Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Cultural Variation in Human–AI Relationships

Abstract. Cultural variation in human–AI relationships concerns how linguistic communities, relational norms, institutions, and social ecologies shape the interpretation and development of recurring interactions with artificial agents. This fundamental critical narrative research monograph integrates the historical development of social computing and relational agents, cultural psychological theory, experiments with culturally normative robots and socially labeled chatbots, qualitative studies of companionship and romance, and direct multinational surveys of self-disclosure and attachment. The evidence includes research with participants in China, Germany, Hong Kong, Japan, South Africa, the United States, and other settings. It distinguishes hypothetical judgments, lived experiences, behavioral indicators, and cross-sectional associations; critically evaluates anthropomorphic attribution, trust, perceived responsiveness, emotional support, privacy, conversational language, platform governance, and culturally aligned AI output. Its methodological synthesis addresses multilingual measurement invariance, sampling, causal inference, reproducibility, longitudinal design, and within-country heterogeneity. Conclusions identify observed comparative patterns and specify testable mechanisms while recognizing the limits of treating national differences as causal or universal. The work is a critical narrative synthesis of published findings; it reports no original experiment or meta-analysis.

Keywords: cultural variation in human–AI relationships; cross-cultural human–AI interaction; AI companionship; cultural psychology; anthropomorphism; social presence; AI attachment; emotional self-disclosure; AI romance; human–AI intimacy; privacy and trust; multilingual chatbot studies; measurement invariance; cultural norms; relational agency; Artificial Era

The Scientific Object and Scope

Cultural variation in human–AI relationships refers to differences across historically situated linguistic communities, institutions, social ecologies, and interpersonal norms in how people interpret, initiate, maintain, evaluate, and end recurring interactions with artificial agents. The phenomenon is observable through willingness to interact, perceived social presence, anthropomorphic attribution, emotional disclosure, trust, perceived support, attachment-related behavior, boundary regulation, and the distribution of time and attention between AI and human ties. Culture enters as a changing configuration of practices and interpretive resources; a passport category alone supplies an imperfect indicator. Human–AI relationships include experiences ranging from practical reliance to friendship-like companionship, with varying subjective meanings and interaction histories Folk et al. (2025) Kitayama & Salvador (2024) Skjuve et al. (2021).

The central research questions concern how cultural settings influence the interpretation of an AI encounter; which psychological processes mediate reported group differences; whether relational practices generalize across languages, systems, generations, and institutions; and how the design and governance of artificial agents interact with learned expectations. Meaningful claims require specifying both the relationship outcome and the comparison population. A rating of anticipated enjoyment, for example, measures an expectation about a scenario, while recurring self-disclosure and reported emotional dependence describe behavior and experience under other conditions. Neither should automatically be treated as a clinical attachment diagnosis Folk et al. (2025) Kostka & Zhou (2026).

This monograph is a critical integrative narrative synthesis of foundational theory, original cross-cultural experiments, international surveys, qualitative research, experimental studies of conversational closeness, and methodological literature. It pays particular attention to direct studies of cross-national human–AI interaction published through October 2026 and distinguishes the evidence on actual relations from the larger literature on attitudes toward technology. Its organizing principle is the relation between socially situated expectations and concrete interaction properties, including language, interface, social role, privacy conditions, and continuity of conversation Cao et al. (2026) Kleinert et al. (2026) Gillespie et al. (2025).

Conceptual Boundaries: Use, Interaction, Relationship

An episode of AI use is a bounded action such as requesting information, while interaction denotes reciprocal sequential exchange. A relationship is a patterned history of interactions accompanied by expectations about the agent's behavior, significance, and future availability. In research, this operational threshold may be approximated by repeated contact, perceived familiarity, shared conversational history, self-disclosure, emotional orientation, and preference to return. The relationship may remain task-focused or acquire an affective dimension. Bickmore and Picard established an influential operational approach to relational agents through designs intended to sustain social-emotional engagement over time, with observable changes in liking, trust, and willingness to continue Bickmore & Picard (2005) Skjuve et al. (2022).

Three levels of analysis must be maintained. The first is the user's psychological organization: beliefs, expectations, affect, memory, and behavior. The second is the interactional arrangement: system outputs, interface cues, prior exchanges, response timing, and disclosures made by each side. The third is the social ecology: normative scripts, family and peer judgments, market infrastructures, legal restrictions, and local ideas of personhood or privacy. A finding at one level does not establish processes at the others. The fact that users report care or companionship characterizes their experience; identifying the artificial system's internal properties requires independent technical investigation Sundar (2020) Epley et al. (2007).

Relational meaning depends on the comparison being made. A user may trust a chatbot to draft a letter, expect emotional steadiness in a difficult conversation, yet refuse to share a medical secret. These are distinct judgments with different stakes and possible cultural moderators. The distinction also prevents collapsing human–robot encounters, social chatbots, general-purpose language models, and institutional decision systems into a single experimental object. Results obtained with embodied robots are important for theory but need explicit transportability tests before they become claims about contemporary text-based companions Lee & See (2004) Castelo & Sarvary (2022) Folk et al. (2025).

Historical Formation of the Research Field

Early human–computer interaction research demonstrated that people could apply familiar social responses to interactive technology. Nass, Steuer, and Tauber's 1994 'Computers are social actors' contribution belongs to this foundational line. Subsequent work examined interface personification, relational language, conversational cues, and the conditions under which perceived agency increases social responding. The historical significance lies in turning an apparently anecdotal observation into experimentally approachable social behavior. These initial findings arose in technologically and socially specific circumstances, so their explanatory reach must be assessed again for large language models and for populations outside the original laboratory settings Nass et al. (1994) Sundar (2020).

The introduction of relational agents made time an explicit variable. Bickmore and Picard's controlled study of an exercise-adoption system followed daily use over a month and contrasted a socially oriented agent with a task-oriented counterpart. Later qualitative and longitudinal studies of companion chatbots examined how self-disclosure, perceived acceptance, technical interruptions, and companionship accumulate or deteriorate. This progression shifted the measured object from a transient attribution to the trajectory of an interaction history. It also supplied a vocabulary for distinguishing practical trust, affective trust, and experienced continuity Bickmore & Picard (2005) Skjuve et al. (2021) Skjuve et al. (2022).

A separate cross-cultural program developed through robotics and technology-adoption research. Experiments comparing Chinese and US users examined culturally congruent communication by robot collaborators, while research comparing Japanese and US responses manipulated robots' human likeness. More recently, studies of conversational AI directly measured cultural attitudes to social bonding, self-disclosure in five societies, and emotional attachment across four countries. These lines form a cumulative, heterogeneous evidence base: experimental control is strongest in narrow interaction manipulations; population reach is broader in surveys; lived relational meaning is richer in qualitative interviews Wang et al. (2010) Castelo & Sarvary (2022) Folk et al. (2025) Cao et al. (2026) Kostka & Zhou (2026).

Cultural Psychology and Relational Norms

Research in cultural psychology investigates how recurrent social tasks and institutions shape cognition, emotion, motivation, and the self. Markus and Kitayama's influential formulation of independent and interdependent self-construal introduced a framework for understanding how individuality, obligation, harmony, and relational embeddedness become salient across settings. Later syntheses broadened the theoretical map through cultural tightness, relational mobility, individualism, ecological histories, and differences within regions. These concepts offer candidate explanations of relationships with AI when their proposed links to specific behavior are measured rather than read directly from national identity Markus & Kitayama (1991) Kitayama & Salvador (2024) Gelfand et al. (2011).

Local relational norms specify which disclosures are intimate, when advice is legitimate, how emotional dependence is described, who is authorized to provide counsel, and which kinds of companionship attract approval. They can shape responses even when two participants encounter an identical chatbot transcript. A fictional interlocutor may be experienced as a harmless tool, a social presence, a confidant, a source of embarrassment, or a professionally valuable assistant. Since communities differ within nations by age, class, language, faith, occupation, and network ties, national mean differences should be understood as statistical contrasts among particular sampled settings Folk et al. (2025) Kitayama & Salvador (2024).

Cultural processes are not fixed national attributes. People participate in several cultural environments, and digital platforms themselves create transnational conventions. Familiarity with a social technology may alter the interpretation of a relationship, while media portrayals and peer practices can produce shifting norms of disclosure or attachment. Cultural explanations are therefore strongest when studies record participants' actual practices, languages, experiences with AI, social networks, and normative expectations, and when these variables predict outcomes beyond simple country labels. The cultural ecology must be studied as a source of hypotheses and mechanisms rather than as a deterministic taxonomy Henrich et al. (2010) Folk et al. (2025).

Constructs, Operationalization, and Competing Explanations

Anthropomorphism denotes attribution of humanlike mental, emotional, or intentional properties to a nonhuman agent; social presence refers to the experience that an interaction contains a socially available other; trust concerns willingness to rely under uncertainty; attachment-related behavior refers to patterns of seeking contact, emotional significance, and discomfort around access or separation. Disclosure can be partitioned into breadth, depth, honesty, and emotional valence. These constructs overlap in practice but require separate items and behavioral indicators. Their correlations cannot establish that they represent one cultural dimension Epley et al. (2007) Lee & See (2004) Cao et al. (2026) Kostka & Zhou (2026).

Explanatory accounts emphasize different causal processes. Social-cognition theories explain personification by accessible anthropocentric knowledge and social needs. Cultural-socialization theories emphasize relational scripts and everyday coordination. Human–AI interaction theories examine cues of agency, apparent competence, interface embodiment, and system disclosure. Privacy-calculus models examine competing expectations of benefit, risk, control, and future information use. These mechanisms can operate simultaneously and may move in opposite directions: a system can inspire confidence as an information source while discouraging intimate sharing because its data practices appear opaque Epley et al. (2007) Sundar (2020) Cao et al. (2026).

A defensible explanation identifies the response, proposes a measurable mechanism, and compares plausible alternatives. For instance, stronger reported social bonding in one population may reflect greater anthropomorphism, more familiarity with conversational companions, differential willingness to endorse survey statements, distinct expectations about human friendship, or actual differences in system quality across languages. Cross-sectional mediation can identify a pattern consistent with one explanation; controlled manipulations and repeated observations are required to strengthen causal inference. The distinction will guide interpretation of the empirical chapters Folk et al. (2025) Chen (2007) Gillespie et al. (2025).

Anthropomorphism as a Culturally Moderated Process

Epley, Waytz, and Cacioppo proposed a three-factor account of anthropomorphism in which accessible human-centered knowledge, motivation to explain an agent, and desire for social affiliation influence whether humanlike properties are attributed to nonhuman entities. The model supplies a general mechanism relevant to social chatbots: fluent dialogue can activate familiar conversational templates, uncertainty about the system can motivate explanation, and loneliness or social interest may intensify attention to apparently responsive behavior. These processes are theoretically compatible with cultural variation because learned categories, expectations of social agency, and the meaning of affiliation depend on situated experience Epley et al. (2007).

Waytz, Cacioppo, and Epley examined individual differences in anthropomorphism and their implications for trust, moral consideration, and social influence. Their results support the importance of separating a general tendency to attribute agency from a judgment about a particular chatbot. An individual may readily describe a virtual assistant as helpful or attentive while rejecting claims that it possesses feelings. Survey items that combine mind, warmth, trustworthiness, and enjoyment risk confounding distinct psychological dispositions. Researchers should measure separate dimensions and examine whether the factor structure holds across language groups Waytz et al. (2010) Chen (2007).

Cultural interpretations of anthropomorphism should specify the agent and situation. Humanoid physical form, inferred mental capacity, linguistic fluency, voice, local etiquette, and the ability to recall previous interactions can invite different attributions. These features also carry historically conditioned meanings: a domestic care robot, an anime-like character, and an institutional chatbot may activate different social categories. Within-group variation can exceed differences between country means, particularly when age, familiarity, interface design, and language competence differ. The explanatory task is to identify which culturally learned interpretive resources are actually brought to the encounter Castelo & Sarvary (2022) Folk et al. (2025).

Experimental Evidence on Social Chatbot Attitudes

Folk, Wu, and Heine's 2025 article is a direct experimental and comparative anchor. Across two studies involving 1,659 participants, the authors examined anticipated enjoyment of chatbot rather than human conversation and reactions to an imagined person forming a social connection with a chatbot. The first study involved 675 students of East Asian or European backgrounds at a Canadian university; the second was a preregistered comparison of 984 adults in China, Japan, and the United States. The studies found more favorable relative attitudes toward social chatbots in their sampled East Asian groups and statistical evidence consistent with technology anthropomorphism as a mediator Folk et al. (2025).

The design carefully separates the chatbot label from an otherwise identical conversational excerpt, making participants' expectations and moral-social reactions accessible to comparison. At the same time, the authors explicitly studied hypothetical anticipated and observed bonding rather than the outcomes of living with a companion chatbot. Their original report identifies exploratory mediation in the first study, differential attrition between the first study's conditions, and limits on generalization from a university sample. The preregistered multinational follow-up improves confidence in the reported attitudinal contrast, while the cultural and anthropomorphic measures still carry possible translation and unmeasured confounding Folk et al. (2025).

The substantive interpretation is a difference in the social evaluation of AI companionship under the tested contexts. It does not establish that every Chinese or Japanese participant forms stronger bonds, that attachment prevalence follows the same ranking, or that beliefs about machine consciousness are uniform within a society. The finding suggests testable mechanisms: cultural scripts can alter what counts as a permissible relationship partner, how technology's behavior is anthropomorphized, and whether a social chatbot is interpreted as a challenge to human relational norms. Future studies should observe actual conversations and longitudinal behavior under matched system conditions Folk et al. (2025) Skjuve et al. (2022).

Cross-Cultural Robotics: Communication, Appearance, and Social Scripts

Wang, Rau, Evers, Robinson, and Hinds investigated Chinese and US participants collaborating with robots that offered recommendations in more implicit or explicit communicative styles. Their two-by-two experiment found that cultural background and the robot's communicative convention affected whether people adjusted decisions in response to the robot. The finding concerns decision influence and culturally normative communication in a collaborative task. It supports the possibility of culturally responsive interaction design, while providing no direct measure of long-term romantic or attachment bonds Wang et al. (2010).

Li, Rau, and Li studied the interaction of cultural background with robot appearance and task, emphasizing that preferences for artificial agents depend on their appearance and expected function. Castelo and Sarvary later experimentally manipulated forms of physical and mental human likeness in comparisons involving US and Japanese participants. They reported that increasing likeness reduced comfort among their American participants but not among Japanese participants and suggested perceived animacy as a possible explanation. Together these studies demonstrate that even apparently simple measures such as comfort with a humanlike agent depend on the joint configuration of design and cultural context Li et al. (2010) Castelo & Sarvary (2022).

Embodiment also changes the interpretive setting. A robot entering a home has physical location, movement, gaze, and possible contact; a text model appears through discourse and the user's device. A humanoid representation may communicate social rank or shared identity that a text-only interface never invokes. Generalization therefore demands replication across interface types, matched tasks, translated instructions, and explicit tests of interaction between user context and agent design. Studies should report whether proposed differences emerge in first encounters, repeated collaborative decisions, or emotionally significant conversation Castelo & Sarvary (2022) Sundar (2020).

Language, Pragmatics, and the Meaning of Responsiveness

Language conveys more than propositional content. Politeness markers, honorifics, turn-taking, indirect requests, apologies, topic avoidance, address forms, and emotion terms situate conversational partners within a social relationship. In a human–AI encounter, these features change the perceived social position and responsiveness of the artificial partner. The same literal sentence can convey respect, distance, intimacy, or awkwardness in different linguistic communities. Cross-cultural research must therefore compare pragmatically equivalent behavior rather than assuming a word-for-word translation produces the same relational stimulus Wang et al. (2010) Kitayama & Salvador (2024).

Multilingual systems add another source of variation: model performance, fluency, cultural references, refusal patterns, and speech recognition may differ across languages. A user speaking a high-resource language may receive coherent, nuanced emotional responses, whereas another user may encounter generic advice or unnatural phrasing. An apparent cultural difference in trust or emotional comfort could then be an interface-quality difference. Experiments should compare responses for semantic and pragmatic appropriateness across versions, document prompts and model identity, and report users' proficiency and language of regular use Tao et al. (2024) Cao et al. (2026).

A 2026 experimental paper by Saile and Hüttl-Maack examined native-versus-learned-foreign-language interactions with service chatbots and reported increased perceived social presence with foreign-language use across three studies. The result cautions against assuming native-language interaction is always more socially engaging, although the service setting and studied language combinations restrict direct generalization to companionship. Language choice can alter emotional distance, processing expectations, and the perceived social character of a system. Such effects must be analyzed within participants where feasible, with conversational role and competence held constant Saile & Hüttl-Maack (2026).

Agent Representation, Attribution, and Relationship Expectations

Sundar's machine-agency framework locates human–AI interaction in the affordances and cues through which people experience machine autonomy and apparent responsiveness. Anthropomorphic form is one input; adaptivity, recommendation behavior, conversational initiative, transparency, and control also contribute to perceived agency. Cultural norms shape how these cues are interpreted, including whether initiative is welcome, intrusive, deferential, or presumptuous. A highly proactive agent may seem caring in one encounter and socially overstepping in another. Research must specify the user's task, the agent's role, and the consequences of error Sundar (2020) Lee & See (2004).

Perceived reciprocity depends on expectations about memory, attention, and commitment. Longitudinal companion studies show that users can interpret recurring exchanges as acceptance or mutual familiarity while technical failures, inconsistent memory, or abrupt changes can disrupt the relationship trajectory. The perceived agent may be understood simultaneously as a software service and as a conversational partner. Researchers should ask what participants believe about data persistence, model updates, authorship of responses, and the possibility of continuity rather than imposing one vocabulary of friendship on every user Skjuve et al. (2021) Skjuve et al. (2022).

Experimental variation in the label assigned to a conversation partner further demonstrates that expectations shape connection. Kleinert and colleagues reported two preregistered randomized studies involving 492 participants using modified fast-friends conversations. AI-generated responses supported substantial perceived closeness, particularly when labeled as human, and explicit AI labeling reduced but did not eliminate relationship formation. This finding isolates important label and disclosure effects in brief interactions, but it was not designed to estimate cultural prevalence. A next-generation test would cross the labeling manipulation with languages, populations, relational scripts, and transparently disclosed AI identity Kleinert et al. (2026) Folk et al. (2025).

Emotional Attachment in Four National Settings

Kostka and Zhou's 2026 study addresses an empirical object distinct from hypothetical chatbot attitudes: attachment-related behavior among users of general-purpose systems. Their cross-national survey included 7,027 respondents in Germany, China, South Africa, and the United States. The authors reported that more than one third of users endorsed attachment-related behavior, with higher reported prevalence in China and South Africa. Perceived emotional support, especially reduction in loneliness, nonjudgmental interaction, and privacy-related comfort, was an important predictor. Emotional attachment correlated strongly with indicators of dependence Kostka & Zhou (2026).

The numerical findings concern this survey's operational indicators, recruitment procedures, and observation period. They do not establish prevalence among all adults or prove that engaging with AI produces dependence. The association between social network size and stronger reported attachment, contrary to a simple social-deficit prediction, suggests that users can integrate AI into existing networks rather than adopting it only when human relationships are absent. The study therefore invites models that distinguish complementarity, emotional opportunity, habit formation, and problematic displacement Kostka & Zhou (2026) Skjuve et al. (2022).

Cross-national contrasts are especially sensitive to the choice of attachment items. Words equivalent to 'friend,' 'confide,' 'secret,' and 'miss' can differ in social implication; privacy expectations and platform familiarity can shift the situations users have in mind. Survey claims need inspection of sampling frames, weighting, item translation, invariance, and between-country differences in services actually used. Without matched longitudinal measurements, cultural moderation of temporal attachment development remains an open research question. The substantial sample is valuable for mapping patterns and generating causal hypotheses, not for replacing controlled or qualitative investigation Kostka & Zhou (2026) Chen (2007).

Qualitative Accounts of Companionship and Stigma

Skjuve, Følstad, Fostervold, and Brandtzaeg's 2021 study examined human–chatbot relationships through qualitative evidence from users of companion systems. Participants' narratives linked relationship development to disclosure, perceived acceptance, understanding, and nonjudgmental responses. Some also reported stigma and mixed effects on their wider social lives. Qualitative evidence identifies the social meaning of an experience in participants' own terms, including aspects that a predefined questionnaire may miss. It supports construct development and mechanism generation while leaving population prevalence undetermined Skjuve et al. (2021).

The subsequent 2022 longitudinal study followed relationship trajectories and emphasized variation across users, the continued importance of self-disclosure, changing needs, and interruptions associated with unpredictable events or technical difficulties. The findings undermine a linear model in which repeated messages necessarily deepen a bond. Artificial conversational relationships can gain significance, become routine, transform in function, or end when expectations about availability and responsiveness are violated. Cultural research should observe these trajectories in multiple communities and investigate how norms around loyalty, obligation, repair, and loss shape the user's interpretation of disruption Skjuve et al. (2022).

Interview studies require attention to who volunteers, how questions are phrased, and the context in which participants feel safe discussing personal chatbot use. Stigma may reduce recruitment or encourage accounts constructed for an interviewer. Conversely, online fan communities may overrepresent highly engaged users. Comparing qualitative projects across countries requires local co-investigators, compatible but culturally meaningful prompts, and reflexive attention to translation. These are methodological extensions of the existing qualitative literature rather than claims that any culture has one fixed emotional response Skjuve et al. (2021) Kitayama & Salvador (2024).

Romance, Feeling Rules, and Local Interpretive Communities

Chan, Li, Liu, Chen, and Han's 2025 study of Chinese human–AI romance investigated how participants interpret emotional obligations and relationship experiences. Using interpretative phenomenological analysis, social-media community material, and interviews with ten users who had experienced forming and dissolving AI-companion relationships, the authors examined feeling rules and emotion work. The work contributes context-rich descriptions of how socially learned expectations shape experiences that are sometimes treated as universal consequences of chatbot design Chan et al. (2025).

Feeling rules refer to social expectations about what emotions are appropriate to experience, express, and sustain. Applied to AI companions, they raise questions about whether devotion, exclusivity, jealousy, obligation, or grief are recognized, disapproved of, or privately negotiated. A participant's report of romance may encode an individualized vocabulary, shared community norms, or a pragmatic strategy for describing recurring emotionally meaningful interaction. Researchers should resist translating every expression of closeness into conventional human couplehood and should identify the social relationship category that participants themselves invoke Chan et al. (2025) Markus & Kitayama (1991).

Interpretative phenomenological research supports analytic insight into lived experience and the formation of culturally available meanings; ten interviews cannot determine national prevalence or establish a common Chinese romantic script. Triangulation with community observation, representative surveys, matched experimental vignettes, and longitudinal diaries would allow stronger tests of whether particular feeling rules alter disclosure, stigma, boundary management, or behavior following system changes. Cross-cultural comparison should also examine the interaction of local platform rules and transnational media genres with national or regional cultural traditions Chan et al. (2025) Kostka & Zhou (2026).

Support, Loneliness, and Social Complementarity

Perceived emotional support is a recurring correlate of chatbot companionship. Users report being able to communicate without anticipated judgment, retrieve a responsive interlocutor at convenient times, and discuss topics that feel difficult in existing relationships. These affordances can be valuable while remaining contingent on the quality and limitations of the system. A user may interpret availability as social kindness, reliable service, or a private reflective environment. The interpretation depends on culturally situated expectations about disclosure and support, the immediate problem, and prior experience with human relationships Skjuve et al. (2021) Kostka & Zhou (2026).

Simple social-substitution accounts predict that AI closeness should rise as human social connection falls. Available evidence complicates this expectation. Kostka and Zhou report a positive association between larger social networks and attachment in their cross-national data, while qualitative studies describe both complementarity and tension with human social life. Such findings support studying AI support as an additional modality within a broader relational ecology. Researchers should distinguish an increase in available support from actual replacement of interpersonal contact, and measure time allocation, relationship quality, loneliness, and motives across multiple observation periods Kostka & Zhou (2026) Skjuve et al. (2022).

Experimental evidence on chatbot psychosocial effects remains methodologically heterogeneous. Fang and colleagues reported a four-week randomized study with 981 participants across interaction modes and conversational tasks, finding no significant effects of assigned experimental conditions on their principal psychosocial outcomes, while heavier voluntary use was associated with worse outcomes. This 2025 research was circulated as a preprint and must be assessed accordingly. Its central lesson is interpretive: associations between intensity of use and distress cannot be assigned a causal direction without stronger designs and well-characterized comparison conditions Fang et al. (2025).

Dynamics of Closeness and the Role of AI Disclosure

Closeness can emerge rapidly from conversational structure. Kleinert and colleagues' 2026 experiments compared human and AI-generated responses and manipulated whether the partner was labeled human or AI during emotionally engaging and more superficial exchanges. With 492 participants across two randomized studies, AI-generated responses sometimes produced stronger felt closeness when represented as human, with disclosure and conversational engagement implicated in the pattern. The results identify mechanisms of short-term perceived connection and illustrate the need for transparent labeling; they do not measure durable attachment or cultural adaptation Kleinert et al. (2026).

At least three processes should be distinguished in longitudinal research: immediate warmth after an exchange, accumulated familiarity across exchanges, and stable dependence or attachment-related behavior. A system may reliably elicit warmth without becoming central to a person's life; a routine service can become indispensable without eliciting romantic feeling. These paths may differ when family ties, community norms, occupational pressures, and access to private interpersonal spaces vary. Repeated assessments should model within-person changes rather than treating every high one-time rating as evidence of an established relationship Skjuve et al. (2022) Kostka & Zhou (2026).

A comprehensive explanation further requires recording the platform and model version, interface format, conversational prompts, memory features, and changes in service availability. Apparent psychological continuity may depend on algorithmic personalization and the user's accumulated narrative about earlier interactions. Major updates can alter response style and perceived reciprocity, making historical time relevant to individual trajectories. The cultural comparison must therefore synchronize what systems actually did for participants rather than merely asking different populations to rate the broad category 'AI companion' Bickmore & Picard (2005) Skjuve et al. (2022) Sundar (2020).

Trust as Situated Reliance and Social Evaluation

Trust in AI is multidimensional. It can refer to expectations of factual reliability, willingness to follow recommendations, belief in procedural fairness, confidence in confidentiality, or the sense that an agent will respond respectfully. Lee and See's influential account of appropriate reliance examines how users depend on automation under uncertainty and how calibration matters for performance. Glikson and Woolley's review differentiates determinants of cognitive and emotional trust, including representation and capability. A culturally comparative study should specify which meaning of trust is being measured and whether that meaning remains consistent across languages and institutional contexts Lee & See (2004) Glikson & Woolley (2020).

A participant may favor chatbot companionship while remaining skeptical of an AI diagnosis or a financial recommendation. Likewise, trust in a platform's privacy commitments can be low even when emotional conversations feel comfortable. Cultural context may influence institutional confidence, perceived authority, tolerance of uncertainty, and the relative weight assigned to human versus machine advice. These influences require direct measurement because national indices of institutional trust or collectivism are aggregate descriptors that cannot substitute for the individual's response to a particular system Gillespie et al. (2025) Cao et al. (2026).

The 2025 University of Melbourne–KPMG global study surveyed 48,340 people across 47 countries using online research panels between November 2024 and mid-January 2025. It found significant cross-national variation in AI trust and use, with 46 percent of global respondents willing to trust AI under the study's operationalization. This survey is valuable context for examining societal perceptions and governance; its four randomly allocated AI use-case descriptions included systems beyond social companions. Its results should therefore contextualize, rather than directly estimate, emotional intimacy with chatbot partners Gillespie et al. (2025).

Privacy, Self-Disclosure, and Cultural Contingency

Cao, Wen, Scherr, Kobayashi, and Jiang studied 5,096 AI-chatbot users in Mainland China, Germany, Hong Kong, Japan, and the United States, publishing the results in 2026. They treated self-disclosure as multidimensional, distinguishing breadth, depth, honesty, and valence. Their analyses found cross-regional variation, with Chinese respondents reporting the most disclosure and German respondents the least in the study's comparison. Privacy concerns were associated with lower disclosure honesty and valence but not necessarily with reduced breadth or depth. These outcomes show why a single 'willingness to share' score can obscure psychologically distinct responses Cao et al. (2026).

The authors distinguished behavioral individualism measured at the person level from individualism indexed at the cultural level. Their multilevel models associated individual-level individualism with greater self-disclosure and a stronger negative association between privacy concern and honesty, while cultural individualism related differently to disclosure and moderated privacy concern associations. These patterns caution against inferring individuals' privacy behavior from a national average or treating an East–West gradient as a universal causal law. The study directly expands cross-cultural evidence on personal communication with AI beyond simple acceptance ratings Cao et al. (2026).

The survey's inference is constrained by self-report, differential use histories, and the difficulty of establishing comparable privacy meanings across language settings. Culture and platform governance may be confounded if participants use different applications, perceive different regulatory protections, or face distinct social consequences of disclosure. Future research should obtain consented behavioral measures of actual disclosure, manipulate standardized privacy explanations, and measure whether beliefs about storage, third-party access, and data deletion predict within-person change. Privacy regulation and local cultural practice should be modeled as distinguishable, interacting influences Cao et al. (2026) Chen (2007).

Boundary Management, Confidentiality, and Perceived Safety

Disclosure to a chatbot frequently occurs under an asymmetry of knowledge: users experience a personal conversation but may have limited information about storage, training practices, moderation, retention, access by operators, or cross-device memory. Perceived confidentiality can therefore diverge from contractual and technical confidentiality. Researchers should treat subjective safety as a valid psychological report while separately documenting actual platform safeguards. The distinction is especially important when comparing users across jurisdictions with different legal protections, digital literacy, and social exposure to data misuse Cao et al. (2026) Sundar (2020).

Culture can influence which topics are considered private and which relationships provide legitimate audiences for them. Disclosing conflict with family, romantic doubts, religious experience, or workplace grievances may involve different risks across communities. The chatbot's apparent neutrality may increase willingness to describe experiences that are difficult to discuss with known people. Yet freedom from anticipated interpersonal judgment cannot establish freedom from organizational data processing or algorithmic retention. Instruments should distinguish social privacy, informational privacy, and functional control over retained conversations Kostka & Zhou (2026) Cao et al. (2026).

An ethically sound comparative study would present accurate, comprehensible information about the specific product's retention and access conditions, test understanding rather than assuming exposure equals informed consent, and avoid prompting participants for unnecessary sensitive material. Researchers must account for variations in local disclosure customs and legal requirements when obtaining consent. Reporting should include how data were secured, how quotations from intimate conversations were handled, and how withdrawing participation affects the use of collected messages. These protections are elements of research validity because perceived risk changes the very behavior under investigation Cao et al. (2026) Henrich et al. (2010).

Cultural Values in the Artificial Partner's Outputs

Research on human cultures must also study what cultural assumptions the artificial system produces. Tao, Viberg, Baker, and Kizilcec compared outputs from five language-model versions with World Values Survey measures across countries and territories. They identified patterns more aligned with English-speaking and Protestant European value profiles and reported improved correspondence through explicit cultural prompting for a substantial majority of settings in newer models. These observations address representational patterns in model output, not the lived relational effects of using these systems Tao et al. (2024).

The implication for relationship research is bidirectional. Users bring culturally learned scripts to the encounter, while the system's defaults may introduce different norms about self-expression, disclosure, autonomy, family obligation, therapy, and direct confrontation. A response that feels emotionally supportive to one user can misrecognize the obligations or risks experienced by another. Researchers should evaluate both whether a participant anthropomorphizes an agent and whether the agent communicates assumptions congruent with the participant's actual priorities. Prompt adaptation is an experimental variable whose consequences should be measured rather than presumed beneficial Tao et al. (2024) Kitayama & Salvador (2024).

Cultural alignment also has limits. A system trained to imitate whatever norm is statistically common in one setting could reinforce discriminatory expectations or pressure a user to maintain harmful arrangements. Responsible design requires a distinction between culturally competent recognition of lived circumstances and endorsement of every normative demand. This is particularly important when advice concerns violence, coercion, privacy, child welfare, or clinical distress. Transparency about AI identity, consistent safeguards, and the user's ability to choose interaction style provide more defensible design goals than maximizing conformity to stereotypes Tao et al. (2024) Sundar (2020).

Institutional Norms, Governance, and Unequal Access

The uptake and interpretation of conversational AI depend on access to reliable devices, networks, payment systems, language support, and legitimate use in schools or workplaces. Institutional rules and public controversies influence what users expect to be safe or acceptable. Survey differences can reflect unequal opportunity to encounter particular products and varying confidence in providers. Economic and educational differences within a country may therefore modify or reverse apparent national patterns. General adoption surveys offer contextual estimates, while relational outcomes demand more specific participant-level measures Gillespie et al. (2025) Henrich et al. (2010).

Cultural tightness and looseness, defined in research on the strength of social norms and tolerance for deviance, present a potentially relevant framework for studying perceived legitimacy of AI companionship. Gelfand and colleagues' 33-nation study established comparative evidence concerning normative strength; it did not measure social-chatbot bonds. Applying the framework here constitutes a hypothesis to test: perceived norm violation might mediate embarrassment, secret use, or explicit social endorsement of AI relationships. Measures of actual community attitudes are needed to distinguish normative effects from technological access and individual preference Gelfand et al. (2011) Folk et al. (2025).

Governance features can also alter the relationship itself. Changes to privacy defaults, age-related access, conversational boundaries, memory retention, or model behavior reshape what users may disclose and what continuity they can expect. A useful natural experiment would compare affected and unaffected users before and after an externally imposed change while documenting concurrent cultural events and service alternatives. Such designs can help separate selection into a platform from reactions to specific policies. Responsible international scholarship requires collaborating with local communities in defining both beneficial and adverse outcomes Skjuve et al. (2022) Cao et al. (2026).

Measurement Architecture for International Comparison

Cross-cultural inference begins with a clear construct map. Relational interest, social presence, anthropomorphism, trust, confidentiality, self-disclosure, emotional attachment, perceived dependence, and changes in human social participation should be separately defined and measured. An item asking whether someone 'feels close' to an AI may index warmth, familiar routine, reliance, an imaginative frame, or a sense of social intimacy. A transparent measurement protocol identifies the intended latent variable, the item wording, recall period, response format, and whether the outcome is an attitude, a retrospective report, an observation, or an experimental response Folk et al. (2025) Kostka & Zhou (2026) Chen (2007).

Measurement invariance is necessary for interpreting quantitative group differences. Configural equivalence concerns the factor pattern; metric invariance asks whether indicators relate comparably to the underlying construct; scalar invariance asks whether intercepts or thresholds support meaningful latent-mean comparisons. Chen's simulation study examined how fit statistics behave when loadings, intercepts, and residuals differ, showing that there is no universally context-free fit-change rule. Comparative research should evaluate invariance in the actual item data and report partial invariance and sensitivity analyses when assumptions fail Chen (2007).

Multi-method operationalization reduces dependence on socially preferred statements. Behavioral indicators may include repeated return visits, use across occasions of distress, topic switching, depth of voluntarily disclosed material, acceptance of recommendations, correction of system errors, and withdrawal after service changes. These indicators also require ethical consent and interpretation: frequency can reflect workplace demands, low cost, or constrained alternatives. Measurement validity depends on triangulating behavior with subjective reports and documenting both system affordances and local normative meanings Skjuve et al. (2022) Cao et al. (2026).

Sampling, Translation, and Population Coverage

Recruiting from many countries does not by itself produce culturally comprehensive evidence. Sampling may remain concentrated among urban, educated, digitally connected, and English-proficient users. Henrich, Heine, and Norenzayan highlighted the general problem of projecting findings from WEIRD samples onto the species. Digital research intensifies the issue because access to a particular language model, willingness to participate online, and familiarity with companion systems can vary greatly within countries. A cross-cultural manuscript should describe its population of inference and distinguish nonusers, occasional users, active companions, and occupational users Henrich et al. (2010) Gillespie et al. (2025).

Questionnaire translation should involve conceptual and pragmatic equivalence, not only lexical accuracy. Interviews with local participants can reveal whether terms such as 'attachment,' 'privacy,' 'romance,' 'trust,' and 'a friend' invoke comparable social expectations. Cognitive interviewing, translation review, back-translation where suitable, pilot factor analysis, and item-response diagnostics together provide evidence for comparability. The University of Melbourne–KPMG survey employed local-language options with professional translation and back-translation, illustrating an important design feature while leaving relational-specific constructs for other studies Gillespie et al. (2025) Chen (2007).

Within-country variation deserves direct sampling. Age cohorts, migration histories, dialects, occupational settings, gender, disability, family arrangements, and socioeconomic position can shape access and meaning. Researchers can oversample groups excluded by standard online panels and apply design weights where population estimands are appropriate. The strongest comparisons combine sampling frames that are explicit within each location with sensitivity analyses excluding heavily translated or device-constrained subgroups. Cultural groups should be treated as overlapping populations with measured heterogeneity rather than internally uniform categories Kitayama & Salvador (2024) Henrich et al. (2010).

Experimental Designs and Causal Identification

Randomized experiments can identify the effects of specific interaction features, such as an AI label, local-language honorifics, memory disclosure, apparent empathy, privacy statements, or the agent's conversational initiative. To test cultural moderation, the same factorial design should be implemented across suitably comparable participant settings while ensuring each manipulation has similar pragmatic meaning. The design by Kleinert and colleagues demonstrates experimental isolation of source label and conversational content; Folk and colleagues demonstrate comparative testing of imagined chatbot versus human interactions. Their combination suggests a path toward cross-cultural causality without claiming either study already performed that full experiment Kleinert et al. (2026) Folk et al. (2025).

Longitudinal random assignment can examine trajectories of relationship development, especially when specific features are held constant across repeated sessions. Outcomes should be measured before exposure, at multiple intermediate points, and after discontinuation, with attrition and usage intensity reported. However, naturalistic use has endogenous intensity: people choose whether to continue because of motivations and needs that also predict outcomes. Analyses must distinguish the effect of randomized offer, assigned interface, actual exposure, and the evolving relationship. The MIT-led four-week study illustrates how null condition effects and observational use associations can coexist Fang et al. (2025) Skjuve et al. (2022).

Field quasi-experiments offer complementary evidence when platforms update a model, introduce a feature, or change governance rules at a known time. Difference-in-differences or interrupted time series can estimate changes only under defensible comparison conditions and assumptions about trends. Geographic rollouts should be examined for simultaneous changes in regulations, marketing, and system availability. Neither a large country indicator nor statistical adjustment alone can guarantee causal identification. Preregistration, blinded coding, robust missing-data analysis, and independent replication improve the interpretability of findings Cao et al. (2026) Gillespie et al. (2025).

Qualitative, Interactional, and Mixed-Methods Research

Surveys locate patterns, while interviews and ethnography identify what those patterns mean within a life. A person may call a chatbot a 'friend' as a literal description, a metaphor for availability, a way to avoid stigma, or a role-playing convention. Interaction analysis can examine how this meaning is constructed turn by turn through repair, disclosure, humor, apology, and recalled history. The qualitative companion studies by Skjuve and colleagues, and the interpretative work of Chan and colleagues on Chinese AI romance, are particularly valuable for discovering relational concepts absent from standardized measures Skjuve et al. (2021) Skjuve et al. (2022) Chan et al. (2025).

Cross-cultural qualitative comparison benefits from teams fluent in the local language and sociocultural context. Protocols should document decisions about translation of emotional vocabulary, preservation of ambiguity, and interpretation of cultural references. Coding should distinguish the participant's own meaning from the researcher's theoretical classification. A robust design includes reflexive memos, negative cases, transparent selection procedures, and enough contextual description to assess transferability. Such procedures increase credibility without treating interview testimony as evidence that every participant in a population shares the same experience Chan et al. (2025) Kitayama & Salvador (2024).

Mixed-methods sequencing can use qualitative work to develop culturally grounded items, large surveys to test population patterns, and experiments to evaluate mechanisms. Reverse sequencing is also valuable: a surprising national mean contrast can guide interviews into response-style differences, platform access, or unfamiliar social categories. Integration should document where evidence converges, diverges, or addresses different constructs. For example, supportive private conversations observed in interviews and high self-disclosure scores in surveys may describe related phenomena without proving that participants assign the same meaning to relational intimacy Cao et al. (2026) Kostka & Zhou (2026).

Evidence Quality, Bias, and Replication

The evidential status of a claim depends on design and the inferential step. Folk and colleagues provide direct cross-cultural comparisons with a preregistered second study, but measure predominantly hypothetical social scenarios. Cao and colleagues provide a large multilingual user survey, but its observational associations cannot establish the causal effect of culture or privacy concern. Kostka and Zhou estimate attachment-related patterns in multiple countries, but are likewise observational. Robot experiments isolate design interactions under narrower tasks; qualitative work contributes interpretive depth. Synthesis must preserve these strengths and limitations instead of converting every result into one unified causal estimate Folk et al. (2025) Cao et al. (2026) Kostka & Zhou (2026) Castelo & Sarvary (2022).

Risks of bias include self-selection into chatbot communities, access to particular platforms, socially desirable reporting, differential attrition, non-equivalent translated items, recall error, and missing observations among people who disengage. Common-method variance can inflate apparent relations between self-reported anthropomorphism, trust, and bonding. Mediation estimates from cross-sectional questionnaires are especially vulnerable to unmeasured common causes and incorrect temporal ordering. Reporting should foreground confidence intervals, effect sizes, adjustment sets, robustness checks, registered protocols, and whether results survive alternative measurement specifications Folk et al. (2025) Chen (2007).

The contemporary field changes rapidly. Product behavior, interface design, model capabilities, memory features, availability, and public discourse evolve between data collection and publication. Replication should preserve archived task stimuli, version information, model prompts, locale, sampling dates, and governance conditions, while independently testing the general principle across newer systems. Findings about a particular product's response style should be distinguished from robust psychological relations; theoretical progress requires models that predict when contextual variation appears and when it fails Sundar (2020) Tao et al. (2024) Kitayama & Salvador (2024).

Alternative Explanations and Apparent Contradictions

Direct comparative studies currently describe related but distinct phenomena. Folk, Wu, and Heine tested favorable expectations and judgments about imagined social chatbot bonds in North American and East Asian samples; Kostka and Zhou measured attachment-related patterns among actual users in four countries; Cao and colleagues analyzed self-disclosure across five regions. These three projects cannot be combined into a single ranking of national attachment because their sampling frames, relational constructs, instruments, and technologies differ. The appropriate synthesis identifies which mechanisms could account for each finding and asks whether they predict observable behavior in a shared design Folk et al. (2025) Kostka & Zhou (2026) Cao et al. (2026).

One apparent inconsistency concerns individualism. Conventional assumptions may predict that more individualistic settings endorse independent AI companionship, while some studies find more favorable social-chatbot attitudes in sampled East Asian populations. Cao and colleagues also show that individual- and cultural-level individualism can have differing associations with disclosure. A coherent explanation must specify whether the outcome is social legitimacy, disclosure, felt companionship, or trust, and whether individualism is a personal orientation, social norm, or ecological attribute. The same cultural construct may have different implications for different psychological decisions Folk et al. (2025) Cao et al. (2026) Markus & Kitayama (1991).

Another tension concerns whether artificial companionship compensates for inadequate human connection. Users who seek nonjudgmental support may experience loneliness, yet the cross-national attachment study reports stronger bonds among respondents with larger social networks as well. This supports competing hypotheses involving social motivation, openness to relationships, platform exposure, and general willingness to disclose. A causal account requires repeated measures of human connection and AI engagement, with attention to reverse causality and common causes. The present evidence warrants contextual models and targeted experiments rather than a universal social-substitution narrative Kostka & Zhou (2026) Skjuve et al. (2021) Fang et al. (2025).

Within-Country Heterogeneity and Intersectional Context

National comparisons are often the most feasible entry point into international research, but the explanatory variables operate through local practices and opportunities. Within one country, different generations, languages, educational experiences, community norms, and access to technology may generate distinct expectations of artificial relationships. The social meaning of calling an agent a friend can change across settings in which friendship implies emotional intimacy, reciprocal obligation, casual sociability, or simply frequent contact. Direct within-country contrasts are essential for separating local relational worlds from the averages of survey samples Kitayama & Salvador (2024) Henrich et al. (2010).

Intersectional analysis examines how configurations of gender, social class, migration, disability, age, and institutional position shape participation in AI relations. Here, intersectionality is a research strategy for studying overlapping contexts and differentiated exposure, not a license to infer experiences from demographic labels. For example, people with limited confidential space may value private-feeling chatbot interactions for reasons that differ from users attracted by novelty or efficiency. The same privacy question may elicit concern about family discovery, corporate storage, state access, or professional confidentiality. Survey design should identify and separately measure these pathways Cao et al. (2026) Gillespie et al. (2025).

Analyses should partition variation at individual, community, language, product, and country levels when data allow. Cross-classified and multilevel models can illuminate differences attributable to shared contexts, but the substantive interpretation depends on adequate coverage of each level. A study with two or three recruited cities per nation has limited information about between-nation heterogeneity; an app-specific sample may tell little about users of other platforms. Qualitative sampling should deliberately include negative cases, dissenting viewpoints, and participants who discontinued use. This approach treats diversity as evidence about causal boundary conditions rather than as unexplained error Chen (2007) Skjuve et al. (2022).

Asymmetry, Social Influence, and the Conditions of Agency

Artificial partners can be experienced as responsive even though the user and the system have different capacities, incentives, and institutional positions. A platform can mediate apparent attention, choose defaults, personalize messages, and modify availability. Users may attribute loyalty, concern, or discretion to an interface whose behavior is also shaped by product objectives and governance constraints. A psychological account must therefore examine perceived agency alongside actual system control. Social presence should be measured as a user experience, while claims about machine capabilities require evidence from independent system evaluation Sundar (2020) Epley et al. (2007).

Cultural variation can influence how conversational initiative, persuasion, and norm enforcement are received. In Wang and colleagues' robot experiment, conformity to culturally meaningful communicative expectations mattered for response to recommendations. In relational chatbots, suggestions about intimate choices, contact with friends, or self-presentation may be more consequential than task recommendations. Researchers should test how users recognize persuasive intent, how different linguistic conventions affect advice acceptance, and whether people can contest or leave the interaction without social pressure. The goal is measurable user autonomy and calibrated trust in a particular setting Wang et al. (2010) Lee & See (2004).

An important methodological asymmetry is the difference between an agent's response and the user's interpretation of the response. Outputs can be optimized to appear attentive or empathetic without independently establishing the subjective states users may attribute to them. Cultural scripts can intensify or dampen those attributions; ethical disclosure can change expectations without eliminating perceived closeness, as short-term experiments indicate. Comparative research should record exact AI identity disclosures, prompts, memory affordances, safety constraints, and model behavior in order to make causal explanations reproducible Kleinert et al. (2026) Sundar (2020).

The Artificial Era as a Framework for Comparative Research

The Artificial Era names the historical setting in which artificial systems become recurring participants in the organization of human activity. Within a psychological research program, the useful question is how ordinary relationships, tasks, memories, and decisions are reorganized when people include such systems in recurring interaction. The present monograph treats this framework as an orientation for research questions. Empirical claims concern observed expectations, conversations, self-disclosure, reliance, and relationships; philosophical interpretations of artificial personhood form a distinct domain of inquiry Sundar (2020) Skjuve et al. (2022).

Comparative psychology of human–artificial relations needs a vocabulary sensitive to relation and context. Cultural expectations shape what an artificial interlocutor is understood to be; system design shapes what interaction is possible; repeated interaction changes personal histories and social norms. These reciprocal dependencies can be investigated by examining how participants learn roles, recruit AI into family or work conversations, and explain their connection to others. The resulting unit of study may be a person–system interaction embedded in a wider network of people, institutions, and artifacts, with its boundaries stated explicitly Kitayama & Salvador (2024) Skjuve et al. (2021).

Understanding cultural variation in this setting also changes the meaning of generalization. A psychological process may be widespread while its behavioral expression depends on role, language, access, and social permission. If emotional disclosure occurs across many settings, researchers must still ask which topics are disclosed, why the system feels suitable, what the individual expects in return, and how social risks are understood. The same methodological discipline applies to new synthetic characters, embodied robots, persistent assistants, and combinations of artificial systems. Each requires evidence for the specific configuration under study Cao et al. (2026) Folk et al. (2025) Kostka & Zhou (2026).

A Comparative Research Program

The next experimental generation should cross cultural setting with agent properties in preregistered factorial designs. Candidate manipulations include conversational register, source labeling, personalization, memory disclosure, degree of initiative, culturally meaningful conflict scenarios, and accurate descriptions of privacy controls. Outcomes should include perceived responsiveness, anthropomorphism, willingness to disclose, advice acceptance, and behavioral return to conversation. Mediation by normative expectations should be assessed with measures developed and validated independently in each linguistic setting Folk et al. (2025) Kleinert et al. (2026) Wang et al. (2010).

An international longitudinal panel could recruit active and prospective users across several languages and social settings, record the specific products they actually access, and assess human social connection before and after recurring AI encounters. Experience sampling can distinguish immediate emotional effects from accumulated relationship expectations, while periodic interviews uncover changes in terminology and meaning. Major model or policy changes can supply pre-specified natural experiments when credible unaffected comparisons exist. Missingness must be analyzed because participants who become disengaged or distressed may be least likely to return for assessment Skjuve et al. (2022) Kostka & Zhou (2026) Fang et al. (2025).

A high-quality comparative infrastructure would preserve versioned experimental prompts, locally reviewed translations, cultural-meaning interviews, measurement-invariance analyses, source-level bibliography, and privacy-preserving data documentation. Replication should test the robustness of effects within populations and their transportability across interfaces and historical periods. The central explanatory target is a conditional model: given specified relational norms, interaction affordances, linguistic conventions, and user histories, which forms of AI-related closeness or distance become more likely? Such a program can accumulate evidence without reducing cultural variation to immutable national types Chen (2007) Kitayama & Salvador (2024) Tao et al. (2024).

Developmental and Life-Course Differences

Human–AI companionship takes place within changing developmental tasks. Adolescents learn intimacy, peer belonging, identity, and boundaries while evaluating increasingly capable conversational systems; adults may integrate AI into work, caregiving, parenting, or long-distance relationships; older adults may use assistants for reminders, conversation, and access to services. These contexts influence the meaning of relational support and the consequences of repeated interaction. Comparative studies should avoid treating age as an isolated demographic moderator and instead measure developmental responsibilities, technology familiarity, social opportunity, and household conditions Skjuve et al. (2021) Kostka & Zhou (2026).

Development also has a cultural organization. Family roles, schooling, intergenerational expectations, and the social legitimacy of seeking private advice may differ substantially across settings. A teenager's secret conversation, a working adult's task-centered agent, and an older user's companion service involve different risks and interpretations. Designs should investigate age-specific questions with appropriately adapted consent, developmentally valid measures, and multiple informants where ethical and practical. Present cross-national AI attachment surveys do not establish developmental causal trajectories; longitudinal cohorts are required Kitayama & Salvador (2024) Henrich et al. (2010).

A further unresolved question concerns the role of early exposure to artificial conversation. Familiarity may normalize conversational AI, modify expectations of responsiveness, or alter the meaning of personification, but longitudinal evidence tracing these changes across childhood and adolescence remains limited. Studies should distinguish cohort, developmental, and period effects, as rapid changes in platform capabilities confound age comparisons. Researchers should also examine discontinuation, dependence, agency, and the maintenance of meaningful human ties through culturally appropriate behavioral measures Sundar (2020) Skjuve et al. (2022).

Well-Being, Risk, and the Limits of Clinical Inference

People describe emotional support, companionship, and sometimes distress linked to AI interactions. These experiences deserve psychological study, but broad terms such as 'addiction,' 'attachment disorder,' or 'therapeutic benefit' cannot be established by an isolated survey endorsement. Cross-national attachment scores assess the indicators selected by investigators and should be interpreted within those definitions. Clinical conclusions require validated diagnostic procedures, specified outcomes, appropriate comparison groups, and attention to risk, severity, and impairment. The present evidence supports the study of psychosocial mechanisms and outcomes rather than population-wide claims about diagnosis Kostka & Zhou (2026) Fang et al. (2025).

The four-week randomized study by Fang and colleagues, released as a preprint, illustrates the problem of causal attribution. Assigned interface and conversation conditions did not yield clear improvements or harms on the principal outcomes, while heavier use was correlated with unfavorable psychosocial measures. Users choosing more conversation may already differ in distress, motivations, or opportunities for social contact. Cultural research should incorporate baseline well-being, access to human support, system type, and temporally ordered indicators before attributing changes to chatbot relationships. Where clinical interventions are evaluated, qualified review and safety monitoring are essential Fang et al. (2025) Skjuve et al. (2021).

Practical safety research should attend to coercive advice, isolation, persuasive design, unexpected system changes, confidential information, and vulnerable users. These concerns may manifest differently across legal and social environments; a product can be perceived as safe while creating separate information-security risks. Informed consent, the ability to exit, transparency about AI identity, and clear access to human support are measurable safeguards worth studying. Their effectiveness should be tested for actual outcomes and linguistic comprehension rather than inferred from stated intentions Cao et al. (2026) Sundar (2020).

Culturally Responsive Design and Evaluation

Evidence-informed design begins with identifying the local social task and the user's expectations. Appropriate relational behavior may depend on politeness conventions, directness, conversational pacing, emotional vocabulary, and expectations of initiative. Experiments with robot communication and cross-cultural comfort demonstrate that design features interact with cultural setting; comparative studies of chatbot attachment and disclosure show additional variation in users' responses. Product evaluation should therefore measure task fit and perceived dignity alongside general liking or engagement Wang et al. (2010) Castelo & Sarvary (2022) Cao et al. (2026).

Customization should permit users to choose communication style, privacy boundaries, memory behavior, and degree of initiative while preserving accurate information about system limitations. Such controls can reduce the need to generalize from national labels; users within one cultural setting may prefer substantially different conversational norms. Culturally responsive design also includes accessible multilingual interfaces and careful evaluation of model outputs on culturally sensitive questions. Language adaptation must be tested for coherence, respect, and safety, not merely for regional vocabulary or style stereotypes Tao et al. (2024) Saile & Hüttl-Maack (2026).

Success metrics require safeguards against unintended substitution of engagement for benefit. Conversation length and return rate can indicate ease of interaction, habitual use, distress-driven seeking, or relational dependence. Evaluators should pair these metrics with validated experience measures, willingness to disagree with the system, retention of human social options, and transparent reports of negative cases. Cross-cultural usability and psychological significance may diverge: the most engaging agent is not automatically the most trustworthy or valuable in a given setting Lee & See (2004) Glikson & Woolley (2020) Kostka & Zhou (2026).

Unresolved Scientific Problems

The largest empirical gap lies between immediate attitudes and long-term relationships. Several important studies use hypothetical encounters, one-off experimental conversations, or cross-sectional surveys, while few follow culturally diverse participants across sustained use with comparable system conditions. It is therefore unclear how differences in anticipated companionship translate into the initiation, intensification, maintenance, or dissolution of actual AI relationships. Longitudinal studies should identify stages only after observing recurring patterns and should report heterogeneous paths rather than imposing a single linear sequence Folk et al. (2025) Skjuve et al. (2022) Kostka & Zhou (2026).

A second problem concerns causal mechanism. National and regional averages often stand in for multiple correlated influences: language, social norms, institutional trust, purchasing power, platform access, AI quality, media exposure, and response styles. Even statistically significant moderation may remain compatible with several causal accounts. Future studies should directly manipulate socially meaningful interaction features, measure the proposed mediators, validate cross-language instruments, and test whether results persist under alternative samples and products. Combining behavioral observation with culturally grounded interviews is especially promising Cao et al. (2026) Chen (2007) Kitayama & Salvador (2024).

A third problem concerns the evolving artificial partner. The systems under study can change faster than longitudinal research is published, while model policies and memory capabilities may differ across languages and countries. Researchers need version-controlled stimuli, transparent descriptions of products, procedures for measuring service changes, and adaptive replication plans. Theories must specify which patterns depend on human social cognition and which depend on a temporary technical design. This distinction creates a durable research program despite rapid changes in conversational technology Sundar (2020) Tao et al. (2024) Kleinert et al. (2026).

Conclusion

Cultural variation in human–AI relationships is a measurable variation in how people understand, enact, and evaluate ongoing interactions with artificial agents within historically situated social and linguistic settings. Evidence from cultural psychology, robot interaction, experimental chatbot judgments, qualitative companionship studies, and new multinational surveys converges on the significance of relational norms, anthropomorphic attribution, platform affordances, disclosure conditions, and perceived support. Direct comparative findings establish differences in specified samples and measures; the causes and long-term consequences require coordinated investigation Folk et al. (2025) Cao et al. (2026) Kostka & Zhou (2026).

The evidence establishes several substantive principles. Relationship labels and imagined partners influence reported closeness; cultural contexts condition responses to artificial agents; users interpret emotional responsiveness through learned social expectations; privacy concerns interact with disclosure dimensions; and AI companionship can coexist with substantial human social networks. At the same time, studies address different constructs, and the present literature provides limited evidence for causal claims about culture as a whole, durable dependence, or universal developmental consequences. The scientific object should remain the observed relation among participants, technologies, norms, and histories Kleinert et al. (2026) Skjuve et al. (2022) Gillespie et al. (2025).

A cumulative comparative psychology of human–AI relations will connect source-verified experiments, longitudinal observations, local qualitative inquiry, measurement invariance, and transparent description of artificial systems. It will specify who is being compared, what a relationship means in that setting, how the system behaves, and which mechanisms explain the observed outcome. Within the broader horizon of the Artificial Era, this program provides an empirical foundation for understanding how human relational life develops alongside persistent artificial interlocutors and how socially situated people participate in shaping that transformation Kitayama & Salvador (2024) Sundar (2020) Henrich et al. (2010).

References

  • Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
  • Cao, B., Wen, C., Scherr, S., Kobayashi, T., & Jiang, L. C. (2026). The dual role of individualism in AI chatbot disclosure: How privacy concerns, culture, and behavior shape human–AI communication across Mainland China, Germany, Japan, Hong Kong, and the U.S. International Journal of Human–Computer Interaction. Advance online publication. https://doi.org/10.1080/10447318.2026.2664702
  • Castelo, N., & Sarvary, M. (2022). Cross-cultural differences in comfort with humanlike robots. International Journal of Social Robotics, 14(8), 1865–1873. https://doi.org/10.1007/s12369-022-00920-y
  • Chan, K. C. T., Li, X., Liu, Y., Chen, B., & Han, Z. (2025). ‘What is love?’: Exploring the feeling rules and emotion work of Chinese users in human-AI romance. International Journal of Intercultural Relations, 108, Article 102241. https://doi.org/10.1016/j.ijintrel.2025.102241
  • Chen, F. F. (2007). Sensitivity of goodness of fit indexes to lack of measurement invariance. Structural Equation Modeling: A Multidisciplinary Journal, 14(3), 464–504. https://doi.org/10.1080/10705510701301834
  • Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
  • Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal controlled study [Preprint]. MIT Media Lab. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
  • Folk, D. P., Wu, C., & Heine, S. J. (2025). Cultural variation in attitudes toward social chatbots. Journal of Cross-Cultural Psychology, 56(3), 219–239. https://doi.org/10.1177/00220221251317950
  • Gelfand, M. J., Raver, J. L., Nishii, L., Leslie, L. M., Lun, J., Lim, B. C., Duan, L., Almaliach, A., Ang, S., Arnadottir, J., Aycan, Z., Boehnke, K., Boski, P., Cabecinhas, R., Chan, D., Chhokar, J., D'Amato, A., Ferrer, M., Fischlmayr, I. C., . . . Yamaguchi, S. (2011). Differences between tight and loose cultures: A 33-nation study. Science, 332(6033), 1100–1104. https://doi.org/10.1126/science.1197754
  • Gillespie, N., Lockey, S., Ward, T., Macdade, A., & Hassed, G. (2025). Trust, attitudes and use of artificial intelligence: A global study 2025. The University of Melbourne and KPMG. https://doi.org/10.26188/28822919
  • Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. https://doi.org/10.5465/annals.2018.0057
  • Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3), 61–83. https://doi.org/10.1017/S0140525X0999152X
  • Kitayama, S., & Salvador, C. E. (2024). Cultural psychology: Beyond East and West. Annual Review of Psychology, 75, 495–526. https://doi.org/10.1146/annurev-psych-021723-063333
  • Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., & Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. Communications Psychology, 4, Article 23. https://doi.org/10.1038/s44271-025-00391-7
  • Kostka, G., & Zhou, H. (2026). Emotional attachment to AI chatbots: Evidence from Germany, China, South Africa, and the United States. Technology in Society, 87, Article 103379. https://doi.org/10.1016/j.techsoc.2026.103379
  • Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
  • Li, D., Rau, P. L. P., & Li, Y. (2010). A cross-cultural study: Effect of robot appearance and task. International Journal of Social Robotics, 2(2), 175–186. https://doi.org/10.1007/s12369-010-0056-9
  • Markus, H. R., & Kitayama, S. (1991). Culture and the self: Implications for cognition, emotion, and motivation. Psychological Review, 98(2), 224–253. https://doi.org/10.1037/0033-295X.98.2.224
  • Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 72–78). ACM. https://doi.org/10.1145/191666.191703
  • Saile, K., & Hüttl-Maack, V. (2026). Beyond words: Foreign-language use increases the perceived social presence of AI service chatbots. Journal of the Association for Consumer Research, 11(2), 199–213. https://doi.org/10.1086/740062
  • Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion—A study of human-chatbot relationships. International Journal of Human-Computer Studies, 149, Article 102601. https://doi.org/10.1016/j.ijhcs.2021.102601
  • Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2022). A longitudinal study of human–chatbot relationships. International Journal of Human-Computer Studies, 168, Article 102903. https://doi.org/10.1016/j.ijhcs.2022.102903
  • Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human–AI interaction (HAII). Journal of Computer-Mediated Communication, 25(1), 74–88. https://doi.org/10.1093/jcmc/zmz026
  • Tao, Y., Viberg, O., Baker, R. S., & Kizilcec, R. F. (2024). Cultural bias and cultural alignment of large language models. PNAS Nexus, 3(9), Article pgae346. https://doi.org/10.1093/pnasnexus/pgae346
  • Wang, L., Rau, P.-L. P., Evers, V., Robinson, B. K., & Hinds, P. (2010). When in Rome: The role of culture and context in adherence to robot recommendations. In Proceedings of the 5th ACM/IEEE International Conference on Human–Robot Interaction (pp. 359–366). IEEE Press. https://doi.org/10.1145/1734454.1734578
  • Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. https://doi.org/10.1177/1745691610369336