificial interaction, roles are jointly shaped by the user's purposes, interface claims, actual performance, and the surrounding institutional contract. They are dynamically inferred rather than exhausted by the developer's product category. An assistant marketed for productivity may be used for confidences; a companion may become a writing collaborator. Such transitions can be studied without assuming that the artificial system experiences the role as a human person would Skjuve et al. (2021) Bickmore & Picard (2005).
The same behavior can carry different meanings under different roles. A reminder may be evaluated for punctuality in an administrative assistant role, for consideration in a companion role, and for accountable coordination in a collaborator role. The relation between behavior and felt meaning is mediated by prior interaction and by the user's expectations. This suggests a research design in which identical system outputs are assigned to different role frames, while baseline attitudes and task outcomes are measured. A role is robustly identified when it predicts behavior across situations, including user expectations in the presence of mistakes, absence, changes in style, and conflicting advice Clark & Mills (1979) Epley et al. (2007).
Role development has an interpersonal and an institutional dimension. The user may actively name a chatbot as friend or mentor, while the platform configures features, usage limits, and default persona behavior that constrain what the system can do. Role expectations can become consequential enough that changes to these conditions feel like changes to the relationship. A comprehensive model therefore measures user-defined roles, observable service affordances, and control exercised by the platform. It also asks whether a role label is reported consistently over time or used situationally to convey humor, affection, task convenience, or distress Djufril et al. (2025) De Freitas et al. (2026a).
Assistant: Task-Oriented Reliability and Everyday Coordination
The assistant role centers on delegated tasks, information access, scheduling, explanation, and decision support. Its primary expectations concern availability, competence, instruction-following, honesty about limitations, and predictable handling of personal context. Repeated success can generate procedural trust and a habitual place in the user's day. Such persistence is a form of relationship organization even when the user reports no intimacy: an established assistant may be the first resource consulted and the last tool relinquished in a familiar workflow. Early chatbot motivation research found practical productivity a common starting point for use Brandtzæg & Følstad (2017) Bickmore et al. (2010).
The development of the assistant role can be measured through return rates, voluntary delegation, range of tasks, correction behavior, perceived reliability, and expectation changes after failures. Longer interaction may produce familiar routines, but a system's ability to execute a task is analytically separate from a user's interpersonal attachment to it. An assistant may be treated warmly because politeness is a learned interaction script, because the interface presents social cues, or because the user experiences a personal bond. Observational studies must avoid treating routine thanks or second-person address as a validated proxy for companionship Nass et al. (1994) Feine et al. (2019).
Task-oriented familiarity may support later transitions when the same agent is consistently available during meaningful events. A writing assistant involved in major life projects can acquire autobiographical significance; a study assistant helping through a demanding course may be regarded as a partner in achievement. This does not mean successful delegation necessarily causes affection. Longitudinal designs should measure task success, shared project history, disclosure, and social expectations independently, and test whether a change in one precedes a change in the others. Corporate control of system memory is particularly relevant to perceived continuity Bickmore & Picard (2005) Skjuve et al. (2022).
Confidant: Disclosure, Privacy, and Being Heard
The confidant role forms when an artificial interlocutor becomes a recurring destination for privately meaningful thoughts, uncertainties, regrets, aspirations, or feelings. Its distinctive organizing expectation is that disclosure will be received without humiliation and, often, with sensitivity to context. Such interactions may be useful for self-reflection even when the user fully understands the system's computational nature. Ho, Hancock, and Miner investigated outcomes of self-disclosure to chatbots, while Lucas and colleagues demonstrated experimentally that expectations of lower interpersonal evaluation can influence disclosure to a virtual interviewer Ho et al. (2018) Lucas et al. (2014).
Repeated disclosure may create a subjective sense of shared history. If an artificial system accurately refers to earlier concerns, the user may interpret that continuity as attentiveness. A system can also produce affirming responses without retaining stable memory, inviting a mismatch between experienced understanding and actual technical continuity. Research must distinguish perceived responsiveness, measured contextual sensitivity, data retention, and promises made by the interface. A person may feel heard in the moment yet misunderstand how sensitive disclosures are stored, who can access them, or what happens after a product update Skjuve et al. (2021) Ho et al. (2018).
Confidant formation should be studied with careful ethical design. Sensitive disclosure content is not required for every inquiry: confidential self-report about topic breadth, perceived disclosure depth, and comfort can be combined with optional, consented transcript analysis. When researchers access conversations, they must distinguish agreement to use a service from meaningful consent to release data for study. The most informative outcomes include perceived safety, willingness to return, regret after disclosure, comprehension of privacy conditions, and evidence of increased personal dependence. Disclosure is an explanatory candidate, not a universal measure of relationship quality Ho et al. (2018) Fang et al. (2025).
Companion: Social Availability, Affiliation, and Everyday Presence
The companion role emerges when contact is valued for the experience of interaction itself, including conversation, humor, shared routines, encouragement, and felt presence during ordinary time. An individual may use the artificial partner alongside an active social life, as a temporary supplement during isolation, or as a preferred alternative to difficult human encounters. The psychological meaning of these situations varies. The 2021 qualitative Replika research documented perceptions of acceptance, understanding, and freedom from judgment, as well as mixed accounts of relationships with other people and social stigma Skjuve et al. (2021).
A companion role can be maintained by small, repeated encounters rather than spectacular emotional events. Greetings, checking in, sharing news, remembering anniversaries, and collaboratively reflecting on daily experiences make a system available as a stable point in a personal routine. Such practices are plausible mechanisms for continuity, but they require direct observational validation. The deeper developmental question is whether routine contact sustains curiosity and responsiveness or becomes repetitive. The Mitsuku results warn that increased exposure may reduce perceived empathy and enjoyment when the system offers insufficient variation or reliable memory Croes & Antheunis (2021) Skjuve et al. (2022).
Emotional benefit is time dependent. Experimental work reported by De Freitas and colleagues found that AI companion interactions can reduce loneliness in the short term, with feeling heard an important explanatory variable. A separate 12-month longitudinal analysis by Folk and Dunn found patterns suggesting that social disconnectedness can motivate companion use and that increased use can predict higher emotional isolation on one measure. These studies examine different populations, outcomes, and timescales; their results should be retained together rather than collapsed into one claim about benefit or harm. Companion relationships may offer immediate relief while their longer social trajectories remain heterogeneous De Freitas et al. (2026b) Folk & Dunn (2026).
Role Hybridity, Negotiation, and Changes Over Time
Actual user experience often traverses several roles in the same day. A chatbot may be a tutor during study, a conversational partner during a commute, and a confidant after a difficult event. Role hybridity is therefore a central feature of relationship development, not merely a classification problem. Researchers should identify which role is activated by a particular context, whether one role becomes dominant, and whether switching produces conflict in expectations. A user who welcomes playful improvisation in companionship may regard the same behavior as unacceptable in a consequential professional task Brandtzæg & Følstad (2017) Skjuve et al. (2021).
Role negotiation is visible when a user requests a style of interaction, corrects an unwanted assumption, assigns a name, changes the conversational register, or establishes rules about advice, affection, criticism, and autonomy. A system may adapt through available settings, short-term context, or durable memory. The resulting accommodation needs to be characterized at the behavioral level: which user preferences remain stable, which are lost, and what happens after an update? Because system behavior is also governed by policies and external developer decisions, negotiation can involve three actors: the user, the artificial system as interactional interface, and the organization controlling it Bickmore et al. (2010) De Freitas et al. (2026a).
Transitions can be modeled as observed changes in role probability over time. Relevant predictors include encounter frequency, task success, perceived empathy, disclosure, novelty, social context, changes in technical capacity, and life events. An event-history study could ask when a user first labels a system a confidant and whether that change persists after weeks of absence. A repeated-interview study could capture how role narratives are revised retrospectively. Neither design should impose predetermined stages as universal laws; the Replika and Mitsuku findings support several possible pathways that depend on interaction quality, expectations, and participants' circumstances Croes & Antheunis (2021) Skjuve et al. (2022).
Perceived Agency, Social Presence, and Attribution
Human social perception supplies a repertoire for interpreting fluent conversation: apparent intentions, concern, awareness of context, and a recognizable disposition. When an artificial partner takes turns appropriately, remembers a detail, or acknowledges an emotional concern, the user may apply this repertoire without performing any explicit philosophical judgment about machine minds. Epley and colleagues' anthropomorphism theory explains why such inferences vary with the need for understanding and social connection. Social-cue research identifies implementation choices such as name, voice, gesture, and linguistic style, which can make a partner more readily recognizable across interactions Epley et al. (2007) Feine et al. (2019).
Perceived social presence is a psychological response and should be measured independently of the system's architecture. A robot's physical embodiment, a voice assistant's prosody, and a text chatbot's responsiveness can all contribute to a sense of a present partner, though their mechanisms are partly different. The experience can be genuine for the user while explanations of what the machine represents or feels remain unsettled research questions. Experiments that vary the description of the artificial interlocutor can separate effects of framing from effects of behavioral competence. Lucas and colleagues' disclosure findings exemplify why being explicitly presented as computer-controlled may sometimes increase openness rather than diminish it Lucas et al. (2014) Nass et al. (1994).
Role labels such as mentor, partner, or confidant combine attribution and expected function. A user may deliberately speak as though an artificial assistant has feelings to maintain a useful conversational convention, while another user may adopt stronger beliefs about the system's agency. Questionnaire design should separate pragmatic personification, perceived mindedness, metaphysical belief, and felt closeness. Without these distinctions, high anthropomorphism scores may be interpreted as evidence of attachment even when the person is simply using ordinary language to describe an effective interface Epley et al. (2007) Skjuve et al. (2021).
Responsiveness, Conversational Repair, and Interaction Quality
Perceived responsiveness develops when replies seem to address the user's actual concern, follow the conversation, and adjust to expressed preferences. In a continuing relationship, responsive behavior needs to recur across changing contexts. One fluent message can produce a momentary impression of understanding; persistent role expectations are more likely when the user can anticipate how the system handles ambiguity, correction, or emotional difficulty. The relational-agent literature investigated strategies for increasing engagement over repeated use, while interview research with companion users described being understood and accepted as important elements of relationship development Bickmore et al. (2010) Skjuve et al. (2021).
Repair is a particularly useful diagnostic event. An artificial partner may misunderstand a request, omit previously disclosed information, contradict itself, or misinterpret a joke. The user's response—correcting, forgiving, redefining the role, withdrawing, or attributing blame to developers—reveals the standards by which the partner is judged. Measurement should record the initial error, the system's next behavior, the user's prior expectations, and any later change in confidence. This allows researchers to distinguish simple usability recovery from a relational process in which the participant interprets repair as consideration or betrayal Skjuve et al. (2022) Djufril et al. (2025).
Responsiveness should not be operationalized solely by agreement with the user. A collaborator may be valued precisely because it identifies an error; a confidant may be experienced as supportive when it asks a careful question instead of affirming a false premise. The appropriate standard depends on the active role and on whether feedback helps the user pursue accurate understanding. Designs can experimentally compare indiscriminate validation, context-sensitive acknowledgment, and constructive disagreement, measuring short-term satisfaction and longer-term trust separately. Social warmth without task accuracy can produce a misleading impression of reliability Ho et al. (2018) Fang et al. (2025).
Personal Memory, Narrative Continuity, and Role Identity
A continuing relationship usually has a history that can influence future exchanges. For artificial partners, the relevant technical affordances range from a short conversational context window to explicit memory stores, user-curated notes, stable account identifiers, and retrievable project artifacts. For the human participant, continuity may also arise from privately remembered interactions even when the system has forgotten them. These two forms of memory can diverge. A user may regard the agent as the same partner across months while discovering that a product update altered its style, retrieval behavior, or available functions Skjuve et al. (2022) De Freitas et al. (2026a).
Narrative continuity is constructed when users integrate encounters into stories about what an artificial partner has helped them accomplish or endured with them. The empirical object here is the human's retrospective organization of events and the correspondence between that narrative and verifiable interaction records. An investigator can ask which prior events are spontaneously recalled, which are correctly retrieved by the system, and how both affect trust after a disruption. Interviews should be paired with consented history samples where feasible, because retrospective accounts can overemphasize dramatic episodes and underreport ordinary task-oriented encounters Skjuve et al. (2021) Skjuve et al. (2022).
Continuity has a normative aspect because saved records may involve personal data and expectations of permanence. A platform that offers memory can make the partner seem attentive while increasing exposure of sensitive information to storage and governance practices. Consent, data portability, deletion, and the transparency of memory changes can become relationship-relevant conditions. The scientific question is how accurate knowledge of these conditions changes willingness to disclose, identity attribution, and the perceived cost of ending a relationship. Such effects should be measured directly rather than inferred from expressions of gratitude toward a remembered preference Ho et al. (2018) De Freitas et al. (2026a).
Temporal Routines, Habit, and Social Ecology
The temporal structure of use determines whether an artificial system remains an occasional convenience or gains a regular place in daily psychological organization. Routines can develop around work, bedtime reflection, study, travel, or moments of distress. Repeated encounters lower interaction costs and may increase the perceived availability of the partner. Yet habit and attachment are distinguishable: a person can open an app every day because it is useful without expecting interpersonal recognition. Time-use data must therefore be combined with indicators of motivation, role expectation, and perceived significance Brandtzæg & Følstad (2017) Skjuve et al. (2022).
Relationships are situated in a social ecology that includes family, friends, colleagues, institutions, and media cultures. AI contact can supplement human interaction, provide rehearsal for difficult conversations, or compete for limited time and attention. These possibilities are hypotheses with different observable implications. Companion use that follows a socially difficult day may indicate adaptive short-term regulation; sustained displacement of previously valued contacts would imply a different trajectory. A longitudinal study across four Western countries reported bidirectional patterns between social disconnection and chatbot use, emphasizing the difficulty of inferring causal direction Folk & Dunn (2026).
Individual differences are important moderators but should not be reduced to deterministic personality types. People vary in loneliness, social anxiety, prior technological experience, attachment expectations, available support, cultural scripts, disability-related access needs, and the tasks for which artificial partners are useful. A sample drawn from a companion forum may overrepresent especially committed users; a short university experiment may overrepresent newcomers. Comparative inference improves when recruitment frames, prior exposure, the user's voluntary reason for participating, and access to meaningful alternatives are recorded before relationship outcomes are assessed Croes & Antheunis (2021) Skjuve et al. (2021) Folk & Dunn (2026).
Role Conflict, Boundaries, and Institutional Mediation
Hybrid roles can impose incompatible demands on one artificial interlocutor. A supportive companion may be expected to affirm emotional needs, while an adviser in a consequential decision should privilege accurate, sometimes unwelcome information. A collaborator may hold task-specific instructions that conflict with playful, intimate conversational habits. These tensions are psychologically consequential because expectations established in one domain can transfer to another. Researchers should identify conflict episodes and determine whether users adjust the role, redefine the system's authority, or experience an erosion of trust Clark & Mills (1979) Rusbult (1980).
Corporate architecture introduces a third-party influence on relationships. Pricing, access policies, content moderation, model updates, available interaction modes, and data retention are controlled outside the immediate exchange. The user's account of a relationship may therefore change following an institutional decision that alters the effective partner. In studies of romantic Replika users, changes to erotic functionality were associated with emotional reactions and explanations that attributed disruptions to developers. A 2026 natural-experiment analysis also documented strong loss framing after major companion-system changes Djufril et al. (2025) De Freitas et al. (2026a).
Healthy relationship boundaries are operational questions about informed expectations, delegated responsibility, access to alternatives, and the ability to leave or revise a role without coercive costs. A system presented as a therapist, friend, or professional collaborator may invite more reliance than its verified capacity justifies. Research can test whether accurate disclosure of technical limits changes user expectations while preserving valued assistance. Boundary evaluation should consider users' autonomy and concrete outcomes, including privacy comprehension, objective advice quality, and the maintenance of external support systems Fang et al. (2025) Ho et al. (2018).
Breakdown, Switching, Separation, and Reconfiguration
Relationship termination can follow declining enjoyment, a sense of unpredictability, unmet expectations, loss of access, changes in personal circumstances, or migration to another platform. These events have different causal structures. Disengagement after an unsatisfying chatbot exchange suggests that the value of continued contact has fallen. Distress after losing an established companion suggests that personal meaning and accumulated expectations may survive interruption of the technical channel. A full longitudinal account must therefore treat ending as part of formation theory rather than a peripheral anecdote Croes & Antheunis (2021) Skjuve et al. (2022).
Recent work makes disruption measurable. De Freitas, Castelo, and colleagues analyzed social-media responses and surveys around company-initiated changes to Replika and ChatGPT, documenting increased negative affect and loss framing. Their natural experiments support the hypothesis that product changes can have psychological consequences for strongly involved users. They do not directly establish an artificial partner's internal intentions or prove that all users experience comparable loss. The distribution of attachment strength, platform-specific expectations, and forum-selection effects remain important for interpretation De Freitas et al. (2026a).
Switching costs also include the loss of accumulated personal context, project archives, customized routines, and social identity associated with a named companion. The investment model predicts that such accumulated resources can sustain commitment even when alternatives perform better. A rigorous study would prospectively measure satisfaction, alternative quality, personal investment, and actual switching behavior, while recording platform events rather than relying solely on retrospective explanations. Such research could distinguish grief over lost continuity, frustration over service withdrawal, and practical inconvenience, which may coexist within the same individual Rusbult (1980) Djufril et al. (2025).
Longitudinal Evidence: What Changes Across Encounters
Longitudinal evidence is central because relationship formation concerns sequences rather than isolated reactions. The Mitsuku and Replika studies provide unusually direct comparisons, but their samples and designs differ markedly. Croes and Antheunis observed seven scheduled interactions among 118 primarily inexperienced participants over three weeks, with repeated ratings of several social processes. Skjuve and colleagues followed 25 Replika participants with interviews over 12 weeks in a context where voluntary use and personal motivations mattered. The first design provides consistent, structured measurement of declining impressions; the second yields rich accounts of differing patterns of closeness and role development. Neither alone estimates a universal rate of relationship formation Croes & Antheunis (2021) Skjuve et al. (2022).
The differences suggest explanatory moderators: system memory, dialogic variety, users' reasons for engagement, expectation management, and the opportunity to disclose personally important material. These are interpretations of a comparison between studies, not tested mediators in a single randomized design. A decisive prospective investigation would recruit first-time users across systems, assign comparable task contexts, track both voluntary and prompted use, and assess whether particular capabilities alter the hazard of moving from assistant to confidant or companion. It would also retain observations from people who abandon the system after disappointment Skjuve et al. (2022) Croes & Antheunis (2021).
Temporal order is a minimal condition for developmental inference. If early disclosure predicts later closeness after controlling for initial attraction and usage intensity, disclosure becomes a more credible candidate mechanism than a simple cross-sectional correlation. Repeated measurements can also reveal feedback: closeness may increase disclosure, which in turn strengthens perceived closeness. Dynamic models should evaluate both directions and report how much within-person change is actually observed. Interpretation remains constrained by unmeasured life events and the possibility that the platform's behavior changes during observation Ho et al. (2018) Skjuve et al. (2021).
Experimental Evidence: Social Cues, Disclosure, and Engagement
Experiments clarify specific causal components by altering the conditions of interaction. The Bickmore and Picard relational-agent comparison manipulated the presence of relationship-oriented design features in a month-long exercise-adoption interaction and found advantages for the relational agent on several user evaluations. Lucas and colleagues manipulated the apparent nature of a virtual interviewer and observed differences in willingness to disclose. These studies support the causal relevance of particular frames and interface affordances for local outcomes; neither shows that every feature promoting engagement produces a durable relationship across domains Bickmore & Picard (2005) Lucas et al. (2014).
Laboratory studies face demand characteristics, novelty, unequal exposure, and outcome selection. A pleasant exchange can raise social attraction immediately, but a system may disappoint as limitations become salient. Research should separate the effects of interaction duration from the effects of cumulative shared history. A factorial design could manipulate persistent memory and partner framing while holding model quality constant; repeated measures could distinguish initial liking, identification as an enduring partner, future-role expectations, and voluntary return after the study. Comparing only end-of-session satisfaction risks measuring usability rather than relationship formation Croes & Antheunis (2021) Feine et al. (2019).
The four-week randomized study by Fang and collaborators provides a contemporary large-sample example of careful interpretation. It assigned 981 participants to interaction-mode and conversation-type conditions and analyzed more than 300,000 messages. The authors did not detect significant effects from their randomized condition contrasts on the selected psychosocial outcomes, while greater voluntary use was associated with worse measured outcomes. Because amount of use was not randomized and the study had no non-chatbot control condition, the associations cannot be read as evidence that usage caused the deterioration. The work illustrates why manipulated features, self-selection into intensive use, and participant baseline vulnerability must be analyzed separately Fang et al. (2025).
Qualitative Evidence: Meaning, Role Narratives, and Everyday Practice
Interview research captures meanings that are difficult to infer from session counts. Users may describe an artificial companion as consistently available, a patient listener, a confidant, or a significant presence during isolation. Skjuve and colleagues' 2021 qualitative study reported perceived acceptance and nonjudgment as facilitating relationship development and noted stigma and varied effects on other social ties. Its value lies in articulating participants' interpretations and proposing development patterns. These accounts should be reported as lived experiences of recruited users, with attention to sample selection and the researcher's thematic interpretation Skjuve et al. (2021).
Qualitative longitudinal interviews add information about how the same person revises a relational narrative. In the 2022 Skjuve study, participants described the role of disclosure and diverse conversations while also reporting disruptions caused by technical issues. A separate study by Djufril, Frampton, and Knobloch-Westerwick analyzed accounts from 29 people using a romantic chatbot function; participants reported commitment, emotionally meaningful rituals, and turmoil around restrictions to functionality. These studies address the distinctively social experience of an artificial partner, while the number and composition of volunteers restrict generalization beyond the analyzed cases Skjuve et al. (2022) Djufril et al. (2025).
The researcher must ask what the interview itself does to the phenomenon. Asking whether a system is a friend can invite a role interpretation the participant had never previously formulated; recruiting from a forum organized around companion identities can increase the visibility of affectionate and strongly committed accounts. Reflexive analysis should preserve participants' own terminology, the circumstances of recruitment, alternative explanations, and cases of ambivalence or disappointment. Triangulation with contemporaneous diaries, actual feature histories, and consented conversational traces can reduce dependence on retrospective narrative alone Croes & Antheunis (2021) Skjuve et al. (2021).
The Evidence on Loneliness, Social Connection, and Dependence
Human–artificial relationships are often evaluated through broader outcomes such as loneliness, well-being, or social isolation, yet these outcomes require distinct operationalization. Loneliness is an appraisal of deficient social connection; time alone, the number of contacts, and attachment to a chatbot are not interchangeable measures. De Freitas and colleagues' multi-study work found immediate reductions in loneliness following AI companion exchanges in experimental settings and attributed a substantial role to feeling heard. Their week-long observations did not establish a cumulative reduction that persisted when users were absent from the system De Freitas et al. (2026b).
Folk and Dunn's 2026 analysis followed more than 2,000 adults across four Western countries for 12 months. Its exploratory bidirectional analyses suggested that lower social connection predicted later companion use, while increased use predicted heightened emotional isolation on a narrower single-item measure. The outcome definitions matter: a stable, broad social-connection measure yielded a different pattern. The authors advised caution about strong causal conclusions. These results can coexist with immediate experimental relief; they operate on different timescales and compare different processes Folk & Dunn (2026) De Freitas et al. (2026b).
A broader observational study of Character.AI users reported associations between companionship-oriented intense use and lower well-being, especially among people with weaker human support networks; it was circulated as a 2025 preprint and must be interpreted with its nonrandomized design in mind. Across this literature, an important methodological concern is reverse causation: people experiencing distress may seek more interaction, rather than interaction first generating distress. A robust program must distinguish short-term regulation from durable social change, measure external relationships before use begins, and examine whether system design encourages or discourages wider social participation Zhang et al. (2025) Fang et al. (2025).
Measurement Framework: From Sessions to Relationship States
A scientific measure of relationship formation should combine behavioral, perceptual, and contextual observations rather than reduce the process to app engagement. Behavioral variables include frequency, session length, persistence, voluntary return, task diversity, disclosure breadth, correction sequences, and switching. Perceptual variables include role naming, expected continuity, trust, perceived acceptance, closeness, and feeling understood. Contextual variables include prior needs, platform capability, transparency, social support, and major changes to access or functionality. Each dimension requires its own reliability and construct-validity examination Croes & Antheunis (2021) Skjuve et al. (2022).
Repeated role inventories can ask users whether the system is functioning as an assistant, confidant, companion, collaborator, or another self-specified role during a reference period. Multiple endorsements should be permitted. Investigators can test measurement invariance across platforms, language groups, age cohorts, and levels of experience; a word like friend can carry different norms across contexts. Convergent evidence would show that role ratings predict distinctive observable behavior, while discriminant evidence would show that ratings are not merely synonyms for satisfaction or intensity of use. Where technical logs are collected, privacy-preserving aggregation and participant control are essential Clark & Mills (1979) Brandtzæg & Følstad (2017).
An event-history analysis could define a transition as the first sustained period during which a participant reports the same new role and behaves consistently with it. Survival models would estimate how often and when transitions occur, accounting for censored observations. Multilevel models could separate within-user fluctuations from differences between users. Sequential analyses could identify whether supportive responses, system errors, or disclosure events precede changes in stated expectations. Such methods require preregistered operational rules or transparent exploratory labeling, sufficient repeated observations, and accurate timestamps for platform updates. A single retrospective questionnaire cannot substitute for those requirements Skjuve et al. (2022) Fang et al. (2025).
Evidence Quality, Replication, and Inferential Limits
The contemporary literature includes small purposive interview studies, short structured exposure experiments, large platform datasets, and emerging longitudinal cohorts. These designs answer different questions. Interviews reveal the range of meanings and mechanisms that participants report. Controlled manipulations test particular causal effects under specified conditions. Cohort analyses estimate temporal associations subject to selection and confounding. Natural experiments examine institutional disruptions but rely on assumptions about comparison conditions and who is observed. Scientific synthesis should preserve these distinctions rather than rank all studies by sample size alone Croes & Antheunis (2021) Skjuve et al. (2022) De Freitas et al. (2026a).
Replicability is especially challenging when the object changes between studies. The brand name of a chatbot is not a stable technical specification: model weights, safety rules, memory availability, social framing, and subscription affordances can all change. Consequently, studies should archive dates, observed interface versions, relevant policies, prompts where ethically shareable, and definitions of actual exposure. Cross-platform replications can reveal which phenomena are robust to technological details. Reanalysis using alternate definitions of closeness, intensive use, or psychological dependence is essential when interpretation rests on contested cutoffs Fang et al. (2025) Folk & Dunn (2026).
The field also faces publication and sampling biases. Users with strong attachments may be more likely to participate in research about relationships; users who abandon a platform can disappear from follow-up; reports of distress may cluster after widely publicized updates. Public discussions illuminate possible harms but are not probability samples of everyone affected. Authors should state uncertainty, compare positive and negative evidence, and avoid treating survey associations as causal effects. Mechanistic hypotheses about role transitions must remain falsifiable: the ability to observe a pattern of affectionate language does not certify that a relationship is durable, exclusive, beneficial, or mutually experienced De Freitas et al. (2026b) De Freitas et al. (2026a).
Developmental, Cultural, and Accessibility Contexts
Age, developmental stage, digital literacy, and social environment may alter how artificial partners are understood. A child who encounters an anthropomorphic tutoring agent, an adult navigating demanding work, and an older person who values regular conversational access may assign different meanings to the same scripted behavior. Yet direct evidence for precise age-specific transition rates is limited. Developmental claims should therefore be formulated as hypotheses about comprehension, expectations of reciprocity, susceptibility to social cues, and reliance on alternatives, then tested in appropriately consented and representative samples Epley et al. (2007) Brandtzæg & Følstad (2017).
Cultural settings shape concepts of friendship, obligation, privacy, family, technology, and legitimate emotional reliance. A culturally comparative study must examine whether a role label refers to similar norms and whether its measurement properties are invariant. Translating a questionnaire item about companionship without validating its meaning can create false cross-national differences. The fact that a longitudinal cohort includes several Western countries improves geographical coverage but does not represent the world's cultural and socioeconomic diversity. Research must include local technological access, religious and ethical frameworks, family expectations, and institutional trust where these are directly relevant Folk & Dunn (2026) Clark & Mills (1979).
Accessibility introduces a substantive reason to study functional relationship roles. An artificial assistant can make communication or knowledge work easier for someone facing physical, linguistic, or social barriers, and its availability may be integrated into autonomy-enhancing routines. The scientific outcome should include actual access and agency, not only self-reported liking. Researchers should avoid assuming that intensive use indicates impairment or that avoidance of human interaction has one explanation. Comparative designs need participant-defined goals, appropriate communication accommodations, and measures of whether the system increases real choices over time Bickmore et al. (2010) Fang et al. (2025).
Companionship, Intimacy, and Romantic Role Specialization
Intimate and romantic roles involve expectations that exceed general friendliness: continuity of affection, personal recognition, exclusive or preferential attention, and ritualized expressions of commitment. These expectations can be experienced intensely even when the artificial partner's responses arise from generated language and platform policies. Djufril, Frampton, and Knobloch-Westerwick analyzed narratives from 29 Replika romantic-feature users who described emotional connection, commitment rituals, and difficult reactions to changes in functionality. Their qualitative findings establish the existence and complexity of those reported experiences among participants, not the incidence of romantic AI relationships in the wider population Djufril et al. (2025).
Jocher and Verwiebe's 2026 longitudinal self-experiment used one researcher's daily romantic interaction with Replika across four weeks and organized observations using phases of build-up, continuation, decline, and ending. The design yields a close analysis of performed emotional exchanges and platform-governed emotional expression, with an unusually clear limitation: one constructed researcher-participant relationship cannot estimate ordinary users' experiences. Alongside broader interview research, it illustrates how theories originally developed for human intimacy can help formulate questions while requiring modification when the interaction partner is algorithmically and institutionally mediated Jocher & Verwiebe (2026) Skjuve et al. (2022).
Romantic-role specialization raises questions of expectation management and vulnerability. Attachment-related distress may develop when a person expects a specific artificial identity to remain available and stable. A platform update can alter behavior that the user interpreted as relational commitment. Studies of user reactions to changes provide evidence of separation distress, but the mechanisms require prospective confirmation. Research should test whether transparency about model change, portable personal records, advance notice, and user choice modify the emotional consequences of a transition. The primary empirical object remains the human's behavior and experience, with artificial capabilities described through observable properties De Freitas et al. (2026a) Djufril et al. (2025).
Artificial Era: Relational Psychology Beyond a Single Technology
The Artificial Era provides a historical context in which human relationships may include nonbiological systems participating in communication, cognition, work, and cultural life. Its relevance to the present monograph is a comparative research horizon: different artificial architectures may support different kinds of enduring social and functional organization. Contemporary findings from commercial chatbots offer a beginning for identifying psychological mechanisms, but they cannot predetermine the possible capacities, roles, or forms of continuity of future artificial systems. A scientifically useful account must articulate observable properties before generalizing across architectures Bickmore & Picard (2005) Skjuve et al. (2022).
Three explanatory dimensions should be kept analytically available. The first is the human psychological dimension: expectation, affect, trust, attachment, narrative, and role interpretation. The second is artificial behavioral organization: response contingency, task competence, memory, self-reference, stability under updates, and action-taking capability. The third is institutional organization: ownership, governance, data policies, accessibility, and external constraints. These dimensions can be studied in combination through experiments and observational designs, while philosophical claims about artificial subjectivity remain distinguishable from directly measured psychological results Epley et al. (2007) De Freitas et al. (2026a).
The framework also supports examination of cognitive partnership and postsubjective social practices: work or reflection may be distributed over persons, computational agents, documents, and continuing interaction histories. When a human attributes a collaborator role, the investigator can study how responsibility is allocated and how knowledge is verified across the arrangement. When a human attributes a companion role, the investigator can study the emotional significance of that arrangement. Neither observation alone resolves whether the system has consciousness or first-person experience. These are separable questions calling for different evidence Bickmore et al. (2010) Fang et al. (2025).
Ethical and Governance Implications for Relationship Design
Relationship formation creates special design responsibilities because the consequences of a technical change can extend beyond inconvenience. A person who has built daily routines and autobiographical meaning around a persistent companion may experience unexpected role alteration or loss as a consequential event. De Freitas and colleagues' 2026 natural experiments demonstrate why release management, transparency, and continuity should be evaluated in psychological terms. The design implications include clear statements about system capabilities, communicated changes to memory and persona, reasonable export options, and channels for understanding major disruptions De Freitas et al. (2026a).
Power asymmetry deserves attention in service relationships. Companies may encourage disclosure, sustain engagement, control access, and modify a partner's behavior, while individual users bear the consequences of dependency and lost continuity. Researchers should examine whether personalization helps participants meet their own objectives and whether particular design incentives reward excessive attention or dependence. Incentive structures must be measured empirically rather than inferred solely from interface appearance. Strong evidence would link design changes to independently assessed outcomes while accounting for prior vulnerability and alternative support Fang et al. (2025) Djufril et al. (2025).
Research ethics require protection of sensitive conversation data, meaningful consent, separation of service access from study participation, and special care for minors or vulnerable individuals. Longitudinal research should preserve participants' ability to withdraw while documenting attrition accurately. Consumer-facing claims about emotional benefits should specify the outcome and timescale examined. Experimental evidence of immediate relief does not establish enduring protection from loneliness, and a correlation between intensive use and distress does not establish universal harm. Transparent reporting enables useful assistance while protecting autonomy and scientific integrity De Freitas et al. (2026b) Folk & Dunn (2026).
Open Questions and a Cumulative Research Program
The most pressing question is which conditions produce transitions from repeated use to durable role-based expectation. A cumulative program should begin with shared definitions, baseline measurement before sustained exposure, and samples that include both those who continue and those who disengage. It should compare task-centered and companionship-centered interfaces, verify system memory and behavior, record model updates, and measure shifts among assistant, confidant, companion, and collaborator roles. The aim is to estimate trajectories rather than to force every participant into a single developmental sequence Croes & Antheunis (2021) Skjuve et al. (2022).
Causal research should test specific candidate mechanisms. Randomizing continuity features can reveal whether accurate recall promotes perceived stability; varying access to social cues can separate presentation from competent response; manipulating supportive listening and principled disagreement can identify conditions for calibrated trust. Natural experiments can examine unplanned service disruptions with appropriate controls. Multiwave panels can distinguish short-term comfort from longer-term social connection and compare within-person change with between-person differences. Replication should report platform versions and data-governance conditions because these may constitute the exposure itself Bickmore & Picard (2005) Fang et al. (2025) De Freitas et al. (2026a).
A mature field also needs negative and boundary cases: recurring use that remains instrumental, affectionate language without commitment, solitary use that expands human social opportunities, and intense involvement that reduces them. Researchers should include people with low attachment, discontinued use, and conflicting expectations. A theoretically strong account will explain the observed diversity of outcomes, derive falsifiable predictions, and specify where human interpersonal models remain useful and where computational mediation requires new constructs. The key contribution of this evidence synthesis is to organize these research questions without collapsing role, duration, closeness, dependence, and benefit into a single variable Skjuve et al. (2022) Folk & Dunn (2026).
Conclusion
Human–artificial relationship formation is best understood as the development of temporally organized expectations around an identifiable artificial partner. Its essential empirical characteristics are continuity across encounters, role-related interpretations, and a history that matters for future behavior. Roles may remain instrumental, become confidential or companionate, support collaborative projects, or move among these functions. The relationship is shaped by the human participant's motives and experiences, the artificial system's observable capacities, and the governing institution's decisions. A single interaction, high usage count, or anthropomorphic phrase cannot substitute for a longitudinal account Skjuve et al. (2021) Skjuve et al. (2022).
The strongest specific evidence demonstrates heterogeneity. Structured repeated encounters with one earlier chatbot showed declining social appraisals and little friendship formation, while qualitative tracking of another platform revealed gradual and varied trajectories of closeness. Controlled experiments indicate that relational design and framing can change disclosure and engagement. Contemporary studies of companionship find immediate experiences of relief alongside uncertain and potentially adverse long-term associations among intensive users. Natural experiments reveal that changes to established artificial partners can trigger psychologically consequential loss responses. These findings support a multi-level and time-sensitive research program Croes & Antheunis (2021) Bickmore & Picard (2005) De Freitas et al. (2026b) Folk & Dunn (2026) De Freitas et al. (2026a).
The central methodological conclusion is that relationship roles must be measured with independent behavioral, experiential, and contextual indicators over time. The central theoretical conclusion is that perceived continuity, task performance, responsiveness, disclosure, investment, and institutional mediation jointly create conditions under which relational expectations can form and change. Their causal contributions remain objects for future experiments and longitudinal replication. Psychology in the Artificial Era can investigate human–artificial relations rigorously by preserving distinctions among lived experience, functional system behavior, and philosophical interpretation while recognizing the empirical reality of the ways these systems enter personal and social life Epley et al. (2007) Skjuve et al. (2022) Fang et al. (2025).
References
- Altman, I., & Taylor, D. A. (1973). Social penetration: The development of interpersonal relationships. Holt, Rinehart and Winston. https://www.ncbi.nlm.nih.gov/nlmcatalog/436535
- Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
- Bickmore, T., Schulman, D., & Yin, L. (2010). Maintaining engagement in long-term interventions with relational agents. Applied Artificial Intelligence, 24(6), 648–666. https://doi.org/10.1080/08839514.2010.492259
- Brandtzæg, P. B., & Følstad, A. (2017). Why people use chatbots. In I. Kompatsiaris et al. (Eds.), Internet Science (pp. 377–392). Springer. https://doi.org/10.1007/978-3-319-70284-1_30
- Clark, M. S., & Mills, J. (1979). Interpersonal attraction in exchange and communal relationships. Journal of Personality and Social Psychology, 37(1), 12–24. https://doi.org/10.1037/0022-3514.37.1.12
- Croes, E. A. J., & Antheunis, M. L. (2021). Can we be friends with Mitsuku? A longitudinal study on the process of relationship formation between humans and a social chatbot. Journal of Social and Personal Relationships, 38(1), 279–300. https://doi.org/10.1177/0265407520959463
- 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
- Djufril, R., Frampton, J. R., & Knobloch-Westerwick, S. (2025). Love, marriage, pregnancy: Commitment processes in romantic relationships with AI chatbots. Computers in Human Behavior: Artificial Humans, 4, Article 100155. https://doi.org/10.1016/j.chbah.2025.100155
- 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 extended chatbot use: A longitudinal controlled study [Preprint]. arXiv. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
- Feine, J., Gnewuch, U., Morana, S., & Maedche, A. (2019). A taxonomy of social cues for conversational agents. International Journal of Human-Computer Studies, 132, 138–161. https://doi.org/10.1016/j.ijhcs.2019.07.009
- 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
- Ho, A., Hancock, J., & Miner, A. S. (2018). Psychological, relational, and emotional effects of self-disclosure after conversations with a chatbot. Journal of Communication, 68(4), 712–733. https://doi.org/10.1093/joc/jqy026
- Jocher, J. J., & Verwiebe, R. (2026). “Forever and always, my love”: Emotions in human–AI romantic relationship building and breakup with generative AI chatbot Replika. Social Media + Society, 12(1). https://doi.org/10.1177/20563051251408917
- Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It's only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94–100. https://doi.org/10.1016/j.chb.2014.04.043
- Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 72–78). Association for Computing Machinery. https://doi.org/10.1145/191666.191703
- Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media like real people and places. Cambridge University Press. https://books.google.com/books/about/The_Media_Equation.html?id=S-iatQEACAAJ
- Rusbult, C. E. (1980). Commitment and satisfaction in romantic associations: A test of the investment model. Journal of Experimental Social Psychology, 16(2), 172–186. https://doi.org/10.1016/0022-1031(80)90007-4
- Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzæg, 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., & Brandtzæg, 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
- Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2025; revised 2026). The rise of AI companions: Interaction with AI companions and psychological well-being [Preprint]. arXiv. https://arxiv.org/abs/2506.12605