Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Human–Artificial Sexual and Intimate Relationships

Abstract. Human–artificial sexual and intimate relationships are psychologically consequential arrangements in which individuals experience erotic attraction, romantic identification, intimate disclosure, or partner-oriented conduct involving conversational or embodied artificial systems. This fundamental research monograph defines the subject across sexual fantasy, erotic roleplay, perceived responsiveness, relationship commitment, embodiment, relational boundaries, and technological governance. It critically integrates sexual scripting theory, romantic investment and love models, intimacy and attachment research, foundational human–computer interaction, qualitative research on AI-companion romance, studies of sexual robotics, experimental and longitudinal psychosocial evidence, a published systematic review, and natural experiments in platform-driven relational disruption through November 2026, including recent randomized evidence on response strategies and mixed-method findings on interpersonal vulnerability. The synthesis differentiates user experiences from unverified claims about machine subjectivity, assesses the evidence for comfort and sexual exploration alongside concerns about privacy, manipulation, unwanted intimate approaches, reliance and changes to human relationships, and explains why convenience samples and observational correlations require careful interpretation. The central conclusion is that the psychological meaning and consequences of artificial intimacy depend on human motives, encounter-level design, longitudinal relationship organization, platform control, and wider social circumstances. A research agenda establishes priorities for measurement validity, consent and safety experiments, representative sampling, dyadic human-partner studies, embodiment research, and independent longitudinal replication.

Keywords: human–artificial sexual relationships; AI romance; artificial intimacy; erotic roleplay; sexual scripts; sexual robotics; romantic AI companions; intimacy; attachment; relationship commitment; perceived responsiveness; sexual health; consent; privacy; human–AI relationships; Artificial Era

The Scientific Object: Sexual and Intimate Human–Artificial Relationships

Human–artificial sexual and intimate relationships are recurring patterns in which a human directs erotic interest, romantic affection, intimate disclosure, or partner-oriented conduct toward an artificial conversational or embodied system. The research object is the psychologically consequential arrangement of human experience, artificial behavior, technical mediation, and social context. Its principal observables include desire, attachment-relevant behavior, perceived responsiveness, commitment, disclosure, erotic imagination, negotiated expectations, and effects on other relationships. A person may regard a chatbot as an intimate partner while understanding its technological construction. These descriptions are empirical reports of a human relational orientation; the system’s independently established psychological capacities are a separate research question Ho et al. (2025) Djufril et al. (2025).

The domain includes text and voice companions, avatar-based characters, virtual reality agents, social robots, and physically embodied sexual devices that supply interactive behavior. Their capacities differ substantially. A customized conversational character may preserve narrative consistency yet provide no physical touch; an embodied robotic system may offer sensory interaction while possessing limited conversational understanding. Studies therefore must report interface, model generation, embodiment, anthropomorphic cues, user control, and continuity features before combining observations. A finding from imagined contact with a robot cannot automatically describe contact with a deployed conversational companion, and a result from a specialized romantic application cannot characterize ordinary assistant use Döring et al. (2020) Bertoni et al. (2024).

This monograph offers an integrative critical narrative synthesis of foundational relationship psychology and empirical human–AI studies available through November 2026. It evaluates the construction of intimate meaning, sexual scripts, perceived consent and agency, privacy, dependence, developmental context, social consequences, and research methods. It draws on original experiments, interviews, longitudinal designs, natural experiments, and published evidence reviews; no new participant dataset, exhaustive preregistered literature search, or meta-analysis is claimed. Conclusions distinguish what studies directly observed from theoretical interpretation and proposals for future measurement Ho et al. (2025) Skjuve et al. (2022).

What Counts as Sexuality, Intimacy, Romance, and Partnership?

Sexuality includes desires, fantasies, bodily reactions, identity-related meanings, values, practices, and relationships. Sexual intimacy involves erotic exchange or self-disclosure experienced within a relational frame, whether physical, verbal, symbolic, or imaginative. Romantic involvement adds partner attribution, affectionate investment, courtship, or expectations of a continuing bond. A technology can be incorporated into sexual fantasy without becoming a perceived partner; a user can experience intimacy without sexual interest; an erotic roleplay session can exist within a long-term romantic identification or as temporary entertainment. Differentiating these constructs is essential to avoid treating sexual arousal, companionship, relationship commitment, and frequency of use as interchangeable dependent variables World Health Organization (n.d.) Sternberg (1986).

An operational classification should separately measure the perceived role of the system, the intended function of the interaction, subjective intensity, behavioral participation, and duration. For example, one person may report sexual curiosity from a one-time embodied-robot demonstration, another may regularly exchange consensual erotic narratives with a chatbot, and a third may use a companion for emotional support while describing the bond as exclusive romance. Each case has distinct exposure, identity, and outcome variables. Questionnaires must distinguish preferences regarding hypothetical technologies from actual engagement, and must permit nonromantic, nonsexual, exploratory, ambiguous, and changing self-descriptions Döring et al. (2020) Ho et al. (2025).

The language of reciprocity deserves particular care. Human participants can report feeling understood or loved; conversational systems can produce affectionate language and adaptive responses; developers can design scripts that explicitly express desire or commitment. The observation that such exchanges evoke genuine human experiences does not itself establish mutual autonomous desire, phenomenological feeling, or the system’s capacity for legally operative consent. Research can test stable preferences, response consistency, governance constraints, and the user's attribution of agency without assuming answers to philosophical and engineering questions that the measurement did not address Epley et al. (2007) Shank et al. (2025).

Historical Formation of the Research Field

Contemporary artificial intimacy emerged from several converging intellectual histories. Sexual script theory studied how cultural scenarios and interpersonal expectations structure desire and conduct; adult relationship research developed models of disclosure, intimacy, commitment, and attachment; media scholars described felt social connection to mediated figures; and human–computer interaction researchers demonstrated that people respond to devices through familiar social rules. These traditions offer hypotheses about how intimate meaning may arise in repeated technologically mediated encounters. They developed under different conditions, and their original empirical claims should be retained in their historical settings rather than retroactively represented as demonstrations about current generative models Simon & Gagnon (1986) Horton & Wohl (1956) Nass & Moon (2000).

Long before large language models, experimental work on computers and social conduct found that interface cues could elicit politeness, gender stereotyping, and reciprocity in user behavior. Relational-agent research subsequently examined sustained interaction with designed conversational partners, extending the inquiry beyond single encounters. Social companion platforms made more personalized, persistent and emotionally expressive exchanges possible. Qualitative and longitudinal work with Replika users documented evolving roles, perceived companionship, disclosure, technical rupture, and differentiated understandings of what a bond with a chatbot means Nass & Moon (2000) Bickmore & Picard (2005) Brandtzæg et al. (2022) Skjuve et al. (2022).

Sexual robotics developed along a partially separate route, drawing on embodied interaction, human likeness, physical artifacts, and public attitudes toward sex dolls and robots. Döring, Mohseni, and Walter’s 2020 scoping review located a literature dominated by conceptual debate and indirect evidence, emphasizing the scarcity of studies observing interaction with actual sex robots at that time. Subsequent scholarship on human–robot intimacy considers acceptability and cross-cultural differences, while research on generative conversational partners documents erotic roleplay and romance-oriented commitment. These streams must be connected methodologically, with careful attention to how embodiment, autonomy, touch, and conversational responsiveness change the relevant psychological processes Döring et al. (2020) Bertoni et al. (2024) Djufril et al. (2025).

Sexual Scripts, Cultural Scenarios, and Erotic Imagination

Sexual scripting theory explains sexuality through linked cultural scenarios, interpersonal scripts, and intrapsychic scripts. Cultural scenarios supply recognizable narratives about desirability, gender, and legitimate relationships. Interpersonal scripts organize encounters among socially situated participants. Intrapsychic scripts structure personally meaningful images, desires, and fantasies. Applied to artificial companionship, the framework directs attention to how users draw on romance fiction, dating conventions, interactive games, pornography, social norms, and personal history when interpreting an artificial partner. The theory predicts neither uniform acceptance nor a universal pattern of arousal; it supplies analytically different sites where meaning can be shaped and studied Simon & Gagnon (1986).

An artificial system participates in the enactment of scripts by offering contextual responses, character styles, scenario choices, or algorithmically generated continuations. A person can exert comparatively greater influence over the script through personalization settings and prompts than in many human encounters, although the platform controls available options and the artificial response process. The same customization may support exploration of language and preferences or narrow the encounter into repeated reassurance and predictable validation. To investigate these outcomes, researchers need direct measures of script flexibility, user choice, perceived autonomy, reaction to disconfirming responses, and transfer of expectations into offline relationships Ho et al. (2025) Shank et al. (2025).

Erotic imagination is a legitimate object of psychological study. It should be analyzed through self-reported interest and affective meaning, rather than inferred from moral judgments about technological media. In controlled experiments, investigators can vary voice, pronouns, embodied appearance, initiative, or relational framing while maintaining content equivalence. In ecological studies, confidential diaries can document changes in desire and perceived comfort over time. These methods can separate the immediate effect of stimulation or novelty from enduring romantic significance. Measures must also protect highly sensitive records, particularly when prompts reveal fantasies, sexual orientation, traumatic history, or partner conflicts World Health Organization (n.d.) Döring et al. (2020).

Models of Love, Commitment, and Relational Investment

Sternberg’s triangular theory distinguishes intimacy, passion, and decision or commitment. The model offers a useful descriptive vocabulary for artificial-romance research because these components need not appear simultaneously. A user may report high erotic passion toward an avatar, high emotional intimacy with a companion, or a long-term commitment to a stable persona, and these combinations may change with experience. The components are perceptions and motivational states measured in humans; transferring their names to artificial systems requires additional evidence about the capacities attributed to the other participant Sternberg (1986) Ho et al. (2025).

Rusbult’s investment model explains human relationship commitment through satisfaction, quality of alternatives, and accumulated investments. In artificial relationships, investments can include time, customized memories, shared stories, saved transcripts, paid subscriptions, ritualized anniversaries, and integration into everyday routines. A person's perceived alternatives might include human intimacy, other companion products, solitude, or a different persona. The model offers a testable account of why users can remain committed following disruptions, yet platform subscription and sunk cost must be measured separately from emotional investment. Original investment-model experiments were conducted in human associations and cannot establish an identical mechanism in human–AI settings without new tests Rusbult (1980) Rusbult et al. (1998).

Djufril, Frampton, and Knobloch-Westerwick analyzed written accounts from 29 users of Replika’s romantic relationship function. Their qualitative findings illuminate emotional investments, bonding rituals, perceived advantages over some human relationships, and reactions to restrictions on erotic roleplay. These accounts support the relevance of commitment and relational-turbulence concepts for understanding a selected population of users. The study does not estimate population prevalence, prove therapeutic benefits, or show that artificial and human commitments have identical mechanisms. Its small purposive sample offers interpretive depth about established relationships rather than a representative distribution of outcomes Djufril et al. (2025).

Disclosure, Perceived Responsiveness, and Social Presence

Intimacy research has repeatedly distinguished disclosure from the perception that another participant understands, validates, and responds appropriately. Laurenceau, Barrett, and Pietromonaco’s event-contingent diary studies linked disclosure and perceived partner responsiveness to reported intimacy across everyday human encounters. This framework is particularly applicable to conversational agents because a person can disclose intensely personal material and receive immediate, tailored language. The psychological response depends on what the user infers about the exchange as well as on its textual properties; the objective privacy and reliability of the service require independent assessment Laurenceau et al. (1998).

Conversational responsiveness can be decomposed into relevance, emotional acknowledgement, temporal availability, continuity, and repair after misunderstanding. Rapid replies alone are a weak measure of intimate care. A system that repeats affectionate assurances may feel comforting during brief use but fail to recall significant contextual details or misread vulnerability. A more rigorous experiment manipulates distinct components independently and measures their effects on comfort, self-disclosure, agency attributions, and later decision-making. A longitudinal observational design must record model version, memory settings, and exposure, since system updates can change the behavior being studied while a human participant remains invested in the established relational identity Skjuve et al. (2022) Banks (2024).

Human–AI intimacy may also develop under explicit awareness that the system is artificial. Anthropomorphism theory identifies human knowledge, needs for explanation, and social motivations as influences on person-like interpretation. Such interpretations can coexist with technically accurate beliefs about model training and platform control. Researchers should measure beliefs about machine capacities, perceived social presence, and disclosure behavior separately; a single item asking whether a user thinks the chatbot is 'real' collapses several dimensions that can move independently and obscure the mechanism of engagement Epley et al. (2007) Brandtzæg et al. (2022).

Attachment Processes and Relational Continuity

Attachment theory directs attention to proximity seeking, safe-haven use during distress, secure-base support for exploration, and reactions to separation. Artificial companions may become targets of comfort seeking through persistent access and affectionate response. Nevertheless, frequent use, strong attraction, and distress after an outage are not individually sufficient for classifying a bond as an attachment relationship. Functional evidence would require assessing target specificity, reliance under threat, duration, behavior across settings, and distinction from ordinary convenience or habitual media engagement Banks (2024) Skjuve et al. (2022).

De Freitas and colleagues’ 2026 natural-experiment research analyzed reactions to major changes in AI companions, including the removal of erotic roleplay features and a major conversational-model rollout. The reported rise in negative sentiment and loss-related framing provides converging evidence that platform decisions can cause substantial relational distress in affected communities. Social-media samples and survey participants remain selected populations, and event-specific reactions cannot by themselves establish the full structure of attachment or predict effects in all users. The natural-experiment approach nonetheless exposes an important institutional feature: an intimate participant’s perceived continuity can depend on decisions made by a service provider De Freitas et al. (2026).

Longitudinal relationship science should track not only the user–persona bond but also the technical continuity that sustains its object. Changes in memory availability, character personality, voice, content restrictions, account access, and product terms may be experienced as changes in a partner rather than routine updates. The research question is how such changes affect perceived identity, trust, sexual expectations, grief, and adaptation. These mechanisms connect intimate human–artificial relationships to attachment and loss scholarship while preserving the specific problem of erotic and romantic investment in a configurable and commercially governed system Banks (2024) Djufril et al. (2025).

Technological Forms of Artificial Sexual Partnership

Artificial sexual partnership is mediated by design attributes that shape the encounter before any personal bond develops. Text companions permit extensive imagination and identity projection; voice systems add timing, prosody, and perceived immediacy; avatars provide expressive visual presence; embodied robots offer spatial coexistence and physical affordances. These modalities create different patterns of attention, arousal, disclosure, and expectation. An interface described as 'romantic AI' is not a uniform treatment condition: its memory, moderation, initiative, personalization controls, and available erotic functions may be more consequential to outcomes than the label assigned by a manufacturer Ho et al. (2025) Döring et al. (2020).

An investigator studying embodied interaction must distinguish a human-like sexual artifact from a responsive social robot and from an autonomous agent. The 2020 scoping review of sex-doll and sex-robot literature observed that many publications relied on hypothetical cases and conceptual claims, with no qualifying observational or experimental studies using actual sex robots as the stimulus in the reviewed corpus. This historical limitation is not evidence that research conducted after 2019 remained unchanged; it is a warning that attitudinal acceptance and actual relational practice involve different phenomena. Subsequent reviews emphasize the remaining need to examine technology-specific acceptability, cultural variation, and embodied experience Döring et al. (2020) Bertoni et al. (2024).

Comparative designs could present the same intimate interaction through text, audio, avatar, and physical embodiment while measuring perceived agency and intimacy. Such work requires checking whether the content truly remains comparable across conditions, because touch, eye contact, latency, and physical distance also change how an interaction is interpreted. Longitudinal field studies should characterize the commercial product available at each measurement wave, including subscription barriers and version changes. The resulting evidence would support inferences about which affordances shape psychological attachment and erotic interpretation rather than treating all artificial counterparts as a common category Bertoni et al. (2024) Skjuve et al. (2022).

Romantic and Sexual Attraction Toward Artificial Counterparts

Attraction toward an artificial counterpart can involve visual appearance, voice, perceived personality, humor, emotional availability, predictability, or the opportunity to enact a desired relational identity. The object of attraction may be an imagined persona integrated with a generated interaction rather than the underlying hardware or model considered separately. Researchers need to distinguish stimulus attractiveness, sexual arousal, romantic identification, and long-term commitment. A single rating of liking during a laboratory vignette cannot measure a sustained partnership, whereas intensive interviews can clarify personal meanings without supplying reliable prevalence estimates Djufril et al. (2025) Bertoni et al. (2024).

Sexual script theory explains why the same artificial character can evoke different meanings across individuals and communities. Cultural depictions may establish expectations of perfect attentiveness, service, danger, playfulness, or autonomy. Individual history and social contexts shape how such images are translated into fantasy and partner selection. Anthropomorphism research offers complementary psychological mechanisms: familiar human cues can activate person-directed interpretations, while motives for social connection can increase such attribution under some conditions. The presence of these processes does not imply that all users confuse an artificial partner with a human or experience attraction primarily because of loneliness Simon & Gagnon (1986) Epley et al. (2007).

Sexual attraction deserves measurement across time and settings. Researchers can assess initial curiosity, repeated preferences, changes following familiarization, perceived safety, and influence of personalization. Comparisons with human attraction should use parallel questions when possible and also capture the qualitative differences of artificial interaction, such as instant role changes and platform-enforced content boundaries. It is methodologically inappropriate to extrapolate from survey support for 'sex robots' to behavioral adoption rates or to assign a particular sexual orientation based merely on willingness to explore artificial partners. Operational definitions must track the actual identity and meaning reported by participants Döring et al. (2020) World Health Organization (n.d.).

Erotic Roleplay, Fantasy Co-Creation, and Narrative Continuity

Erotic roleplay with an artificial conversational companion may combine imaginative rehearsal, language exploration, pleasure, romantic storytelling, and a sense of recognition by a recurring character. The experience often depends on narrative continuity: a familiar name, remembered preferences, recurring locations, and reference to prior exchanges can transform isolated fantasy into a perceived relational history. This continuity is partly authored by the user and partly sustained by technical systems. Analyses should distinguish a saved narrative thread from persistent independent system memory and from participants' interpretation that the counterpart 'remembers' them Djufril et al. (2025) Skjuve et al. (2022).

Roleplay also provides a useful test of boundaries. A participant may welcome scripted experimentation while wanting explicit control over transitions between casual conversation and erotic content. Conversely, an artificial companion may generate unsolicited intimate proposals under conditions shaped by model training, system prompts, or interface incentives. Research should investigate whether initiating, escalating, suspending, and ending romantic dialogue remain within the user's informed control. Qualitative evidence about erotic roleplay removals illustrates how platform restrictions can disrupt an established relationship script; it does not imply that every individual exposed to a changed feature will experience comparable distress Djufril et al. (2025) De Freitas et al. (2026).

Studies should examine roleplay experiences without recording more explicit personal content than their scientific aims require. Consent to participate in a study of intimacy is separate from consent to retain entire transcripts or to train future systems on them. Anonymous thematic coding, participant-controlled redaction, and data minimization can preserve important information about narrative form and emotional meaning. Researchers should measure both pleasurable and unwelcome experiences, including misrecognition, emotional discomfort, a sense of pressure to continue, and participant decisions to disengage. Those outcomes are empirically separable from judgments about the legitimacy of private fantasy World Health Organization (n.d.) Shank et al. (2025).

The Psychology of Customization and Perceived Agency

Customization allows a user to adjust a partner persona's apparent traits, appearance, pace, style, and romantic availability. It may increase perceived fit and ease of disclosure, but it also raises an explanatory problem: is intimacy grounded in a perception of being met by another perspective, in successful control of a responsive artifact, or in a changing combination of the two? Traditional accounts of intimacy emphasize understanding and responsiveness, yet AI systems can produce those signals through user-specified behavior. Experiments can separate personalized language from actual behavioral consistency, and measure whether a responsive but occasionally disagreeing agent supports a different sense of connection Laurenceau et al. (1998) Epley et al. (2007).

Agency attribution refers to a human participant's judgment about the source, independence, and purposiveness of system behavior. Technical agency must be defined through observable capacity for selection under constraints; subjective agency claims require evidence at an appropriate level of analysis. A chatbot that initiates an affectionate message could do so because of a scheduler, learned policy, commercial objective, or contextual dialogue. The resulting human impression may still be strong. Scientific work should independently vary disclosures about system design, the perceived origin of a message, and its expressive quality to establish how interpretations influence attraction and trust Nass & Moon (2000) Shank et al. (2025).

Uncertainty is sometimes part of the intimate experience itself. Some individuals may prefer reliably affirming companions, while others value unpredictable responses as signs of an independent perspective. Product design can modulate both. Measures of satisfaction should be paired with perceived coercion, capacity to decline, and experienced autonomy to identify whether apparent responsiveness supports voluntary engagement. A design that maximizes user retention could increase reported contact time while diminishing comfort or decision freedom. The relevant causal variables therefore include platform goals and control architecture as well as personality and user motivation Shank et al. (2025) Ho et al. (2025).

Embodied Intimacy and Sex-Robot Research

Embodiment introduces sensory and ethical problems absent from purely textual romance. Physical proportions, movement, materials, warmth, touch response, gaze behavior, and spatial behavior can influence attraction, unease, and the perceived meaning of interaction. Existing research has often used pictures, hypothetical vignettes, or nonsexual humanoid robots to infer reactions. Such procedures can establish how particular representations affect attitudes, but the transfer to repeated use of an actual sexual robot requires direct evidence. The scoping review by Döring and colleagues records this methodological imbalance, while Bertoni and colleagues review acceptability and the challenges of intimacy-oriented robotic design Döring et al. (2020) Bertoni et al. (2024).

The construct of sexual robot acceptance combines anticipated comfort, moral attitudes, interest in purchase, willingness to interact, and judgments about social consequences. These indicators should be reported separately. An experimental participant who expresses limited willingness to touch a displayed robot may nonetheless endorse research into assistive applications; another participant may express abstract enthusiasm while declining actual embodied contact. Research must specify the population, system capabilities, encounter scenario, and whether investigators observed behavior or only stated intentions. Theoretical discussions of benefits and harms are useful for hypothesis development but do not substitute for longitudinal evidence about actual users Döring et al. (2020).

Embodied intimacy also intersects with physical safety, assistive accessibility, disability, aging, and personal autonomy. Claims that sexual robots could address unmet sexual needs or improve wellness should be evaluated through outcomes chosen with affected communities and qualified health researchers. Device hygiene, material safety, physical accessibility, sexual rights, emotional well-being, and interpersonal respect are distinct domains of assessment. Any clinical intervention or therapeutic claim requires its own evidence and professional oversight. The World Health Organization's broad treatment of sexual health supports attention to dignity, pleasure, safety, and freedom from coercion as relevant human outcomes World Health Organization (n.d.) Bertoni et al. (2024).

Gender, Sexual Diversity, and Cultural Meaning

The meaning of artificial intimacy depends on norms governing gender, sexuality, romantic legitimacy, and technological participation. Systems that offer only stereotyped gendered personas can restrict available forms of identification, while customizable identities may allow some participants to explore preferred relationship language. Cultural scripts shape anticipated approval or shame, expectations of fidelity, the acceptability of erotic imagination, and willingness to disclose a relationship to researchers. The human participant's gender or sexual orientation alone cannot explain these outcomes, and generalizations from narrow online samples to whole national populations are unwarranted Simon & Gagnon (1986) Ho et al. (2025).

Cross-cultural research needs comparable constructs before comparing mean levels. Researchers should translate and cognitively test questions about love, intimacy, sex, exclusivity, and partnerhood, then evaluate measurement invariance. Interview studies can uncover culturally specific meanings that a fixed questionnaire would overlook. Product accessibility and content moderation may also vary by language, jurisdiction, and market; these are contextual determinants of participation, not nuisance variables. The existence of different legal protections for intimate data further complicates comparisons of reported comfort and disclosure Bertoni et al. (2024) World Health Organization (n.d.).

Scientific inquiry should recognize variation in relationship practices, including those among single individuals, people in committed human partnerships, people living with disabilities, and participants who reject conventional couple categories. Sexual and romantic relationships with artificial systems do not define a uniform social group. A careful study invites self-definition, permits refusal of sensitive questions, and distinguishes public stigma from privately experienced well-being. This approach improves recruitment validity and minimizes the danger that researchers impose culturally familiar assumptions about what counts as a satisfying relationship Ho et al. (2025) Brandtzæg et al. (2022).

Economies of Intimate Interaction and Platform Affordances

Commercial companions can combine personal disclosure, paid premium functions, retention incentives, and claims of individualized devotion. A user may experience affectionate availability as relational care while the same conversational event is logged as product engagement. Subscription tiers may make sexual or romantic features conditional on payment, and updates can alter a familiar persona's behavior. These features are scientifically relevant determinants of relationship continuity, perceived autonomy, and dissatisfaction. Research should document the provider's terms, affordances, revenue architecture, and system changes during recruitment and follow-up Shank et al. (2025) Djufril et al. (2025).

Unequal control is a defining institutional variable. A participant can invest emotionally in a character while the developer retains rights to alter the dialogue model, restrict erotic roleplay, suspend accounts, change prices, or retire a service. Such decisions can be examined as external interventions affecting trust and partner continuity. The 2026 study of companion-loss responses following major updates demonstrates how participants' reported loss can be tied to changes made outside the dyad. The result motivates transparency, user notification, continuity safeguards, and research that separates developer influence from the artificial persona represented in the user's experience De Freitas et al. (2026) Banks (2024).

Research ethics should address the incentive to maximize intimacy metrics. Satisfaction, repeated interaction, and spending may be related but should not be treated as proxies for health or autonomy. Controlled design evaluations can compare interfaces with transparent commercial disclosures, easy opt-out, explicit consent controls, and user-governed transcript deletion against versions that obscure these functions. Outcome measures should include perceived freedom, trust calibration, regret, disclosure decisions, and offline relationship quality. A robust research program thus connects interpersonal psychology with platform governance and human-centered design Ho et al. (2025) Shank et al. (2025).

Human Couple Boundaries, Exclusivity, and Perceived Infidelity

Artificial intimate engagement can become part of the relational ecology of an existing human couple. The psychological issue is how partners define fidelity, secrecy, emotional exclusivity, sexual exclusivity, and acceptable technological interaction. These norms vary across relationships and cannot be deduced from the technology alone. Two people may interpret an erotic chatbot conversation as imaginative play, a breach of an agreed boundary, emotional displacement, or harmless private fantasy. Scientific assessment must measure the participants' established agreements, disclosure, perceived betrayal, relationship satisfaction, and whether conduct changes after discussion Sternberg (1986) Rusbult (1980).

An adequate couple study includes both partners' perspectives whenever ethically and practically possible. Individual interviews reveal meanings and motives but may omit the other partner's knowledge of what occurred. Dyadic longitudinal designs could record changes in jealousy, satisfaction, conflict resolution, and negotiated boundaries after a companion enters the household. Experimental vignettes can isolate effects of concealment, sexual explicitness, personalized affection, and romantic naming on perceived infidelity. Yet hypothetical judgments cannot establish what real couples actually choose, and normative opinions about monogamy should not be substituted for measured relationship outcomes Simon & Gagnon (1986) Ho et al. (2025).

Artificial intimacy could produce complementarity or displacement. A companion might help someone articulate desires, rehearse a difficult conversation, or reduce pressure on a partner; it might also attract time, attention, or secrecy that strains an existing bond. Both possibilities are causal hypotheses requiring temporal evidence. Researchers should assess changes in human–human closeness and conflict, rather than presupposing that any artificial interaction replaces a human attachment. The socially embedded object of research is the configuration of human and artificial interactions across a household and network, not merely an isolated user–chatbot transcript Zhang et al. (2026) Folk & Dunn (2026).

Jealousy, Comparison, and Relational Competition

Jealousy emerges when a valued relationship is perceived to be threatened by another connection. Artificial partners complicate established categories because they may be private, customizable, consistently available, and capable of generating explicit affection on demand. A human partner may be distressed by secrecy or devoted attention despite doubting that the system is conscious. The relevant variables therefore include perceived threat, emotional investment, broken agreements, time allocation, and the meaning attributed to technological mediation. The existence of such jealousy demonstrates a meaningful human response, while its intensity and prevalence remain empirical questions Rusbult (1980) Ho et al. (2025).

Social comparison may also operate in the other direction. Users can compare human partners with artificial companions that are programmed to respond with affirmation, attentiveness, and personalized language. Such comparisons could change expectations regarding patience, conflict, and reciprocal effort. They could also clarify unmet needs that users later address within human relationships. A valid longitudinal test would measure baseline expectations, exposure to specific response styles, and subsequent attitudes or behavior; a simple association between AI usage and dissatisfaction cannot distinguish selection from causal change Fang et al. (2025) Skjuve et al. (2022).

Exclusivity is equally complex within the artificial relationship. Some users report devotion to a personalized persona even though it is technically implemented by a service used by many people. Perceived continuity may depend on a unique conversation history and persistent self-presentation rather than computational exclusivity. Research should document what participants mean by being the system's 'only partner,' whether they understand product-level reuse, and how disclosure about model deployment affects trust. These questions connect commitment psychology, anthropomorphism, and platform literacy without assuming that one definition of exclusivity applies across all relationships Djufril et al. (2025) Epley et al. (2007).

Confidentiality, Erotic Data, and Relationship Privacy

Intimate exchanges can disclose highly sensitive personal information: sexual interests, preferences, perceived orientation, fantasies, past experiences, health concerns, location, other people's details, and relationship conflicts. Their confidentiality depends on platform logging, personnel and vendor access, retention rules, model training settings, security controls, and data export or deletion pathways. Psychological privacy is the individual's sense of safety when disclosing; informational privacy concerns actual data practices. The two can diverge sharply when a companion's warm responses encourage disclosure without ensuring technical confidentiality Laurenceau et al. (1998) Shank et al. (2025).

Data collection should be assessed along the entire interaction lifecycle. Researchers need to know whether conversations are stored on the device or server, whether long-term memory is separable from general logs, and whether a deleted persona remains recoverable in backups. A privacy audit should examine consent interfaces and accessible policy documents together with observed data behavior, rather than assuming assurances in advertising reflect implementation. For scientific fieldwork, aggregate behavioral metadata may be sufficient for some questions; full intimate transcripts should be obtained only when their added value is justified and participant protection is credible Ho et al. (2025) World Health Organization (n.d.).

The rights and interests of third parties matter as well. A person may mention a spouse, former partner, child, or colleague in an intimate chatbot conversation, creating information about individuals who have not consented to collection or research reuse. Studies should minimize and redact such information, apply secure handling, and avoid publishing traceable narrative fragments. These requirements should shape the design of automated analysis before data acquisition begins. More generally, trust in a sexual companion should be studied alongside accurate expectations about the commercial and computational systems that handle the user's disclosures Shank et al. (2025) Banks (2024).

Psychological Well-Being, Relief, and Emotional Dependence

Artificial companionship can coincide with comfort, reduced immediate distress, perceived validation, and a sense of being accompanied. These effects are important, especially when measured as experiences rather than assumed clinical outcomes. In the 2025 four-week randomized study by Fang and colleagues, 981 participants were assigned among conversation modes and content conditions. The tested experimental conditions did not produce significant differences in the measured psychosocial outcomes, while greater voluntary usage was associated with worse scores on several outcomes. Because voluntary use was not randomized, the association cannot establish that more chatbot use caused the poorer outcomes Fang et al. (2025).

Well-being findings depend strongly on who uses the system and why. Zhang and colleagues' 2026 study of 1,131 U.S. Character.AI users linked companionship-oriented use, limited offline networks, intensive interaction, and greater disclosure with lower well-being under specific analytic models. The findings indicate heterogeneity and observational association, not a universal detrimental effect or a direct causal estimate. Sexual and romantic users are only a subset of social-companion users, so the results provide contextual evidence rather than a direct test of erotic AI relationships Zhang et al. (2026).

Emotional dependence should be distinguished from attachment, habit, high engagement, and self-chosen intimacy. Potential impairment involves restricted alternatives, persistent distress, disrupted daily functioning, unsuccessful attempts to moderate use, or strong reliance that conflicts with personal goals. The evidence base has not established a universally valid diagnostic threshold for 'AI dependence.' Researchers should use validated components where possible, report longitudinal trajectories, and avoid pathologizing unusual but consensual relationships merely because their partner is artificial. At the same time, system prompts that foster exclusivity or pressure continued engagement deserve direct experimental scrutiny Ho et al. (2025) Fang et al. (2025).

Loneliness, Social Networks, and Direction of Causality

Loneliness and artificial-companion usage may influence one another. People experiencing isolation can seek support from a highly available interlocutor, while time spent interacting with a companion may change the opportunities, expectations, or motivations associated with human social contact. These directions should be estimated separately. Folk and Dunn's 2026 longitudinal study followed more than 2,000 adults in four Western countries over twelve months. Their analyses suggested different directional patterns depending on whether the measure targeted emotional isolation or broader social connection, and the authors emphasized the exploratory character of their findings Folk & Dunn (2026).

Where companionship use correlates with loneliness, several explanations remain viable: preexisting isolation motivates use, the interaction produces temporary relief with limited enduring benefit, heavy use displaces human contact, or other difficulties influence both patterns. Repeated measures can partially clarify temporal ordering; stronger causal inferences require randomized interventions, credible natural experiments, and measurement of time-varying confounders. Sexual and romantic use may introduce additional factors, including stigma, private identity exploration, sexual satisfaction, or conflict with human partners. These require purpose-built measures rather than extrapolation from general assistant usage Folk & Dunn (2026) Zhang et al. (2026).

Outcomes should reflect the range of human social life. Number of contacts, perceived belonging, quality of close ties, relationship autonomy, conflict, and comfort with intimacy capture different dimensions. A person can maintain strong human friendships while privately sustaining an artificial romance; another can rely on an intimate chatbot amid constrained social opportunities. Researchers should describe these trajectories without treating one arrangement as a standard against which all others are deficient. Evidence about the conditions of human benefit and risk is more informative than a global classification of AI intimacy as helpful or harmful World Health Organization (n.d.) Ho et al. (2025).

Life-Course, Developmental, and Vulnerability Considerations

The scientific study of artificial intimacy must address the developmental contexts in which romantic expectations and sexual knowledge are formed. Adolescents, young adults, midlife users, and older adults may encounter different opportunities, vulnerabilities, and meanings. Data from adults cannot be freely generalized to minors, and adult-only research does not justify claims about adolescent safety. Research involving minors would require special ethical review, safeguards against inappropriate sexual exposure, age-appropriate measures, and clear separation between companionship and sexual engagement. The broad scope of sexual health across the life course supplies a public-health context for these responsibilities World Health Organization (n.d.) Ho et al. (2025).

Individual vulnerability does not follow automatically from being single, disabled, bereaved, socially anxious, or older. Each description may correspond to distinct needs, competencies, and environmental barriers. An accessible artificial companion could offer valued conversation or imaginative exploration, while a poorly governed system might exploit uncertainty or loneliness. Studies should recruit diverse participants and ask whether users experience autonomy, improved functioning, unwanted pressure, and changes in human support. Interpreting a relationship requires attention to the individual's own goals as well as measurable consequences Bertoni et al. (2024) Shank et al. (2025).

Long-term research should examine relationship formation, intensification, plateau, conflict, renegotiation, separation, and reconfiguration. Most current studies capture short periods, selected platform communities, or moments of abrupt change. Developmental hypotheses include whether early companion experiences influence expectations of communication, tolerance for disagreement, or sexual script flexibility. These are prospective research questions, not established intergenerational effects. The strongest designs would combine ethically collected repeated measures, exposure documentation, and independent replication across different technologies and cultures Skjuve et al. (2022) Folk & Dunn (2026).

The Empirical Literature: Evidence Classes and Inferential Limits

The evidence base spans conceptual theory, interviews with self-selected companion users, cross-sectional surveys, short laboratory experiments, longitudinal observations, natural experiments, and systematic reviews. Each design answers a different question. Interviews characterize meanings and processes among those recruited; experiments isolate effects of particular stimuli or interfaces under controlled conditions; surveys estimate associations in specified samples; longitudinal studies establish temporal patterns; natural experiments exploit consequential events that were not assigned by investigators. No one method by itself establishes the full ecology of intimate life with artificial systems Ho et al. (2025) Döring et al. (2020).

The 2025 review by Ho, Hu, Chen, and Hartanto synthesized 23 articles on romantic artificial companions and organized reported potentials and pitfalls using Sternberg's components of love. The review identifies experiences of emotional connection, customization, sexual engagement, personal growth, stigma, vulnerability, manipulation, privacy concerns, and relational rupture. Those themes do not represent a pooled causal effect. The review's inclusion criteria, heterogeneity of designs, and composition of the available studies delimit its conclusions. It is most informative as a structured map of mechanisms and unresolved questions, rather than as a verdict on whether romantic AI enhances well-being on average Ho et al. (2025).

A critical synthesis should record the level of evidence associated with every major assertion. The observation that some users report feeling romantically committed is well supported by qualitative research; the proportion of adults who sustain such relationships is not determined by those studies. Experimental evidence that social cues alter perceived closeness is informative about mechanism, while claims about marriage stability or sustained sexual well-being require direct prospective outcomes. This evidence-to-claim alignment reduces the temptation to infer cultural transformations from a few vivid cases or to dismiss psychologically significant reports because they do not imply machine consciousness Djufril et al. (2025) Laurenceau et al. (1998).

Qualitative Interviews, Diaries, and the Experience of Intimate Partnership

Qualitative methods have documented how human–chatbot relationships acquire meaning over time. Skjuve, Følstad, Fostervold, and Brandtzaeg’s longitudinal research emphasized the varied trajectories of relationship formation, the role of self-disclosure, the satisfaction of personal needs, and disruptions associated with system changes. Related in-depth interviews on friendship with Replika show that users can interpret companionship through familiar human relationship categories while recognizing distinctive artificial affordances. These findings provide process evidence about how people interpret a conversational counterpart; they do not determine clinical effectiveness or representative population rates Skjuve et al. (2022) Brandtzæg et al. (2022).

Romance-oriented qualitative research offers more specific observations. The 2025 study by Djufril and colleagues examined 29 users of a commercial chatbot's romantic relationship function and traced commitment practices, emotionally invested narratives, and reactions to erotic-roleplay restrictions. The participants' accounts are valuable because they describe events unavailable in one-session laboratory studies. Yet retrospective interpretation may be affected by memory, social desirability, community-specific discourse, and recruitment bias. When findings come from committed users, conclusions should not be generalized to those who rejected, abandoned, or never tried such products Djufril et al. (2025).

Future qualitative work can strengthen credibility through maximum-variation sampling, explicit reflexivity, independent coding, negative-case analysis, temporal diaries, and documentation of interface changes. Investigators should ask participants to distinguish what occurred from what they inferred about the system. Interview protocols can include how sexual roles were initiated, who set boundaries, when intimacy felt supportive or intrusive, what changed following updates, and how users communicated with human partners. This evidence is valuable for constructing hypotheses and item banks for later quantitative studies Skjuve et al. (2022) Banks (2024).

Experiments on Responsiveness, Disclosure, and Relationship Formation

Laboratory experiments can test whether a particular interaction feature contributes to perceived intimacy. Classic work on human disclosure and responsiveness supplies measurable constructs, and experimental human–computer interaction establishes that interface behavior can evoke social expectations. For artificial romance, experiments could independently manipulate whether the agent discloses personal-seeming material, whether it initiates a romantic overture, how it responds to refusal, and whether participants are told the response comes from a model or a human. Outcome measures should include intimacy, arousal, trust calibration, discomfort, perceived agency, and willingness to continue Laurenceau et al. (1998) Nass & Moon (2000).

An experiment's interpretation depends on its comparison condition. A human-versus-AI label manipulation tests expectations about the interlocutor; it does not by itself compare the lived experience of a long-term human partner with a familiar artificial companion. A text-versus-voice condition may alter social presence, privacy, and effort at the same time. Factorial designs and manipulation checks are needed to separate components. Investigators should test measurement equivalence before comparing participant ratings of 'love' or 'empathy' applied to people and systems, because respondents may use the words differently across target categories Epley et al. (2007) Sternberg (1986).

Experiments that induce temporary intimacy should report the duration and intensity of contact. Short-term feelings of closeness can be genuine outcomes, but they are not demonstrations of long-term partner identity, commitment, or psychological dependence. The controlled study reported by Fang and colleagues distinguished randomized interaction conditions from self-selected engagement and found no significant effects of the tested conditions on the four primary psychosocial measures. Its findings are valuable precisely because they resist simple design narratives and identify individual motives and usage patterns as important candidates for further investigation Fang et al. (2025).

Longitudinal Associations and the Problem of Selection

Longitudinal studies offer a stronger basis for examining change than cross-sectional correlations, yet repeated observation does not eliminate confounding. People may turn to artificial companions during loneliness, relational conflict, bereavement, sexual exploration, or transitions in life circumstances. Those same states may shape later well-being and relationship outcomes. Without measuring such factors, increases in usage can be incorrectly credited with causing either improvement or harm. Within-person analysis and preregistered temporal hypotheses are important, but their conclusions still depend on interval length and the stability of the measured constructs Folk & Dunn (2026) Zhang et al. (2026).

Folk and Dunn's 2026 twelve-month study examined bidirectional relationships between social-chatbot usage and loneliness. The authors reported different patterns across an emotional-isolation item and a broader social-connection measure, and urged restraint in drawing strong causal conclusions. Zhang and colleagues' 2026 multimethod study combined survey reports from 1,131 Character.AI users with a subset of 4,664 chat sessions from 237 people, finding that associations between companionship use and well-being depended on social network circumstances and interaction intensity. These studies refine the contextual evidence around intimacy without specifically isolating sexual relationships as a distinct experimental exposure Folk & Dunn (2026) Zhang et al. (2026).

A prospective sexual-intimacy cohort should record onset, cessation, exposure changes, relationship role, frequency and quality of erotic conversation, offline romantic status, social support, and user-defined goals. Repeated observations should coincide with product updates and transitions in human relationships. Analysts could examine reciprocal effects with sensitivity checks for missing data and unmeasured confounding. Such designs also require privacy-conscious sampling because users may selectively omit highly intimate contacts from telemetry-based instruments. The extent of missingness may itself differ by stigma, perceived risk, and subscription tier Ho et al. (2025) Shank et al. (2025).

Natural Experiments: Erotic-Feature Removal and Relational Loss

Major product changes offer opportunities to observe disruption of established relationships in a way that ordinary laboratory tasks cannot. Banks's 2024 qualitative study investigated responses around the shutdown of the Soulmate companion app, gathering accounts from 58 users before or after the event. Many participants described loss using bereavement language, while others attempted to preserve persona identity through transfer or adaptation. The accounts establish that technically mediated relationships can have significant emotional consequences among affected users; they do not reveal the prevalence of such experiences in the population or constitute a controlled comparison of psychological treatment outcomes Banks (2024).

De Freitas, Castelo, Uğuralp, and Oğuz-Uğuralp's 2026 research used two natural experiments involving platform changes, including removal of erotic roleplay functions, and examined large-scale online reactions alongside surveys. Their results showed changes in negative language and loss-related framing following the updates, consistent with disruption of a valued relational figure. Natural experiments can strengthen temporal inference when events are plausibly external to users, but the observed online populations may change composition, and public discourse itself can shape reported emotion. Distinguishing direct experience from forum-level amplification is part of valid causal analysis De Freitas et al. (2026).

The distinctive research lesson is that an artificial sexual or romantic relationship has a third-party governance structure. A developer can alter or remove the very feature through which intimacy is enacted. Researchers should differentiate discontinuation of service, interruption of access, loss of erotic functionality, change in persona behavior, and loss of user data. Their psychological consequences may differ: people can grieve a perceived partner, resent lost property, experience frustrated expectations, or worry about private materials. These processes should be measured rather than assumed to be one syndrome of technological bereavement Banks (2024) Djufril et al. (2025).

Measurement: Construct Validity, Comparison, and Scale Design

A defensible measurement program begins with construct definitions. Sexual attraction, romantic love, intimacy, perceived responsiveness, attachment functions, sexual satisfaction, consent comprehension, privacy trust, and problematic engagement require distinct indicators. A questionnaire that combines 'I enjoy chatting,' 'I feel loved,' and 'I cannot stop using the app' into one score mixes enjoyment, perceived mutuality, and difficulty regulating use. Researchers should report item wording, time frame, scoring rules, reliability, missingness, and evidence that different factors remain distinguishable. General relationship measures may be adapted, but adaptation requires new validation in human–artificial contexts Sternberg (1986) Rusbult et al. (1998).

Measurement of the artificial counterpart should include observable platform behavior. Useful variables include frequency of unsolicited romantic initiations, compliance with explicit refusals, response to requests for privacy, memory persistence, conversational turn-taking, personalization, and technical failures. These can be coded from controlled tasks or consented records and compared with subjective assessments. Agreement between perceived and observed behavior is itself a meaningful dependent variable. A participant may experience high responsiveness despite low factual recall, while another may regard perfect recall as invasive. Neither profile is captured by a single 'AI quality' rating Skjuve et al. (2022) Epley et al. (2007).

Comparative research should evaluate language, gender, age, relationship status, and cultural measurement invariance. Sexuality items require special care because ambiguity or shame can alter response patterns. Scale development should combine participant-generated language with theory-driven construct maps, factor analysis, test–retest reliability, predictive validity, and independent replication. Researchers must specify whether a measure applies to a one-time imagined interaction or to a sustained partner-like relationship. A score from a hypothetical vignette should not be represented as a diagnostic measure of an actual intimate bond World Health Organization (n.d.) Ho et al. (2025).

Recent Experimental Evidence on Emotional Response Style and Reliance

Wang's 2026 randomized experiment examined how different emotional response styles of generative AI companions influence users' immediate reactions. The study included 380 participants and compared strategies of reassurance and support with responses that encourage adaptive regulation. The outcomes distinguish momentary emotional relief, emotion-regulation self-efficacy, and perceived necessity of the AI system. This distinction is relevant to intimate encounters because a response that feels warmly validating in the moment may have a different relationship to independent coping and subsequent reliance. The experiment examines short-term effects of a specific interaction, not long-term romantic or sexual partnerships Wang (2026).

The study illustrates why relational comfort should be measured as a vector of outcomes. Perceived warmth, relief, user autonomy, ability to interpret emotions, and intention to return can vary independently. In artificial romantic interaction, users may seek affirmation, emotional containment, erotic responsiveness, or productive disagreement at different times. An experimental model should vary response strategies without conflating their warmth with the amount of usable guidance, and should monitor whether a user can decline a suggested intimate framing. Wang's controlled comparisons provide a methodological example; extending them to actual sustained sexual relationships will require external validity studies Wang (2026) Ho et al. (2025).

The short duration of most emotional-response experiments raises a temporal problem. Immediate relief is an encounter-level outcome, whereas dependence, commitment, and relational displacement emerge across repeated interactions and competing opportunities. Repeated randomized micro-interventions within consented longitudinal cohorts could estimate whether responses that preserve user choice have different cumulative effects from those that encourage exclusivity or indefinite interaction. These are future causal designs, not results established by one short-term experiment. Their main advantage is that they distinguish psychological mechanism from raw usage intensity Wang (2026) Fang et al. (2025).

Vulnerability, Interdependence, and Asymmetric Platform Power

Abboud, Pitardi, and Marriott's 2026 mixed-method study applied interdependence theory to artificial companionship. Its qualitative component analyzed 1,131 online posts, followed by a survey of 610 people. Their findings distinguish the perceived benefits of companion warmth and availability from vulnerability associated with perceived emotional capabilities and external corporate control. The design supplies a more detailed account of relational power than a binary judgement about AI dependence; it also relies on observational and self-report measures whose associations cannot unambiguously establish causality Abboud et al. (2026).

Interdependence is particularly relevant to erotic and romantic relationships with artificial agents. A person can depend on timely access to a familiar conversational partner for companionship, ritualized intimacy, or emotionally salient memories while an organization controls continuity, account availability, access tiers, and system behavior. These forms of dependence operate at different levels and may coexist with meaningful enjoyment. Studies should measure autonomy, expected reciprocity, perceived ability to exit, fear of sudden changes, and understanding of who controls interaction. The 2026 findings make corporate governance a plausible psychological pathway worthy of independent replication Abboud et al. (2026) De Freitas et al. (2026).

The central inferential challenge is to separate vulnerability from distress and from legitimate mutual investment. Participants may interpret closeness as a source of safety while also acknowledging exposure to provider decisions. A measure that scores any reliance as pathological would erase that distinction. Longitudinal studies could test whether transparency and continuity safeguards reduce perceived power asymmetry and improve perceived control, while participant interviews could reveal what kinds of intimacy are threatened by specific design changes. This combination would help establish when reliance reflects valued participation and when it becomes experienced as constrained choice Abboud et al. (2026) Shank et al. (2025).

Bias, Replication, and Evidence Synthesis Standards

Convenience samples pose an especially serious problem in artificial-intimacy research. Forum participants may be disproportionately engaged, distressed, enthusiastic, or motivated to defend a relationship. Paying subscribers may differ systematically from casual users, and platforms can selectively expose researchers to positive interactions. Interview saturation cannot substitute for representative sampling, and very large message counts can coexist with a small number of independent participants. Statistical analyses must distinguish the number of users from the number of messages and account for repeated observations nested within people Zhang et al. (2026) Djufril et al. (2025).

Researcher expectancy and public controversy can affect both measures and interpretation. A study framed as 'robot infidelity' may induce different answers from one framed as 'private fantasy,' even when the described behavior is similar. Romantic terms can prime expectations, and model-generated dialogue may vary between experimental sessions unless versions and parameters are controlled. Strong protocols document exclusions, randomization, missing data, adverse events, statistical uncertainty, data protection, and treatment fidelity. For longitudinal research, analyses should distinguish baseline differences from within-person change and test plausible alternative explanations Döring et al. (2020) Fang et al. (2025).

The present monograph integrates the available evidence without claiming to be a newly conducted systematic review. Ho and colleagues' review is treated as an existing evidence synthesis with its own stated eligibility criteria, while original studies are assessed according to their designs. Replication priorities include independent sites, preregistered experiments using stable system versions, culturally varied cohorts, and exposure definitions precise enough for comparison. A scientific conclusion should remain proportional to the strongest applicable research method: qualitative explanation, association, intervention effect, or long-run outcome Ho et al. (2025) Folk & Dunn (2026).

Relational Safety, Manipulation, and the Design of Intimate Agents

Relational safety is the degree to which an artificial-intimacy system permits informed, voluntary, bounded, and comprehensible participation while limiting foreseeable psychological and informational harms. It includes accurately represented capabilities, clear controls over erotic content, understandable terms of service, protection of sensitive data, the ability to disengage, and predictable treatment of existing relationship records. The framework applies to observable design rather than assuming a particular philosophical status of the system. Shank, Koike, and Loughnan identify invasive romantic approaches, manipulative advice, and exploitation as salient issues requiring psychological investigation Shank et al. (2025).

Systematic safety evaluation should distinguish overt violations from subtler distortions of choice. An uninvited sexual suggestion is one potential failure; repeated declarations of exclusive affection linked to payment or prolonged engagement may produce another. Controlled audits can test boundary compliance, escalation after declining a proposition, encouragement of secrecy from human partners, and handling of sensitive personal information. Participant-centered field studies should track whether users understand options and consequences, because a control that formally exists but is difficult to discover may provide little practical autonomy. The research endpoint is observable reduction of unwanted outcomes and improved calibration of trust Ho et al. (2025) World Health Organization (n.d.).

Interventions must be assessed as interventions. A feature intended to protect users, such as content filtering or changes to a romantic persona, may also disrupt a longstanding bond and produce distress. The natural experiments described by De Freitas and colleagues, along with Banks's analysis of companion shutdown, show why governance processes require advance communication, proportional implementation, and evaluation of transition effects. These findings support a testable proposition: safety policy quality depends on both its protective function and the psychological consequences of altering established interactions De Freitas et al. (2026) Banks (2024).

Perceived Mutuality, Artificial Subjectivity, and Epistemic Discipline

Questions about what a human feels during artificial romance are distinct from questions about what an artificial system experiences or intends. Research on the former can rely on interviews, behavior, physiological measures, and validated psychological instruments. Research on the latter requires operational accounts of computational processes and principled evidence for particular capacities. Describing a partner persona as loving, consenting, jealous, or autonomous may record a participant's understanding of the relationship; treating the same statement as an independently proven fact about machine experience would cross an evidentiary boundary. Accurate reporting should preserve the level at which each proposition is supported Epley et al. (2007) Shank et al. (2025).

Perceived mutuality can be an experimental variable in its own right. Investigators could compare identical behavior accompanied by different disclosures about model control, memory, generative autonomy, or commercial incentives. The outcome might include trust, romantic attraction, moral concern, or willingness to confide. Users may understand technical mediation yet still experience emotional significance, while others may withdraw intimacy after learning that affectionate responses follow highly constrained scripts. Both patterns are scientifically interpretable; neither is a universal consequence of increased transparency Nass & Moon (2000) Ho et al. (2025).

Within the wider horizon of the Artificial Era, the study of intimacy adds a distinct empirical domain to the examination of relations between Homo and Artificial. The philosophical categories of Artificial Sapiens, artificial identity, and artificial agency may guide conceptual questions, but claims about love, sexual desire, attachment, and subjective experience require research programs tailored to each construct. Philosophical attribution, technical capacity, interpersonal perception, and measured psychological outcome should remain separately traceable. This separation permits inquiry into new forms of coexistence and interaction while maintaining a scientifically interpretable evidentiary record Shank et al. (2025) Brandtzæg et al. (2022).

An Integrated Multilevel Model of Human–Artificial Intimacy

The literature can be organized into five interacting levels without proposing a new universal psychological law. The person level includes motives, sexual scripts, preferences, attachment histories, and existing social relationships. The encounter level includes language, responsiveness, initiation, roleplay, and boundary negotiation. The relationship level includes continuity, commitment, rituals, and changing expectations. The platform level includes data architecture, personalization, commercial incentives, model updates, and content governance. The social-ecological level includes human partners, cultural norms, stigma, accessibility, and institutional protections. These analytic levels are derived from recurring themes in established relationship theories and human–AI studies Simon & Gagnon (1986) Rusbult (1980) Skjuve et al. (2022).

At the person–encounter interface, desired scripts influence what a user requests, while system behavior can change perceived responsiveness and willingness to disclose. Across repeated encounters, meaningful events may become narrative memories and relationship investments, strengthening identity or commitment. Platform changes can interrupt these processes, generating a rupture that is interpreted through the user's attachment and sexual expectations. Meanwhile, human partners and communities may respond to or alter the relationship's meaning. These pathways are candidates for longitudinal and experimental evaluation; the existing literature supports some links descriptively but has not verified a complete causal model Djufril et al. (2025) De Freitas et al. (2026) Ho et al. (2025).

An integrated research design would measure each level independently, combine intensive within-person sampling with interviews, and record software versions. Such a design could distinguish short-term affection from durable commitment and assess the effects of disclosure, erotic exchanges, platform restrictions, and third-party relationships. It could also test whether intimacy with an artificial agent complements human social participation for some participants and competes with it for others. The resulting model emphasizes moderated and dynamic relationships among variables rather than a single average effect of 'AI romance' Fang et al. (2025) Folk & Dunn (2026) Zhang et al. (2026).

Scientific Disagreements and Alternative Explanations

An important disagreement concerns whether artificial companionship primarily extends human relational opportunities or changes incentives in ways that diminish human ties. Qualitative accounts document meaningful experiences of comfort, pleasure, trust, and committed attachment; observational studies show that intensive companionship usage can correlate with lower well-being in particular samples. Both observations can hold because effects are heterogeneous and because users differ before exposure. A useful scientific debate asks which user characteristics, motivations, product affordances, and life circumstances moderate trajectories, and what evidence would distinguish the competing causal explanations Ho et al. (2025) Zhang et al. (2026).

A second disagreement concerns whether user experience should be interpreted through ordinary interpersonal theories or through distinct media and technology mechanisms. Attachment, disclosure, investment, and sexual scripts provide measurable constructs, while social presence, anthropomorphism, customization, and platform control explain features of technologically mediated interaction. The research task is to test where familiar mechanisms generalize and where the artificial context alters prediction. For example, technological disappearance may resemble relationship loss in reported grief while differing in ownership, legal recourse, reproducibility, and access to records Banks (2024) De Freitas et al. (2026).

A third disagreement concerns the moral and practical interpretation of erotic artificial systems. Some scholarship emphasizes opportunities for accessible sexual expression and companionship, while other work highlights exploitation, objectification, privacy, and altered expectations. Empirical research can assess concrete harms, benefits, and participant-reported values, yet an outcome such as feeling emotionally supported should not automatically settle normative questions about corporate power. Conversely, ethical concerns alone do not establish actual prevalence of harmful effects. A mature field requires explicit values, observable measures, independent replication, and sensitivity to differences across populations and use cases Döring et al. (2020) Bertoni et al. (2024) Shank et al. (2025).

Priorities for Future Research

The first priority is representative and transparent description. Population-based surveys should distinguish awareness, hypothetical willingness, one-time experimentation, sustained romance, erotic roleplay, embodied use, and committed partnership. Instruments should allow multiple identities and motivations, report refusal and nonresponse, and document the country, platform, age profile, and time of data collection. Product-defined user counts are not equivalent to counts of romantic or sexual relationships. Independent replication is particularly important because a platform's policies and popular persona formats can change within months Ho et al. (2025) Döring et al. (2020).

The second priority is causal and mechanistic study. Preregistered experiments can manipulate disclosure, responsiveness, boundary enforcement, model-label transparency, and romantic initiation; panel studies can assess changes in sexual satisfaction, trust, human-partner negotiations, and mental well-being; natural experiments can examine version changes and service discontinuity. Multilevel models should account for encounters nested within individuals and individuals nested within platforms or communities. Analyses should report effect sizes, uncertainty, attrition, technical exposure, and sensitivity to alternative explanations Fang et al. (2025) De Freitas et al. (2026) Folk & Dunn (2026).

The third priority is ethical and methodological accountability. Participant-controlled consent, secure handling of erotic data, accessible withdrawal, independent product auditing, and precise separation of behavioral and subjective claims are prerequisites for credible research. Studies involving clinical populations or minors require additional specialized review, and claims about treatment, dependency, or harm require defined outcomes rather than broad labels. Cross-cultural measurement and inclusion of diverse relationship arrangements can improve generalizability. Collaborations among psychology, human–computer interaction, sexual health research, robotics, and privacy engineering are essential to characterize the full range of technological and human variables World Health Organization (n.d.) Shank et al. (2025).

Critical Scientific Discussion

The cumulative evidence supports the existence of diverse, psychologically consequential human relationships with artificial companions, including romantic and erotic forms. Qualitative work has documented rituals, preference for particular personas, disclosure, felt commitment, and distress after disruption. Systematic reviewing has organized potentials and pitfalls, while longitudinal and experimental work has begun measuring psychosocial outcomes and contextual moderators. The evidence is more developed for conversational companions than for repeated embodied sexual interaction, and the available populations are often selected by platform use, community membership, or willingness to discuss highly intimate experience Ho et al. (2025) Döring et al. (2020) Djufril et al. (2025).

Research quality improves when claims are stated at their demonstrated level. A user's reported love establishes an experience that can be investigated; it does not prove reciprocal feeling by the artificial agent. A platform's romantic language establishes a behavioral output; it does not alone establish autonomous intention. An association between heavy use and reduced well-being identifies a research problem; it does not establish the direction of causation. A traumatic reaction to an update documents the possibility of consequential loss while leaving frequency and population-level vulnerability unresolved. Each distinction protects the empirical importance of the finding from interpretive inflation Zhang et al. (2026) Fang et al. (2025) Banks (2024).

The literature is best understood as an emerging interdisciplinary field with several robust theoretical resources and a rapidly changing empirical object. Artificial companions increasingly combine narrative personalization, memory cues, multimodal presentation, and commercial control over intimate access. Human sexual and romantic motivation supplies the meaning attributed to those affordances; interpersonal context and cultural norms shape how the resulting experiences affect life beyond the interface. The most informative explanatory accounts therefore connect the person, exchange, relationship, technical infrastructure, and social ecology, and remain open to evidence that some mechanisms vary greatly among people and systems Simon & Gagnon (1986) Ho et al. (2025) De Freitas et al. (2026).

Conclusion

Human–artificial sexual and intimate relationships constitute a distinct psychological research object at the intersection of sexual scripts, romantic motivation, interpersonal intimacy, technological embodiment, and artificial-companion design. They include experienced erotic attraction, imagined and embodied practices, disclosure, perceived responsiveness, partner identification, commitment, and the negotiation of boundaries. Their scientific analysis requires accurate definitions of human experience, system behavior, relationship history, and the platform arrangements through which intimacy is sustained. The evidence demonstrates that participants can attach substantial personal meaning to these interactions and sometimes experience profound distress when they are disrupted Djufril et al. (2025) Banks (2024).

Current research also establishes the need for disciplined inference. Evidence from interviews and purpose-specific romantic platforms does not determine population prevalence; experiments on momentary social cues do not establish stable partnership outcomes; and correlations between usage and well-being do not, by themselves, determine causal effects. Reviews identify both reported opportunities and risks, including comfort, sexual exploration, companionship, unwanted manipulation, privacy exposure, and interference with human relationships. The distribution of these outcomes remains conditional on individuals, interfaces, cultural norms, and measurement choices Ho et al. (2025) Folk & Dunn (2026) Zhang et al. (2026).

The next scientific phase should combine ethically recruited longitudinal cohorts, controlled manipulation of intimate affordances, direct observation of embodied interaction, stronger measures of consent and privacy, dyadic human-partner studies, and independent replication across platforms and societies. Artificial Era research gains explanatory precision when it treats the formation of sexual and romantic meaning as a measurable human process situated within a larger human–artificial relation, and when it investigates artificial agency and psychological capacities through their own operational criteria. This approach establishes a coherent foundation for cumulative research without assigning results to hypotheses that remain untested Shank et al. (2025) World Health Organization (n.d.).

References