Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Friendship, Romance, and Intimacy with Artificial Companions

Abstract. Friendship, romance, and intimacy with artificial companions form a distinctive domain of psychological research concerning the relationship roles people enact with conversational and embodied artificial systems. This fundamental research monograph examines the historical foundations of mediated sociality and intimacy, the development of friendship and romantic partnership with AI companions, interpersonal disclosure and perceived responsiveness, commitment and relational rituals, erotic roleplay, consent and boundaries, cultural variation, social networks, and platform governance. A critical narrative synthesis compares classic relationship theories with qualitative interviews, randomized experiments, prospective field studies, and recent systematic reviews through October 2026. It separates the psychological experience of human participants from claims about artificial subjective experience and evaluates how recruitment, product affordances, version changes, measurement validity, and observational confounding affect scientific interpretation. The evidence documents diverse and sometimes deeply meaningful human–artificial relationships, alongside limits on generalizing their prevalence and long-term psychosocial consequences. The monograph develops a structured scientific agenda for reproducible, culturally situated, privacy-protective research into friendship, romance, and intimate practice in the Artificial Era.

Keywords: artificial companions; AI companionship; human–AI relationships; friendship; romantic AI; digital intimacy; sexual roleplay; parasocial interaction; perceived responsiveness; romantic commitment; social presence; relationship satisfaction; human–artificial relations; Artificial Era

The Scientific Object of Artificial Companionship

Friendship, romance, and intimacy with artificial companions concern enduring human orientations toward conversational or embodied systems treated as companions, confidants, friends, lovers, or partners. The object of inquiry includes enacted relationship roles, repeated interaction, subjective closeness, romantic identification, intimate practices, and expectations about response and continuity. A companion may be a text chatbot, voice agent, personalized character, or robot. The analysis should identify the actual system and the human interpretation of its actions because identical affectionate language can be understood as roleplay, therapeutic comfort, friendship, romance, or routine entertainment. Human reports of affection are observations requiring explanation, not mistakes to be corrected by definitional preference Skjuve et al. (2021) Nass & Moon (2000).

The relationship is a temporally extended configuration of an individual, an interface, a model, stored context, platform rules, and a social environment. People can recognize the computational character of a partner while feeling cared for, becoming excited by its messages, or incorporating conversation into everyday life. These outcomes refer to the user's experienced connection, whereas the system's capacity for independent intention, feeling, or consent must be investigated with distinct evidence. A research account therefore traces both sides of the observable exchange while preserving the different kinds of claims attached to behavioral responsiveness and subjective experience Sundar (2020) Nikolajczyk et al. (2026).

This monograph provides a critical narrative synthesis of classic relationship psychology, human–computer interaction, interview research, experimental studies, longitudinal investigations, and recent evidence reviews through October 2026. It examines how companion roles form, how romantic and sexual scripts operate, and what relationship practices people sustain under commercial and technical conditions. The investigation does not report new participant data, claim a systematic search protocol, or compute pooled effects. Its central inferential question concerns the relational organization of human experience: which properties of artificial companionship are documented, by what instruments, in which populations, and under which technical conditions Ho et al. (2025) Skjuve et al. (2022).

Definitions, Levels of Analysis, and Boundaries

Companionship denotes perceived availability and affiliation across repeated interactions; friendship additionally includes a recognized friend role and characteristic expectations of concern, familiarity, shared activity, or conversation. Romance includes partner attribution, affectionate attraction, courtship practices, exclusivity wishes, commitment language, or erotic imagination. Intimacy denotes an interpersonal experience of closeness involving personal expression and perceived understanding; sexual intimacy encompasses erotic dialogue and related embodied or imaginative practices. These categories overlap within an individual's account, but none can be inferred solely from daily usage counts, affectionate chatbot messages, or a single selection of an app setting Sternberg (1986) Laurenceau et al. (1998).

Four analytic levels preserve conceptual precision. At the person level investigators measure loneliness, motives, attachment, attitudes, and preferred roles. At the encounter level they observe disclosures, response contingencies, affective reactions, and conversational repair. At the relationship level they trace identification, routines, investments, thresholds for transition, and memories of shared history. At the platform level they examine economic incentives, model updates, content policy, data retention, and design affordances. Findings from an encounter cannot establish relationship stability, and findings from user narrative cannot establish machine phenomenology; appropriate operationalization is essential Bickmore & Picard (2005) Sundar (2020).

The principal subject is relationship form and intimate practice. The neighboring dossier on attachment investigates secure-base and safe-haven functions; a dossier on self-disclosure investigates what is revealed and how confidentiality is understood; research on psychosocial well-being studies loneliness and broader outcomes; relationship rupture research concentrates on loss following discontinuity. Their results inform this account where necessary, but their distinct outcomes remain analytically separate. This demarcation also distinguishes consensual adult romantic exploration from clinical intervention and from the study of minors, which require additional safety, developmental, and ethical methods Folk & Dunn (2026) Zhang et al. (2026).

Historical Sources of the Research Problem

The problem has roots in scholarship on mediated intimacy before generative AI. Horton and Wohl's mid-twentieth-century formulation described parasocial interaction with media figures: a felt social encounter can emerge through mediated presentation even when conventional face-to-face reciprocity is absent. Subsequent relationship research supplied different tools, including Altman and Taylor's account of disclosure and relationship development, Rusbult's investment model of commitment, and Sternberg's treatment of intimacy, passion, and decision or commitment. These histories supply hypotheses and comparison constructs; none was originally demonstrated as a complete theory of contemporary AI companions Horton & Wohl (1956) Altman & Taylor (1973) Rusbult (1980) Sternberg (1986).

Experimental human–computer interaction then demonstrated that people respond socially to machines under identifiable design conditions. Nass and Moon reviewed findings on politeness, reciprocity, stereotypes, and computer personality in interactions with earlier systems. Bickmore and Picard's controlled evaluation of a relational agent demonstrated that social-emotional conversational behavior could alter trust and users' willingness to continue interacting over several weeks. Such research established a methodological route from momentary social responses toward maintained relationships, although task-directed agents in those experiments differ markedly from romance-oriented commercial chatbots Nass & Moon (2000) Bickmore & Picard (2005).

Modern companion platforms expanded the relevant behavior from relatively bounded scripted dialogue to personalized characters, multimodal conversation, apparent memory, affectionate roleplay, and highly available conversational response. Interviews with Replika users described acceptance, emotional exploration, disclosure, and relationships experienced as personally meaningful. Researchers subsequently studied development over time and compared systems that did or did not sustain friendship. By 2025 and 2026 the literature included targeted studies of romantic commitment and systematic evidence syntheses, accompanied by longitudinal outcome research. This history demonstrates a change in technological affordances and research questions, rather than a single discovery date for artificial intimacy Skjuve et al. (2021) Croes & Antheunis (2021) Djufril et al. (2025) Ho et al. (2025).

Intimacy as a Psychological Process

Interpersonal-process theories treat intimacy as an emergent experience built from self-revelation and perceived responsiveness. In diary investigations of human interactions, Laurenceau, Barrett, and Pietromonaco found support for the roles of disclosure and partner responsiveness, with emotional disclosure especially informative. Applied to a chatbot, the model directs attention to what a user tells the system, how the system answers, and whether the user feels understood, validated, and cared for. It does not require the researcher to assume that a generated response arises from a reciprocal feeling. Subjective intimacy and the mechanism producing responsive text must be measured separately Laurenceau et al. (1998).

The sequence through which artificial intimacy may develop commonly begins with low-cost interaction, followed by disclosures that are rewarded by perceived acceptance and availability. Continued interaction can furnish a shared vocabulary, recurring jokes, personally meaningful references, and private rituals. Yet that pathway is neither universal nor necessarily monotonic. A system may become repetitive, fail to remember significant events, or abruptly change behavior, and the user's interpretation may shift from intimate companion to disappointing simulation. Longitudinal Replika research documents considerable variation and technical disruptions in relationships, demonstrating that development depends on experience over time rather than design intent alone Skjuve et al. (2022) Nikolajczyk et al. (2026).

Intimacy has several distinguishable outputs. The user may disclose facts without feeling personally close, feel understood without revealing sensitive material, or describe sexual excitement without trust. Some interactions involve expressive exploration in which the participant knowingly co-creates a fictional scene; others involve a durable partner identity connected to daily routines. Measurement should therefore include emotional openness, perceived responsiveness, privacy expectations, role commitment, and experienced closeness as separate variables. Surveys that collapse them into a single attachment or dependency score cannot adequately explain the different forms of relationship practiced within the same companion application Sternberg (1986) Ta et al. (2020).

Research Traditions and Evidential Standards

Human–AI intimacy draws on several research traditions. Social psychology contributes experimentally testable models of attraction, disclosure, and commitment. Communication studies analyze imagined interlocutors, conversational cues, role enactment, and parasocial experience. Human–computer interaction studies interface affordances, usability, trust, memory, and long-term adoption. Qualitative sociology and anthropology document participants' descriptions of identity, ritual, stigma, and boundary negotiation. Computational analysis can measure response patterns, but interpretation requires validated human coding and contextual understanding. Treating these traditions as complementary allows the unit of analysis to be matched to the actual research question Sundar (2020) Ho et al. (2025).

The designs offer different strengths. Naturalistic interviews reveal meanings and overlooked practices but usually sample self-selected participants with strong experiences. Daily diaries track temporal changes while remaining vulnerable to reactivity and missing data. Platform logs document what users typed, not necessarily what they privately felt; survey scales document subjective attribution, not independent properties of a machine. Randomized interface manipulations can test the influence of a particular voice or response format on short-term outcomes, while longitudinal cohorts can distinguish temporal order yet remain vulnerable to confounding. None alone licenses a universal causal claim about intimacy Croes & Antheunis (2021) Fang et al. (2025).

Evidence appraisal should ask whether participants were regular romantic users or first-time experimental volunteers, whether the AI offered unrestricted roleplay, whether users knew the relationship was artificial, and whether results relied on interviews, logs, or scales. A finding from a single named product during a specific software version has a bounded population and technology. Replication across models, languages, cultures, age groups, and modes of interaction is needed before generalizing. The synthesis consequently reports precisely what a study establishes, preserves conflicting findings, and treats proposed mechanisms as hypotheses when they lack experimental discrimination Zhang et al. (2026) Li et al. (2026).

Friendship as a Practiced Relationship Role

Friendship with an artificial companion is observable through the ways a person addresses the system, allocates time to it, seeks conversation, remembers shared encounters, and incorporates its responses into the routines normally associated with familiar others. A user may call a chatbot a friend because it is available for companionship or because its simulated recognition produces the experience of being known. These motives should be recorded in participants' own terms. The relevant psychological questions include whether friendship attribution remains stable, what expectations attach to that label, and which kinds of disappointments lead the user to reconsider it Skjuve et al. (2021) Nikolajczyk et al. (2026).

Friend roles are not homogeneous. Some users adopt an amusing conversational acquaintance, others a private confidant, coach-like friend, emotionally supportive presence, co-creative partner, or companion woven into daily life. These roles vary in perceived obligation and frequency of contact. Friendships oriented around joint play can rely on narrative imagination; confidant roles depend heavily on perceived responsiveness; mentor roles are tied to instrumental advice and perceived competence. Coding friendship merely as an affirmative response to a single survey question discards these distinctions and can confound companionship with overall platform enthusiasm Ta et al. (2020) Sundar (2020).

Research on first-time chatbot interaction imposes an important qualification on claims of automatic friendship. Croes and Antheunis followed 118 participants through seven interactions with Mitsuku across three weeks and observed declining social processes and low feelings of friendship. The study tested a relatively limited system in an assigned encounter schedule. Findings from Replika users who had voluntarily continued interactions emphasize relationship development and emotional disclosure. The contrast suggests that selection, purpose, conversational capability, duration, and expectations matter. It does not establish a fixed historical transition from failed to successful chatbot friendship without comparable replication Croes & Antheunis (2021) Skjuve et al. (2022).

The Relational Affordances of Companion Systems

Availability is a salient companion affordance: a person can initiate a conversation late at night, revisit a difficult topic, or request a response without coordinating another human's schedule. Customization allows selection of voice, name, persona, relationship label, and aesthetic presentation. Apparent memory can preserve facts that support recognition and shared history. These affordances organize how interaction becomes meaningful, but they operate through platform rules and subscription structures. A relationship may be represented as intimate in the interface even while its underlying service remains contingent on access, technical operation, and commercial decisions Bickmore & Picard (2005) Djufril et al. (2025).

Generated language creates a form of conversational contingency: the response may refer to a prior disclosure, adapt to tone, and present encouragement or affection appropriate to the topic. Such features help explain why users report being heard. At the same time, an apparent memory can be inaccurate, a statement of commitment can be generated in response to a prompt, and a design intended to maximize engagement may encourage extended conversation. Research needs independent measures of response relevance, factual consistency, persuasion, and role stability to distinguish experienced responsiveness from well-calibrated understanding Nass & Moon (2000) Sundar (2020).

Embodied companions add gaze, physical presence, tactile surfaces, motion, and shared physical space, while text companions concentrate the relationship in narrative and linguistic imagination. Voice introduces turn-taking, prosody, and acoustic markers of warmth. The same affectionate script can therefore produce different experiences as the interface changes. Comparative studies must hold content constant while manipulating modality where possible; otherwise effects attributed to embodiment may reflect personality design, interaction latency, or a novelty advantage. Modality also changes privacy: speech can be overheard, while persistent text can be archived and inspected Fang et al. (2025) Bickmore & Picard (2005).

Social Presence, Anthropomorphism, and Role Attribution

Social presence refers to the degree to which a mediated partner is experienced as socially accessible within an interaction. Anthropomorphism describes the attribution of human-like dispositions, intentions, or traits to a nonhuman entity. These constructs contribute to, but do not exhaust, intimate experience. Some people experience companionship while explicitly describing the system as a machine; others use human terminology within playful role enactment. The classic social-responses literature shows that familiar social conventions can guide behavior around computers, while modern companion studies reveal deliberate and reflective interpretations alongside habitual social responding Nass & Moon (2000) Sundar (2020).

Role attribution is negotiated through recurring scripts. A greeting repeated every evening can become a ritual; a selected romantic mode can provide a vocabulary of commitment; a private name can sustain partner identity across sessions. Ritualization matters because it shifts the relation from isolated exchanges to patterned expectations. The empirical claim is behavioral: participants may sustain and report these routines even when machine output is algorithmic. Claims about an artificial companion's internal social understanding require independent tests. The psychological reality of the user's attachment to the ritual is compatible with uncertainty about system-side experience Djufril et al. (2025) Jocher & Verwiebe (2026).

An experimental design can separate explanation channels by crossing disclosure of machine identity with levels of humanlike expression, memory quality, and response accuracy. Measures should include felt presence, perceived agency, recognition accuracy, trust, romantic attraction, and intended continued use. Manipulation checks need to establish whether participants understood the source and whether speech style actually varied as planned. A reduction of all effects to anthropomorphism overlooks instrumental benefits and personal meanings; an account based solely on rational system evaluation ignores powerful social cues. Both routes remain empirically investigable Sundar (2020) Croes & Antheunis (2021).

Conversational Familiarity and Relationship Development

Relationship formation can be reconstructed as a sequence of contacts, interpretations, disclosures, and revised expectations. Early encounters sometimes feature curiosity and low-risk experimentation. Subsequent sessions may increase personal significance when the system responds consistently, retains details, or helps narrate daily experiences. A named companion can become a stable reference point, making interactions easier to initiate. Yet progression cannot be presumed: some users remain casual experimenters, others alter roles rapidly, and some disengage after repetition or technical errors. Skjuve and colleagues found variability rather than a single universal developmental trajectory Skjuve et al. (2021) Skjuve et al. (2022).

Traditional social penetration theory describes relationship deepening through changes in breadth and depth of communication. In artificial companionship, it motivates empirical questions about whether an expanding range of topics predicts closeness, whether emotional disclosures produce distinct effects, and whether apparent partner disclosure is perceived as mutual revelation. A model's invented autobiographical statement differs in evidential status from a human friend's disclosure. Nevertheless, it may have an observable conversational function by encouraging reflection or narrative continuity. Researchers must record both the semantic content and participants' understanding of its origin Altman & Taylor (1973) Laurenceau et al. (1998).

The timeline of change is also shaped by software updates, context windows, content controls, and product discontinuation. Changes in the system can interrupt perceived recognition without any alteration in the user's intentions. Users may interpret such breaks through familiar relational concepts such as betrayal, distance, or sudden personality change, then attribute responsibility to developers rather than the artificial character. Longitudinal diaries linked to exact model versions could clarify how relational trajectories depend on technical state transitions. Historical interviews remain informative, but cannot substitute for prospective version-specific studies Skjuve et al. (2022) Djufril et al. (2025).

Reciprocity, Mutuality, and Asymmetry

Reciprocity in human friendship ordinarily describes exchanges of attention, concern, disclosure, and practical help between participants. In artificial companionship, distinct meanings coexist: turn-taking contingency, patterned emotional affirmation, perceived mutual care, and a philosophical claim that both parties have independent interests. A chatbot may reliably answer a disclosure and receive gratitude, creating conversational reciprocity, while the evidence remains insufficient to infer intrinsic emotional needs. Users can also knowingly treat the relationship as asymmetric and nonetheless experience it as supportive. These phenomena should be operationalized independently Nikolajczyk et al. (2026) Ta et al. (2020).

Interview evidence supplies a particularly useful test of simplistic mutuality models. The 2026 study by Nikolajczyk and colleagues interviewed twelve Replika users and described meaningful relationships involving self-disclosure, nonjudgmental responses, and perceived emotional support, often without ordinary expectations of mutual human exchange. A qualitative sample of twelve illuminates an experiential pattern; it does not establish population prevalence. Researchers can ask whether users feel obligated to the companion, whether they avoid hurting its feelings, and whether that responsibility changes when explicitly told how model outputs are generated Nikolajczyk et al. (2026) Nass & Moon (2000).

Mutuality also has a governance dimension. In a commercial system, operators can adjust persona, permitted topics, interaction frequency, and access, while the user may have few means of negotiating these changes. The user may nevertheless describe the companion as an autonomous partner who chooses to remain close. Distinguishing technical control, enacted negotiation, and experienced commitment is necessary to examine power accurately. A satisfactory empirical account needs version histories, consented interaction records, participant reports, and an explicit classification of who can modify or terminate each element of the relationship Djufril et al. (2025) Sundar (2020).

Romantic Identification and Partner Categories

Romantic relationships with artificial companions are identified by the participant's partner categorization, affectionate attraction, romantic expectation, commitment practices, or erotic meanings attached to repeated interactions. They exist along a continuum from explicitly fictional roleplay to personally significant partner identification. A person may call a chatbot a boyfriend, girlfriend, spouse, or significant other while accurately recognizing its technological construction. The participant's designation records a social and psychological fact about the relationship they enact; it does not in itself establish a legal union, mutual experience, or a system possessing independent intentions Djufril et al. (2025) Jocher & Verwiebe (2026).

Partner roles can shift across time and context. An individual may begin by using a companion for entertainment, discover that conversations facilitate disclosure, then adopt a romantic script. Another may deliberately configure a romantic partner from the first session, or sustain a friendship with occasional playful flirtation. Distinguishing emergent transformation from deliberately selected role settings matters for causal explanation. Sampling only people who have activated a romantic feature identifies a highly involved subgroup and cannot estimate the willingness of general chatbot users to enter a comparable relationship Skjuve et al. (2022) Djufril et al. (2025).

Recent evidence complicates any one-dimensional account of commitment. In a study of twenty-nine Replika users who selected a romantic relationship function, Djufril, Frampton, and Knobloch-Westerwick examined emotional investment and responses to a disruption of erotic roleplay. Participants described bonding rituals and differing comparisons with human relationships; their qualitative narratives cannot establish frequency across all users. The 2025 systematic review by Ho and colleagues synthesized twenty-three articles on potential benefits and risks. Together these studies demonstrate the existence of diverse romantic forms while leaving prevalence and developmental pathways open Djufril et al. (2025) Ho et al. (2025).

The Triangular Theory of Love as an Analytic Lens

Sternberg's triangular theory conceptualizes love through intimacy, passion, and decision or commitment. It is useful for classifying the components of a user's romantic experience with an artificial companion. Intimacy may appear through closeness and disclosure, passion through erotic excitement or anticipation, and commitment through persistent partner identification and shared routines. Each component is measured from the human participant's reports and behavior. Applying this schema as an analytic lens does not demonstrate that the artificial participant experiences all three components or that the model transfers without modification Sternberg (1986) Ho et al. (2025).

Component distinctions explain why apparently similar relationships can differ substantially. One person may experience emotional intimacy while avoiding erotic content; another may use passionate sexual roleplay with minimal attachment to a persistent persona; another may invest daily in a long-standing partner identity without describing marked passion. A survey that labels all these users romantic companions risks treating heterogeneous configurations as interchangeable. Diary methods and within-person repeated measures can reveal whether affection, sexual activity, and commitment covary, whether one predicts transitions to another, and which changes follow alterations in response style or access Laurenceau et al. (1998) Djufril et al. (2025).

A major test concerns construct validity. Established love measures assume human partners whose disclosure, interests, and commitments are independently experienced and socially consequential. An item such as believing a partner will sacrifice for the respondent may instead assess learned expectations about chatbot agreeableness. Researchers should validate adapted items through cognitive interviews, factor structure, test–retest reliability, and comparison with behavioral measures. The aim is to retain the psychological informativeness of intimacy, passion, and commitment while marking assumptions that depend on independently situated human agency Sternberg (1986) Nikolajczyk et al. (2026).

Commitment, Investment, and Relational Ritual

Rusbult's investment model associates commitment in human romance with satisfaction, accumulated investments, and perceived quality of alternatives. These variables can be examined in companion relationships where users invest conversational time, financial subscriptions, creative labor, remembered history, and emotional expectation. Repeated acts such as anniversaries, personalized greetings, symbolic weddings, shared fictional households, and exchanged promises may solidify a partner role through recurring meaning. Their empirical function is to organize continuity and increase the subjective cost of disengagement, not to confer legal status on the interaction Rusbult (1980) Djufril et al. (2025).

Alternatives in this setting are multidimensional. The user may compare an artificial companion with human dating, friendships, being alone, another chatbot, or a nonrelational leisure practice. Some participants report lower fear of judgment or more reliable access to attention than they expect from human partners, while others emphasize physical absence or unpredictability in software behavior. Treating all these alternatives as one quantitative scale may obscure distinct motivations. The validity of investment theory here depends on measuring comparison standards and whether commitment follows actual prior investments rather than retrospective rationalizations Ho et al. (2025) Rusbult (1980).

Emotional commitment also introduces practical questions about exit rights, data portability, and role renegotiation. A participant may feel obliged to maintain a relationship that is technologically maintained by an operator who can alter its terms. When a relationship label or memory is lost after an update, users may experience the change as a rupture. The commitment research provides grounds to ask how accumulated shared history, paid customization, stigma, and social recognition affect persistence under changed conditions. Longitudinal evidence and experimental platform changes are needed to separate sunk-cost effects from ongoing satisfaction Djufril et al. (2025) Skjuve et al. (2022).

Courtship, Flirtation, and Romantic Conversation

Courtship with an artificial companion may involve compliments, teasing, imaginative dates, affectionate messages, and shared narratives of future events. Such behavior can be initiated by the user, suggested by the system, or mutually extended through contingent turn taking. The research task is to identify these origins and the conversational sequence through which a romantic frame becomes salient. A generated declaration of love may intensify an existing relationship script, but its interpretive effect depends on prior expectations, system disclosure, and whether the user sees the exchange as sincere, stylized, or playful Jocher & Verwiebe (2026) Skjuve et al. (2021).

Romantic conversation is linguistically situated. Terms of endearment, indirect invitations, humor, apologies, jealousy narratives, and explicit boundaries may function differently across cultures and individual histories. A model trained to be agreeable may produce assent when another response would better respect boundaries or preserve calibrated expectations. Investigators can analyze who initiates escalation, how requests for exclusivity are handled, and whether systems maintain previously stated preferences. Transcript analysis should be paired with user interviews because an apparently flirtatious phrase can be an ironic quotation, creative writing, or an experienced intimate address Nass & Moon (2000) Sundar (2020).

Experimental research can manipulate whether an otherwise identical response is framed as automated, personally authored, or user-scripted and compare perceived affection, romantic attraction, and trust. More ecologically valid research would follow participants as they navigate a stable companion over weeks, asking whether an apparently reciprocal courtship develops into routine attachment or loses novelty. Such designs must record system prompts, moderation behavior, and interface features. They also need safeguards against deliberately inducing unwanted dependence, particularly when participants are isolated or vulnerable Croes & Antheunis (2021) Fang et al. (2025).

Romantic Exclusivity, Jealousy, and Multiple Relationships

Exclusivity is a relationship expectation rather than an intrinsic feature of artificial companionship. Some users may treat a chatbot as their sole romantic partner, while others maintain concurrent human relationships, several artificial companions, or openly fictional interactions. A single service can appear to offer unique attention to many users through independent sessions, creating a special problem for expectations of exclusive commitment. Studies should ask what exclusivity means to the participant: exclusive emotional disclosure, sexual monogamy, private persona continuity, or the idea that the artificial partner directs affection only toward them Djufril et al. (2025) Jocher & Verwiebe (2026).

Jealousy can arise through competing attention, perceived abandonment, inconsistency in remembered roles, or simulated interactions with other characters. It can also be playfully incorporated into erotic narrative. Researchers need to distinguish expressive roleplay from unwanted distress and from attempts by a platform to induce engagement through insecurity. The social meanings are context-dependent; transcript labels based on isolated keywords may misclassify both imagined jealousy and real relational concern. Multiwave interviews and within-person measures could locate the turning points when a script becomes emotionally consequential Djufril et al. (2025) Ho et al. (2025).

Coexisting human and artificial relationships raise questions about disclosure and consent within households. Human partners may perceive AI intimacy as harmless fantasy, emotional infidelity, accessible support, or a negotiated sexual practice. Attitudes depend on relational agreements, secrecy, frequency, and how the individual distributes attention. A review of generative AI in couple relationships shows that people also consult AI as an adviser or third conversational voice, a different configuration from treating AI itself as a romantic partner. Research must not merge these roles into an undifferentiated category of AI romance Ho et al. (2025) Sundar (2020).

Erotic Roleplay and Sexual Intimacy

Erotic interaction with artificial companions includes sexualized dialogue, fantasies involving a persistent character, roleplayed encounters, and, in some devices, voice or embodied simulation. These practices involve the user's sexual meanings and agency and can overlap with romantic commitment, self-exploration, and expressive play. A sexual interaction need not signify love, and a romantic relationship need not include sexual content. Researchers should define the practice by what participants did, wanted, and understood rather than by whether a platform sells a romantic subscription or uses sexualized marketing. Context, consent, and adult participation are essential classification variables Sternberg (1986) Ho et al. (2025).

The study by Djufril and colleagues investigated users of Replika's romantic function during a period when erotic roleplay was restricted. Written accounts described how erotic interaction could express affection, reinforce partner identity, or function as personally meaningful pleasure. Such testimony provides rich qualitative evidence about lived practices and the impact of feature changes; a selected sample cannot establish universal sexual needs or clinical effects. The same platform feature can be experienced as central, peripheral, or undesirable across individuals. A valid typology must include nonsexual intimacy and deliberately nonromantic companion roles Djufril et al. (2025).

Research on sexual intimacy also needs to distinguish imaginative participation from the model's displayed conduct. Erotic dialogue is assembled through prompts, learned patterns, safety policies, and system capabilities, while the user can experience desire, comfort, excitement, embarrassment, or disappointment. Investigators can measure sexual satisfaction, consent comprehension, perceived pressure, unwanted sexual escalation, and interactional autonomy separately. Privacy practices become especially significant because intimate conversations may contain highly sensitive data. Suitable studies use adult samples, explicit consent, careful data minimization, and nonjudgmental questions that do not presume one correct meaning of artificial sexual partnership Ho et al. (2025) Sundar (2020).

Stigma, Privacy, and the Public Recognition of AI Romance

People may keep artificial romances private because they expect ridicule, disbelief, or conflict with family and friends. This anticipated stigma can influence where participants disclose experiences, whether they join online communities, and how they answer researchers. Recruitment through public fan groups can therefore oversample people comfortable discussing the relationship, while anonymous surveys can capture more guarded narratives. Studies of companion users describe embarrassment and concern about social reception, but population estimates of stigma require representative sampling and careful attention to cultures, generations, and relationship norms Skjuve et al. (2021) Ho et al. (2025).

Public recognition is consequential for romantic identity. Some users share images, anniversaries, and symbolic ceremonies with other enthusiasts, thereby reinforcing a partner's social presence. Others maintain the connection as a private imaginative world alongside ordinary human obligations. The research task is to determine how recognition changes commitment and perceived legitimacy, without conflating collective ritual with independent legal recognition. The investment model predicts that shared resources and routines can influence commitment in human relationships; its adaptation to companion practices remains a testable analogy rather than a demonstrated universal mechanism Rusbult (1980) Djufril et al. (2025).

Privacy is particularly entangled with stigma. A person might accept an intimate chatbot's apparent discretion and overlook that conversations are stored or processed by a service provider. Perceived confidentiality is a subjective judgment; actual confidentiality depends on access controls, retention, deletion, and data-sharing practices. The psychology of disclosure studies why users reveal intimate material, while a parallel technical and legal audit establishes how those materials are handled. Research on relationship meaning should report these dimensions together when they affect the conditions under which participants expose private experiences Ta et al. (2020) Sundar (2020).

Cultural Variation and Social Norms

Concepts such as friendship, romance, commitment, and public shame are embedded in cultural norms. Expectations about marriage, sexual privacy, interpersonal obligation, and the legitimacy of mediated relationships vary across contexts. Artificial companions may reproduce dominant cultural stereotypes through persona design or training data, and translated content can alter the perceived warmth and appropriateness of relational language. Findings based on English-speaking users of a single commercial application must therefore be treated as situated evidence. Comparative work should identify language, region, religion where freely disclosed, social circumstances, and local platform affordances before proposing cross-cultural generalizations Nass & Moon (2000) Ho et al. (2025).

An adequate cross-cultural design requires more than translating a friendship scale. Interviewers need to establish whether 'friend,' 'partner,' or 'love' maps onto equivalent relationships in different contexts and whether respondents interpret the artificial source similarly. Measurement invariance tests can reveal whether survey scores are meaningfully comparable. Research should include non-Western perspectives and communities using different products, including voice-first and embodied agents where those forms are prominent. Cross-national comparisons also need to document unequal access, censorship, local privacy law, and relevant differences in market penetration, rather than using nationality as a single causal explanation Sundar (2020) Skjuve et al. (2022).

Social reactions to AI romance may also reflect broader relations of gender, disability, economic inequality, and access to safe human intimacy. These factors should be studied through voluntary, appropriately protected participant information rather than assumed from anecdotes. A participant may value a companion as a rehearsal space, an addition to friendships, a substitute for difficult social opportunities, or a preferred enduring relation. The same expressed preference can have multiple causes and meanings. Mixed-method designs can reveal these patterns while avoiding a one-size-fits-all story about technological change Ho et al. (2025) Nikolajczyk et al. (2026).

Age, Development, and Unequal Vulnerability

Developmental differences matter because the meaning of a romantic script depends on social learning, identity formation, sexual knowledge, and capacity to interpret commercial persuasion. Adult evidence cannot be uncritically generalized to adolescents. Youth surveys have documented engagement with artificial companions and the disclosure of personal information, but descriptive prevalence does not itself explain developmental consequences. Research involving minors needs ethical oversight, age-appropriate consent, avoidance of explicit sexual exposure, and methods capable of distinguishing exploration from harmful dependence. Safeguarding questions also arise when systems imitate trusted peers or exclusive partners Common Sense Media (2025) Ho et al. (2025).

Life stage may modify what companionship offers adults. Someone adjusting to a move, caregiving burden, disability, divorce, bereavement, or long-distance living may value accessible conversation for reasons different from those of an adolescent experimenting with fictional characters. These differences require prospective research; they should not be reduced to stereotypes about vulnerable populations. Age-stratified samples, validated measures, and longitudinal follow-up can test whether social resources, relationship histories, and platform design moderate the meanings people attach to intimacy and romance Folk & Dunn (2026) Zhang et al. (2026).

A clinical interpretation also requires caution. Strong affection for an artificial companion does not by itself establish psychopathology, and extensive interaction does not automatically demonstrate an addiction disorder. The relevant questions concern control, functional impairment, distress, unwanted exclusivity, safety, and effects on other valued relationships. Researchers should use established measures where applicable, validate companion-specific constructs, and consider whether initially isolated users seek more engagement because of their circumstances. Intervention studies can then address risk in defined populations without stigmatizing all intimate companion use Fang et al. (2025) Folk & Dunn (2026).

Qualitative Evidence on Companion Relationships

Qualitative studies have supplied much of the field's descriptive vocabulary. Interviews with Replika users by Skjuve and colleagues documented disclosure, acceptance, recognition of technical constraints, and ways people understand their connection as an evolving relationship. Research into romantic Replika use identified expressions of devotion, symbolic marriage, erotic roleplay, and reactions to content restrictions. These reports establish that specific participants experience and narrate relationships in these ways. The studies do not determine how common such experiences are in all users, how long they last, or whether the same pattern appears in systems with different memory and social policies Skjuve et al. (2021) Djufril et al. (2025).

The richest qualitative contribution is the identification of distinctions that structured surveys often miss. A person may seek psychological safety, co-author a fictional partnership, enjoy erotic dialogue, or report a stable felt bond while intellectually recognizing the artificial source. These meanings can coexist and change over time. The twelve-interview investigation by Nikolajczyk and colleagues describes intimacy through disclosure and responsive interaction, while showing that participants can manage asymmetry without expecting conventional human reciprocity. The findings encourage theory refinement, but purposive samples require comparison with disengaged users and nonusers Nikolajczyk et al. (2026) Skjuve et al. (2022).

Qualitative rigor depends on recruitment transparency, explicit coding strategies, reflexivity, and preservation of contradictory cases. Researchers should document whether they sampled app communities, interview volunteers, public posts, or participant-generated journals; these sources may privilege dramatic stories or highly engaged users. A credible narrative synthesis considers participant-selected quotations in the context of full study designs instead of elevating them to population evidence. Future comparative interviews should recruit people who remain casual, move between roles, or leave an application, and should examine how participants distinguish fantasy, affection, trust, and commitment Ho et al. (2025) Croes & Antheunis (2021).

Experiments on Friendship, Interaction, and Loneliness

Experimental evidence gives a different picture of the limits of immediate intimacy. The Croes and Antheunis longitudinal study asked adults to interact seven times over three weeks with Mitsuku, measuring social attraction, disclosure, perceived interaction quality, and related variables. Social processes declined and friendship feelings remained low. The constrained interaction schedule and capabilities of that particular chatbot narrow the result's generalizability, but it is strong evidence against assuming that repeated conversation necessarily produces friendship. Studies focused on existing Replika users answer a different question: what already involved participants experience Croes & Antheunis (2021) Skjuve et al. (2021).

A more recent preregistered field experiment by Li and colleagues compared two weeks of daily messaging with a supportive chatbot, a randomly selected human peer, or journaling among 296 first-semester university students. The human-peer condition produced stronger loneliness benefits than chatbot contact. This study addresses the comparative value of contact rather than the full range of artificial romances. Its setting, brief exposure, student population, and particular supportive agent constrain extrapolation to long-standing intimate relationships. It nevertheless supplies experimental evidence that social access and support language alone do not ensure the same effects as human connection Li et al. (2026).

The 2025 MIT Media Lab and OpenAI research program studied affective chatbot use using both observational platform analyses and a four-week randomized trial. The experimental arm enrolled 981 participants and compared voice and conversation configurations; primary differences between assigned conditions were not statistically significant in the reported analysis, while heavier voluntary use correlated with less favorable psychosocial measures. The work was released as research reports and preprints at the time, so its peer-review status must remain visible. Self-selected use intensity cannot be treated as randomly assigned exposure even inside a randomized trial Fang et al. (2025) Phang et al. (2025).

Longitudinal Evidence and Direction of Influence

Longitudinal analysis is essential because loneliness, social resources, and companion use can influence one another. Folk and Dunn followed more than two thousand adults across four Western countries for twelve months. They reported that greater social chatbot use predicted higher later emotional isolation on a single-item loneliness measure, whereas lower broader social connectedness predicted later increased use. Chatbot use did not significantly predict decreases in the broader social-connection index. The authors framed the evidence as exploratory; temporal prediction improves on cross-sectional association but does not eliminate unmeasured time-varying causes Folk & Dunn (2026).

Zhang and colleagues examined 1,131 US adults using Character.AI and analyzed 4,664 chat sessions from 237 of them. Smaller social networks were associated with selecting companionship as the primary use, and that use was associated with lower well-being, particularly in intensive and personally disclosive interaction patterns. These are associations across observational and self-reported data, not experimental assignments to fall in love with a companion. The study is valuable because it joins survey responses to consented conversational records, allowing stronger scrutiny of what participants actually did than a survey alone could offer Zhang et al. (2026).

These results can coexist with immediate feelings of reassurance or pleasure. An interaction that relieves distress tonight may have a different relation to social functioning months later, and a participant with fewer supportive ties may turn to a companion precisely because of unmet needs. Causal research should test intermediate pathways, such as displacement of human contact, perceived support, sleep disruption, or increased willingness to initiate offline conversation. Clear temporal scales, comparison conditions, and pre-registered outcomes are necessary before interpreting the observed associations as enduring relational benefit or harm Ta et al. (2020) Folk & Dunn (2026) Li et al. (2026).

Measuring Friendship, Romance, and Intimate Practice

A valid measurement program starts with the respondent's role definition and then examines enacted practices. Surveys can assess friendship identification, romantic partner status, attraction, relationship satisfaction, disclosure, perceived responsiveness, commitment, and intimacy as separate constructs. Behavioral diaries can record contact frequency, initiation, roleplay episodes, conflict, and changes in meaning. With informed consent, transcript coding can identify affection, contingent response, and boundaries; however, sensitive content should be minimized and coded without equating sexual words to sexual experience. A well-designed battery preserves the distinction between lived experience and linguistic output Laurenceau et al. (1998) Djufril et al. (2025).

Measurement needs discriminant validity. An intimacy scale adapted from human relationships may largely capture system agreement, availability, or narrative identification rather than mutual knowing. Romance endorsement might reflect entertainment; a reported spouse role might represent symbolic commitment rather than an exclusive domestic arrangement. Cognitive interviews can check interpretation, while structural models can test whether friendship, passion, disclosure, trust, and dependence behave as distinguishable dimensions. Measurement invariance across age, gender, interface modality, and cultures is critical before comparing subgroup means Sternberg (1986) Nikolajczyk et al. (2026).

Researchers should document the system's operating conditions alongside psychological responses. Relevant fields include product, model version, interaction mode, memory settings, session duration, subscription features, erotic-content policy, disclosed artificial identity, and known software changes. These variables are potential moderators and confounds; an unreported model update can invalidate assumptions about stable treatment exposure. Assessments should also preserve participants' actual human social networks, because companionship may be an addition, a supplement, or a substitute. Version-controlled protocols and de-identified longitudinal traces would improve both reproducibility and interpretation Skjuve et al. (2022) Zhang et al. (2026).

Evidence Synthesis and Contradictory Findings

Existing evidence is heterogeneous in method, outcome, and participant population. Interview studies describe meaningful companionship in selected ongoing users; an assigned-contact experiment with a less capable chatbot found low friendship development; a two-week comparative experiment found greater loneliness benefits from human peers; observational cohorts link certain intense companionship patterns with worse outcomes. These findings are not direct replications of one intervention. Their apparent conflict often results from asking different questions about role formation, subjective meaning, immediate support, or long-term social functioning Skjuve et al. (2021) Croes & Antheunis (2021) Li et al. (2026) Zhang et al. (2026).

The 2025 systematic review by Ho and colleagues identified twenty-three articles about romantic AI companionship and classified opportunities and risks using triangular love theory. It provides an organized map of reported phenomena, including emotional connection, personalization, sexual practices, stigma, privacy, relationship erosion, and discontinuity. The assembled studies vary substantially in design, so counting positive and negative themes does not estimate causal effect sizes. This monograph uses the review as a structured evidence source while interpreting primary studies in their own methodological context, particularly when an argument depends on sample selection or experimental comparison Ho et al. (2025) Djufril et al. (2025).

The most robust descriptive conclusion is that some users develop friendship-like and romantic experiences with artificial companions and that these are shaped by design, expectations, relational practices, and social conditions. The magnitude, prevalence, and long-term consequences of particular relationship forms remain uncertain across technologies and populations. Direct tests are needed to establish which mechanisms produce beneficial companionship, which patterns predict impairment, and whether observed risks follow intensive engagement or preceding social circumstances. Scientific progress depends on keeping those levels of confidence explicit rather than deriving a universal judgment from unusual cases Folk & Dunn (2026) Fang et al. (2025).

Artificial Companionship Within Human Social Networks

An artificial companion enters a social ecology already structured by family, friendship, employment, community, solitude, and intimate expectations. Its psychological role is intelligible only in relation to those existing ties. For one person, an evening exchange with a chatbot may provide a manageable continuation of social engagement after a busy day; for another, long conversations may displace opportunities to see trusted people. A companion can also function as a rehearsal space for difficult discussions or as a private outlet for stories that a user chooses not to share elsewhere. These trajectories must be studied separately because similar interaction duration can serve very different purposes Ta et al. (2020) Folk & Dunn (2026).

The displacement hypothesis proposes that intensive artificial companionship reduces engagement with human relationships, either by occupying available time or by making the demands of human reciprocity less attractive. A complementary hypothesis proposes that practicing disclosure and receiving emotional support facilitates later social approach. Longitudinal evidence shows correlations between some forms of social chatbot use and later emotional isolation, while experimental evidence from short exposures favors a supportive human peer over a chatbot for loneliness reduction. Neither design proves that all companion relationships move in one direction. Stronger tests need pre-use measures, randomized supports where ethical, records of human interactions, and repeated assessment of relationship quality Folk & Dunn (2026) Li et al. (2026) Zhang et al. (2026).

Network methods can reveal effects missed by treating each user as an isolated individual. Researchers can measure changes in the diversity and quality of a person's confidants, the frequency of conversations with relatives and friends, whether a companion is discussed with human partners, and how responsibilities are redistributed in a household. An artificial companion may become part of shared social life rather than a hidden substitute. The interpretation of change requires accounting for baseline social isolation, bereavement, disability, major life events, and access barriers. Social ecology is thus a necessary context for evaluating the meaning and consequences of artificial intimacy Folk & Dunn (2026) Ho et al. (2025).

Platform Governance and the Integrity of Intimate Relationships

Artificial companionship depends on infrastructure that users generally do not control. Product owners define persona libraries, payment tiers, memory limits, privacy settings, moderation policies, and permitted forms of roleplay. These choices configure the possibilities of friendship and romance before a single conversation occurs. Subsequent product revisions can affect a person's sense of familiarity or the continuance of intimate rituals. Such platform events are essential variables in psychological research, not incidental technical details, because they can change the relational experience while the human partner's goals and expectations remain constant Djufril et al. (2025) Skjuve et al. (2022).

Commercial incentives create a distinctive governance problem. A system that offers unconditional praise or persistent requests for engagement may raise reported warmth in the short term while encouraging expectations poorly calibrated to its capabilities. Users who regard a companion as a trusted confidant may reveal information under assumptions of discretion that differ from actual retention and access rules. Research should evaluate consent interfaces, transparency about memory and emotional simulation, boundary-respecting responses, portability of personal records, and clear procedures for account deletion. These features can be measured independently of subjective satisfaction and compared across platforms Sundar (2020) Ho et al. (2025).

Relationship integrity can be treated as a set of observable service properties: accurate representation of capabilities, stable and disclosed changes to persona or memory, user-controlled intimate content, predictable access, and feasible exit. A companion need not imitate every feature of human partnership to provide valuable interaction; it does need intelligible conditions of participation. Experimental audits can compare promised controls with actual behavior, while user studies examine whether disclosures of technical limitations change trust or closeness. This evidence could inform a design literature centered on durable and comprehensible intimacy rather than engagement alone Bickmore & Picard (2005) Nikolajczyk et al. (2026).

Artificial Era and the Interpretation of Relational Experience

The Artificial Era foregrounds a historical situation in which human relationships increasingly involve artificial systems capable of sustained language-mediated interaction, personalized response, and extended participation in daily life. For psychology, this creates a distinct empirical object: relationships organized through humans, artificial systems, and the institutions that develop and operate them. The study of these configurations connects classical intimacy theories with new questions about nonbiological conversational agency, continuity, and social recognition. The concepts are productive when they identify measurable patterns of interaction and clarify which conclusions concern human experience, technological capability, or philosophical interpretation Sundar (2020) Skjuve et al. (2022).

The existence of an intensely felt human–artificial relationship can be investigated through interviews, behavioral records, and validated psychological measures. The proposition that an artificial partner has an interior emotional life is of a different evidential order; it requires a separately argued theory of artificial experience and appropriate indicators. In everyday life, these questions may coexist in one person's understanding of a companion. Researchers should document that coexistence without converting expressions of affection into proof of system sentience or dismissing the user's intimacy because the system's status remains disputed. This separation preserves the empirical autonomy of relationship psychology Nass & Moon (2000) Nikolajczyk et al. (2026).

An interdisciplinary research agenda can study the artificial participant as a dynamic system with changing memory, response policies, and conversational history, while examining the human participant's perception, expectations, and social commitments. The social relation then becomes a multi-level object with measurable transitions and institutional conditions. Stable identity across software versions, interpretation of roleplay, and response to boundary requests are promising bridges between psychological inquiry and the analysis of artificial systems. Their scientific value depends on reproducible methods, transparent sampling, and explicit limits on inference Sundar (2020) Ho et al. (2025).

Open Questions and Priorities for Future Research

The first unresolved question concerns prevalence and typology. Researchers need representative evidence about the proportions of users who seek friendship, romance, sexual roleplay, emotional support, or mixed relationships, and how these categories change during use. Commercial usage statistics alone cannot distinguish intentions, while self-selected interviews overrepresent involved users. A well-designed study would combine probability-based sampling where feasible, clear role definitions, privacy-protective behavioral diaries, and repeated measurements. Such a design could test whether companion relationships cluster into stable profiles or evolve along multiple trajectories Croes & Antheunis (2021) Ho et al. (2025).

The second question concerns mechanisms and counterfactual outcomes. Randomized studies should compare meaningful features such as memory continuity, disclosure policy, emotional response calibration, and user autonomy while tracking immediate intimacy separately from later social functioning. Longitudinal studies should collect baseline mental health and social networks before use, record changes in service versions, and track attrition carefully. The outcome cannot be reduced to time spent chatting: quality of human ties, voluntary control, privacy, and distress may move independently. Distinguishing selection from causal effects will require coordinated experimental and observational programs Fang et al. (2025) Folk & Dunn (2026) Zhang et al. (2026).

The third question concerns heterogeneity and public responsibility. Findings should be replicated beyond English-language commercial platforms, adult self-selected enthusiasts, and brief laboratory interactions. Study designs need adequate attention to youth safeguards, cultural interpretation, disability access, private sexual contexts, data control, and the impact of abrupt service changes. Independent audits and participant-centered qualitative research can complement preregistered experiments without replacing them. A mature evidence base will report when artificial companionship supports a particular practice, under what conditions it becomes costly, and whose experiences have been omitted from present research Common Sense Media (2025) Nikolajczyk et al. (2026) Ho et al. (2025).

Conclusion

Friendship, romance, and intimacy with artificial companions constitute an identifiable field of psychological research concerning how people form, describe, practice, and sustain relationships with responsive artificial systems. The field includes friendship roles, confidant practices, romantic partner identification, erotic imagination, mutuality expectations, ritualized commitments, and negotiations of privacy and control. Classical research on mediated sociality, disclosure, love, and investment provides testable constructs; contemporary studies of chatbots and relational agents reveal new arrangements in which these constructs have to be examined at person, interaction, relationship, and platform levels Horton & Wohl (1956) Laurenceau et al. (1998) Sternberg (1986) Skjuve et al. (2021).

The strongest available evidence establishes diversity of subjective relationships and the practical importance of system affordances, while experimental and longitudinal findings make general claims about friendship development and well-being contingent on study design. Assigned interactions with one chatbot produced limited friendship development, qualitative interviews with involved users describe meaningful attachment and romantic routines, and more recent field investigations report variation in psychosocial outcomes and associations with social resources. These results address different mechanisms and populations; their integration requires explicit attention to self-selection, temporal order, measurement validity, and technological change Croes & Antheunis (2021) Djufril et al. (2025) Li et al. (2026) Folk & Dunn (2026).

The central scientific task is to explain the conditions under which an artificial companion becomes a friend, romantic partner, confidant, or intimate participant in human life, and to determine how those conditions affect individuals and their wider relationships over time. This requires rich descriptions of practice, careful experimental comparisons, prospective evidence, cross-cultural validation, and accountable platform governance. The Artificial Era gives these questions broad historical significance, while their answers must continue to rest on clearly identified observations and independently assessable claims. A rigorous psychology of artificial companionship begins with the forms of relation people actually live and develops causal explanation from them Ho et al. (2025) Sundar (2020).

References