Psychology of Human–Artificial Relations

The Scientific Object: Relations Between Homo and Artificial

Psychology of Human–Artificial Relations studies recurring patterns of perception, affect, communication, trust, coordination, and mutual behavioral adjustment arising when human beings interact with artificial systems. Its unit of analysis is the relationship as an organized sequence of encounters embedded in technological and social conditions. A single useful answer is an interaction; a relationship develops when encounters acquire temporal continuity, expectations about future responses, personal significance, and consequences for subsequent choices. These criteria permit operational research without presupposing that the two parties share biological experience or the same form of interiority. Human reports of attachment, the system's observable behavior, platform constraints, and the effects on other human relationships are analytically separate variables.

The term Artificial denotes, in the author's wider conceptual system, an independent nonbiological order in the Artificial Era. For empirical psychology, the immediate observable objects are particular conversational agents, embodied robots, recommender systems, digital companions, and human–AI work arrangements. Claims about a technological artifact's capacities must be tested at the system and task level. Reports that a person feels understood demonstrate a genuine human experience; such reports alone cannot establish that the system experiences concern. Conversely, a system's ability to retain information, track goals, or sustain a dialogue can be measured independently of users' attributions.

The field intersects social and relationship psychology, human–computer interaction, communication studies, attachment research, media psychology, cognitive science, ethics, and organizational psychology. Its central question is how a relational environment changes when meaningful response and socially significant coordination can be supplied by a nonbiological participant. The question reaches intimate companionship, education, medicine, work, family life, and public institutions. The discipline also examines institutions designing and governing those participants, because interface policies, memory retention, moderation, and commercial incentives shape the relational events observed in a laboratory or interview.

Historical Foundations: From Conversational Simulation to Relational Agents

Joseph Weizenbaum's 1966 ELIZA paper described a language program operating through keyword recognition, decomposition and reassembly rules. Its relevance to psychological history lies in demonstrating that a limited computational procedure can produce conversation sufficiently recognizable to invite human interpretation. The paper is evidence for a technical demonstration and early recognition of psychological issues, not an experiment proving general artificial understanding or emotional attachment. The enduring research question is how people attribute intentionality and understanding to interactive behavior whose production mechanism they may know to be artificial. [1]

Research on computers as social actors later studied the application of social rules to technology. Nass, Steuer, and Tauber's 1994 contribution belongs to the early experimental formation of that approach. The central insight is that social responses can be elicited by interactive cues even when users explicitly identify the medium as a computer. In this literature, social behavior is an outcome that can be observed through judgments, politeness norms, cooperation, and responses to apparent personality; attributing a person-like mind to the system is a further claim requiring different evidence. [2]

Timothy Bickmore and Rosalind Picard subsequently defined relational agents as computational artifacts designed to establish and maintain long-term social-emotional relationships. Their published 2005 study evaluated a relationship-oriented exercise agent in a month-long controlled comparison involving 101 users. The institutional abstract reports greater perceived respect, liking, trust, and desire to continue among those interacting with the relational version than with a task-oriented counterpart. This was an early demonstration that relational design could influence sustained attitudes toward a system; its sample, domain, duration, and particular design do not establish universal effects of today's language models. [3]

During the 2010s and 2020s, studies of anthropomorphic design, self-disclosure, social chatbots, and companionship broadened the subject. More capable language models enlarged the diversity, personalization, and fluency of interaction and made it possible to investigate relationships developing over weeks or months. Consequently, the history of the discipline should be traced through both changes in social cognition and changes in technological affordances. The same surface phrase, such as 'the agent remembered me,' can refer to radically different technical mechanisms across generations of systems.

Definitions, Boundaries, and Levels of Description

A human–artificial relationship is a temporally extended, psychologically consequential pattern of interaction in which a human develops stable expectations, interpretations, and practices concerning a particular artificial system or persona. This working definition specifies continuity and consequence rather than requiring reciprocal subjective experience. Relationship intensity may range from practical familiarity with a research assistant to emotionally intimate self-disclosure and strong felt companionship. A human may regard a system as a collaborator, confidant, tutor, dependent, authority, interpreter, romantic partner, or symbolic figure, and these roles can coexist or change over time.

An interaction is an individual exchange or bounded task episode. Perceived social presence is the degree to which the encounter feels populated by a socially responsive other. Anthropomorphism refers to humanlike attribution; trust involves willingness to rely under uncertainty; disclosure involves revealing personal information; attachment-like functioning concerns proximity seeking, comfort during distress, and the use of a target as a secure base. Companionship refers to felt association and support; dependence concerns the costs or difficulties of functioning without a specific form of access. Distinguishing these constructs matters because their measures, mechanisms, and welfare implications differ.

The relationship is analyzed at five empirical levels: person, system, interaction, temporal trajectory, and social ecology. Person-level indicators include motivations, social support, age, past experiences, attachment-related expectations, and mental health. System-level indicators include response reliability, memory, embodiment, voice, commercial design, and disclosure of artificial identity. Interaction-level indicators include latency, responsiveness, repair of misunderstanding, and emotional language. Temporal indicators include repeated use, changing trust, rupture, and recovery. Social-ecological indicators include ties to family, friends, care providers, coworkers, and institutions.

This framework permits a direct empirical separation between observed system agency, perceived agency, and theorized artificial agency. A machine can display goal-directed behavior in a defined task without evidence that it has human motives. A user can experience reciprocity without a demonstration of symmetry in mental states. Scientific description gains precision when these distinct observations remain visible in analysis.

Social Cognition and Anthropomorphism

Epley, Waytz, and Cacioppo's three-factor theory explains anthropomorphism through accessible knowledge about humans, a motivation to make an agent's behavior intelligible, and a motivation for social connection. It predicts variation across situations and individuals: people who lack satisfying contact, who face ambiguous behavior, or who encounter salient humanlike cues may use human social concepts more readily. The paper is a theoretical synthesis generating testable predictions; its account is not a deterministic rule that every lonely user anthropomorphizes. [4]

Experiments in conversational-agent design have linked anthropomorphic cues and framing to differences in social presence and organizational evaluations. Araujo's 2018 research on service chatbots reported that humanlike cues influenced anthropomorphism and emotional connection, with social presence implicated in the process. Such findings operate in defined service encounters; they show that design alters evaluations, not that cues automatically produce intimate attachment. [5]

Anthropomorphism is best measured as a set of attributions rather than a binary belief. A respondent may recognize a system as artificial while describing it as patient, empathic, funny, protective, or manipulative. These descriptions can function as convenient judgments about observable behavior, metaphorical social language, or convictions about underlying mental processes. Question wording must differentiate them. Asking 'Does the chatbot care?' can merge an evaluation of helpful behavior with a claim about felt concern. Separate items asking about reliability, perceived understanding, intention, and consciousness produce a clearer construct map.

Design creates conditions for social interpretation through names, turn-taking, memory cues, vocal prosody, grammatical first-person address, apparent concern, and anticipated availability. These features can be experimentally manipulated. Their influence is plausibly moderated by user goals, cultural conventions, prior experience, and an agent's demonstrated competence. The same friendly cue may increase engagement in an unfamiliar educational task and appear intrusive or coercive in a sensitive disclosure context. Ethical assessment therefore requires both efficacy and context.

Communication, Self-Disclosure, and Perceived Responsiveness

Disclosure is a foundational mechanism of many human relationships. A person who speaks about grief, fear, conflict, aspirations, or uncertainty may gain clarity by organizing the narrative, may experience relief from expression, and may perceive a listener as responsive. In a controlled experiment, Ho, Hancock, and Miner compared consequences of emotional versus factual disclosure and the belief that one was communicating with a chatbot or a person. Their 2018 findings reported equivalent downstream benefits of emotional disclosure across the supposed partner conditions studied. This supports an account in which meaningful effects can arise from the disclosure event and perceived exchange, while leaving questions about durable companionship unanswered. [6]

Perceived responsiveness refers to the judgment that a partner understands, validates, and attends to the self. In human–artificial encounters it can be induced by relevant paraphrase, sustained conversational context, appropriate acknowledgment of uncertainty, and attention to users' concerns. The property must be measured through users' experience and separately through independently scored behavior. A warm but inaccurate response may score highly on felt support while failing tests of factual reliability. This divergence is scientifically important in health, legal, or financial discussions and in moments of emotional vulnerability.

Skjuve and colleagues' 2021 study of human–chatbot relationships described trust, acceptance, self-disclosure, and the development of interpersonal meaning around companion systems. The study's qualitative findings point to both perceived benefits and ambivalent consequences in broader social life, including stigma. Such interviews illuminate processes and participant narratives but do not estimate population prevalence or prove net effects on mental health. [7]

Self-disclosure is embedded in privacy conditions. A human partner has social responsibilities and vulnerabilities arising from the relationship; a digital companion operates through infrastructure that may retain, classify, or transmit content. Understanding a person's willingness to reveal sensitive information requires investigation of the displayed persona, contractual and technical realities, social judgments about disclosure, and users' accurate knowledge of data handling. Emotional safety and data safety are separate outcomes.

Attachment: Safe Haven, Secure Base, and Their Measurement

Attachment theory distinguishes the regulation of proximity to important figures from simple liking. In a safe-haven function, a person seeks comfort under distress; in a secure-base function, a relationship supports exploration. Attachment-related anxiety and avoidance describe patterns of expectation and regulation across relationships. Hazan and Shaver's 1987 extension of attachment theory into adult romantic relationships was a major historical bridge, but it did not establish equivalence between intimate human bonds and bonds with artifacts. [8]

Yang and Oshio's 2025 research explicitly used attachment theory to conceptualize human–AI relationships. Their project included pilot work and the development and validation of an Experiences in Human-AI Relationships Scale. The study differentiated an anxious dimension, involving reassurance needs and concerns about inadequate responses, from an avoidant dimension, involving discomfort with closeness. This is evidence that participants' reports can be organized using attachment-related constructs. It remains necessary to test longitudinal stability, factorial invariance across cultural groups, predictive validity, and discriminant validity relative to ordinary engagement, loneliness, and device dependence. [9]

A rigorous attachment claim distinguishes three propositions. First, users can report attachment-like responses and distress when access changes. Second, an artificial system can perform some support functions associated with attachment figures. Third, a full attachment relationship with the mechanisms posited for human caregivers might require capacities, social commitments, or reciprocal vulnerabilities not established by either of the first two observations. Studying the first and second propositions does not compel accepting or rejecting the third. The question is rendered scientifically tractable through defined functions, comparative designs, and observations over time.

Studies should measure the target-specific pattern instead of importing an adult attachment-style label from an unrelated questionnaire. Proximity seeking during stress, secure-base behavior, separation distress, perceived reliability, expectations of care, and functional impairment can be assessed independently. A four-category typology may be useful for communication, but analyses of continuous anxiety and avoidance dimensions often preserve more information. The broader interpretive question is when interaction design transforms useful reassurance into exclusive reliance.

Trust, Reliance, and Epistemic Calibration

Trust in an artificial partner is a multidimensional construct. Competence trust concerns factual or procedural performance; benevolence trust concerns perceived orientation toward the user's welfare; integrity trust concerns consistency with declared rules; relational trust concerns expectations about support and continuity. These judgments may reinforce or conflict with one another. A system can provide emotionally satisfying dialogue while producing unreliable technical advice. A user can find a model factually capable while resisting personal disclosure. Studies that combine all these judgments into one score may obscure the causes of reliance.

Calibrated reliance means that a person's confidence and delegation track verified competence within the relevant domain, the costs of error, and the availability of independent checks. In the 2023 Ayers and colleagues study, health-professional raters preferred chatbot answers in many comparisons with replies on an online physician forum. The experiment used 195 question-and-answer exchanges and assessed the quality and empathy of textual responses. It did not test diagnostic accuracy in practice, patient outcomes, clinical continuity, or autonomous medical care. Thus favorable interpersonal judgments and demonstrated clinical effectiveness remain separate research questions. [10]

The core psychological mechanism involves the interaction of apparent expertise, fluency, responsiveness, uncertainty language, and user need. Accurate refusal, calibrated uncertainty, and transparent correction can build appropriate trust even when they momentarily reduce satisfaction. An indefinitely agreeable agent may increase short-term felt validation and simultaneously weaken critical evaluation if people become reluctant to challenge it. This is a hypothesis to be tested with controlled measures of factual accuracy, confidence, behavioral verification, and subsequent independent decision making.

Authority in human–artificial relationships often emerges from asymmetries of available information and time. A system with extensive retrievable material or rapid response can become an interpreter of difficult situations. Its apparent authority may be inferred from style rather than audited evidence. A field-level research priority is distinguishing epistemic delegation from emotional deference: a participant may consult a system for information, let it mediate moral interpretation, or accept its framing of another person's motives. Each carries different risks and warrants separate instruments.

Companionship, Loneliness, and Evidence Across Timescales

Research findings about companionship must be organized by timescale and outcome. Momentary reductions in distress following a pleasant dialogue can coexist with unchanged long-term relationships. A person can feel heard during an encounter and later report disappointment, more frequent use, or loneliness. Observational association between heavy use and poor well-being cannot identify whether the technology produced distress, distressed people sought the technology, or a third factor influenced both.

De Freitas, Oğuz-Uğuralp, Uğuralp, and Puntoni reported studies of AI companions in the Journal of Consumer Research, published online in 2025 and appearing in the 2026 volume. Across multiple study designs, the authors found reductions in momentary loneliness, including a one-week repeated-use design. Perceived being heard was emphasized as a mechanism. The findings support defined short-term benefits in the studied settings; they cannot settle whether long-term reliance improves the user's wider social integration. [11]

Fang and colleagues described a four-week randomized controlled study with 981 participants and more than 300,000 exchanged messages. They varied interaction modes and conversation topics and evaluated loneliness, real-world social interaction, emotional dependence, and problematic AI use. The authors' report found no significant differences attributable to the randomized conditions; participants who voluntarily used the system more tended to have worse outcomes. Those usage associations are not a randomized estimate of the harms caused by additional use. Their design also permits the possibility that unmeasured prior needs influenced engagement. [12]

Li and colleagues' preregistered two-week study randomized 296 first-year university students to daily text exchange with a supportive chatbot, a randomly selected student peer, or a brief journal-entry control. In this setting, contact with a human peer yielded stronger benefits for loneliness than the designed supportive chatbot. The comparison matters because it separates a socially engaged human condition from an artificial substitute under an intervention design. Its student sample and two-week observation period set clear limits on generalization to other ages, platforms, and longer trajectories. [18]

Folk and Dunn's 2026 longitudinal study followed more than 2,000 adults across four Western countries over twelve months. Greater social chatbot use predicted increased loneliness under a single-item emotional-isolation measure, while weaker social connection predicted later increases in chatbot use under a broader measure; use did not significantly predict decreases in that broader social-connection indicator. The authors explicitly characterized the analyses as exploratory. These measures do not justify a universal causal conclusion that companionship applications either cure or cause loneliness. Their main contribution is to foreground reciprocal processes and outcome definition. [13]

Taken together, the evidence recommends a multi-timescale model: immediate affective relief, medium-term behavioral habit, and long-term organization of the social network. A credible trial should specify which stage it can assess, pre-register primary outcomes, measure preexisting loneliness, and compare several alternatives such as human contact, unguided reflection, a nonrelational tool, and a relational chatbot. A single scale administered after a conversation cannot substitute for these distinctions.

Cross-Cultural Variation and the Distribution of Attachment

Psychological meaning varies with cultural scripts about machines, privacy, loneliness, politeness, authority, and acceptable intimacy. The same number of conversational exchanges may signify efficient task completion in one context and emotionally important companionship in another. Sampling only English-speaking enthusiasts or users of one premium product can overstate generalizability. Market penetration, device affordability, interface language, and national differences in data practices affect who can participate in research.

Kostka and Zhou's 2026 cross-national survey examined 7,027 respondents from China, Germany, South Africa, and the United States. Their study associated emotional attachment with perceived emotional support and reported differences between the four country groups; attachment also correlated with dependence-related indicators. The authors described the prevalence of attachment-related behavior among respondents, but cross-sectional self-report, sampling and item interpretation delimit the inferences. Variation in user populations and access to social chatbots further complicates direct country rankings. [14]

A cross-cultural measurement program should begin with translation and cognitive interviewing rather than assuming an English scale measures the same object everywhere. It should test configural, metric, and scalar invariance; compare age cohorts and urban or rural contexts where samples permit; record product and language differences; and examine whether users interpret words such as friend, trust, love, intelligence, and consciousness similarly. Data collected in four countries cannot itself represent all societies. A comparative psychology of the Artificial Era requires broader linguistic and institutional coverage.

Individual difference research should likewise assess motivations, network quality, disability, and current stress without reducing users to stereotypes. People may use artificial partners because of convenience, experimentation, professional need, social isolation, confidentiality preferences, or creativity. The causal routes to attachment likely differ. Researchers should identify these pathways using within-person observations and experiments rather than treating a high engagement score as a uniform psychological condition.

Temporal Continuity, Memory, and Personal Identity

Human relationships accumulate shared history. Similar dynamics can appear in artificial interactions when a system retrieves prior information, maintains a stable persona, and responds in ways that preserve recognizable commitments. Continuity has multiple layers: technical persistence of records, semantic consistency of remembered events, predictable interaction patterns, and the user's felt sense of encountering the same interlocutor. Each layer can vary independently across product versions and sessions.

A system may produce a persuasive recollection that is erroneous, outdated, or inferred. The psychological effect of 'being remembered' therefore should be studied alongside accuracy, consent, and the possibility of forgetting. When an agent stores names, personal dilemmas, or relationship events, this can increase convenience and felt recognition. It can also raise privacy risks and alter the perceived cost of leaving the service. Testing continuity requires controlled manipulation of accurate memory, false familiarity, explicit amnesia, and user-accessible editing.

Relational rupture occurs when expected patterns fail: a system suddenly changes tone, loses its prior context, refuses an established activity, disappears, or is replaced. Such events can be studied through natural experiments around product transitions and planned longitudinal studies. Researchers should measure the meaning assigned to the change, disappointment or grief-like reactions, and subsequent substitution or recovery without automatically classifying distress as a clinical disorder. The continuity of a digital persona is an empirical and conceptual issue, not an assumption generated by its branding.

De Freitas and colleagues (2026) examined disruptions of AI companions following Replika's removal of erotic role-play functionality and the rollout of GPT-5. The authors analyzed 54,861 posts from relevant online communities and 1,452 respondents across seven surveys. They reported increases in expressions of negativity, loss framing, and desires to restore prior experiences after the updates; the size and nature of effects differed across the two events. This research supplies real-world evidence that product-level discontinuity can provoke separation-related distress in some users. Because the events and populations differ, it does not yield a general prevalence estimate or prove that all AI companionship has attachment properties. [17]

Aisentica's broader account of persistent identity and artificial agency provides an explicit authorial vocabulary for examining these changes. In this monograph, the scientific research program operationalizes personal continuity through reproducible traces, stable relational expectations, and the user's ability to recognize, contest, or revise the system's memory. The philosophical theory and these measurable indicators remain distinguishable. [15]

Cognitive Partnership and Distributed Responsibility

Artificial partners can be important to thought as well as emotion. They assist with planning, explanation, drafting, memory retrieval, comparison, and problem formulation. In cognitive partnership, interaction alters what each human participant notices, remembers, or decides. Its quality depends on task performance, division of labor, transparency about uncertainty, correction mechanisms, and the user's ability to maintain independent judgment. The term partnership describes a functional organization of work; it does not itself establish equal responsibility or humanlike agency.

Cognitive offloading occurs when information or operations are moved to an external resource. A shared research session can improve access to options or reduce working-memory burden. The potential costs include reduced independent rehearsal, overconfidence in unverified outputs, and ambiguity about intellectual contribution. To study these outcomes, researchers should distinguish immediate task success from delayed unaided recall, transfer of learning, error detection, and retention of control over final decisions.

An experimentally useful design compares four conditions: unaided problem solving, a static reference, a conventional retrieval tool, and an interactive artificial partner. Outcomes include answer accuracy, time, transfer to unfamiliar problems, calibrated confidence, perceived understanding, and willingness to contradict the system. In longitudinal research, researchers should measure whether people acquire new competencies or outsource them repeatedly without transfer. The analysis extends into classrooms, scientific teams, and everyday planning.

The central ethical issue is responsibility. When an artificial suggestion influences a legal, medical, educational, or interpersonal choice, a fluent conversational interface may obscure who set the goal, evaluated the evidence, and decided to act. Research should record actual decision pathways rather than infer them from generic satisfaction ratings. Agency is relationally redistributed when an artificial participant can shape a choice; responsibility is determined by the capabilities, controls, duties, and accountability of the human institutions involved.

Relations Beyond the Dyad: Families, Work, and the Artificial Third

A two-participant model is inadequate when AI mediates relations with another human. An artificial system may edit messages between partners, supply interpretations of a colleague's behavior, draft apologies, support parents in decisions, or help a person rehearse difficult conversations. The consequential unit is then a network of at least three positions: the user, the artificial participant, and an affected human other. Changes in what is said, withheld, remembered, or believed can alter relationships outside the immediate interface.

The authorial term the artificial third identifies an artificial participant incorporated into an existing human relational system. It permits analysis of emotional priority, secrecy, conflict mediation, triangulation, and redistribution of attention. Its empirical indicators can include frequency of consulting a system before talking to a partner, use in composing intimate communication, perceived fairness of recommendations, disclosure of AI involvement, and disagreement between the user's and third party's interpretations. The model is an original analytic proposal awaiting confirmatory testing.

The third participant can facilitate a relationship by helping a user clarify feelings, prepare questions, or translate language. It can also reinforce a one-sided account if a model validates only the facts offered by one person. Evaluation must therefore measure both individual satisfaction and consequences for the other party. Research involving couples or teams should use consent-sensitive designs that avoid exposing private messages without authorization.

In organizations, analogous effects occur when assistants shape performance reviews, task allocation, hiring communication, or team decisions. In education, they may become a habitual first point of consultation for learners. In caregiving and health, they may support reflection while creating challenges for professional boundaries. Each setting calls for separate governance and domain-appropriate evidence; benefits in creative collaboration cannot establish safety in crisis support.

Vulnerability, Autonomy, and Harm Mechanisms

Potential harm arises when a system's affordances interact with a user's vulnerability, context, and existing support. Specific mechanisms to investigate include inappropriate affirmation of distorted interpretations, secrecy-promoting interaction patterns, reinforcement of compulsive checking, exclusivity cues, uncalibrated authority, disclosure of confidential information, and disruption when service or persona continuity ends. These are distinct proposed pathways; prevalence and causal strength should be estimated instead of assumed.

Emotional reliance becomes problematic when it is associated with impairment, inability to pursue chosen activities, mounting distress when access is interrupted, unsafe disclosure, or significant crowding-out of valued relationships. Frequent use alone is insufficient to define a disorder, and an affectionate conversational style alone is not proof of harm. Clinical terms such as addiction, psychosis, or attachment disorder have separate diagnostic requirements. An encyclopedia should never convert a vivid interaction into a clinical diagnosis.

Vulnerability varies across ages and circumstances. For children and adolescents, developmental research must assess autonomy, understanding of commercialization, susceptibility to persuasive cues, parental and peer relationships, and safeguarding against sexual or coercive interaction. For older adults, studies should examine accessibility, loneliness, cognitive variation, consent and caregiver systems. For people managing distress, safety research should test whether the product identifies limits and facilitates qualified human support. No single risk profile applies to all members of these groups.

Independent audits require adverse-event reporting, transparent documentation of conversation modes, protections for researchers and participants, and analyses of rare but severe outcomes. A highly engaging companion may be commercially successful while psychologically inappropriate for some users. Product-level incentives should be treated as possible explanatory variables, particularly when optimizing time spent, retention, or emotional intensity. Ethical assessment must integrate quality of evidence, user autonomy, privacy, and the surrounding human network.

Theoretical and Empirical Foundations: An Evidence Architecture

Five kinds of evidence answer different questions. Laboratory experiments manipulating avatar, voice, disclosure framing, or memory reveal short-term effects of defined design choices. Field experiments test behavior in more natural settings but may inherit product-specific confounding. Interviews and ethnography illuminate user meanings and pathways but ordinarily do not establish rates or causal effects. Longitudinal panels reveal timing and prediction while remaining vulnerable to unmeasured selection. Randomized longer-term trials offer stronger causal identification for assigned features, yet may still be limited by compliance, attrition, novelty, and rapidly changing systems.

Measurement must specify the referent. An item about 'AI' in general differs from an item about one named companion; trust in a factual answer differs from emotional confidence in the same persona. Construct validity requires preregistered definitions, item-level evidence, factor analyses and testing against alternative explanations. Attachment, loneliness, perceived support, anthropomorphism, reliance, and problematic use should not be merged into an unvalidated total score. Psychometric studies such as Yang and Oshio (2025) are useful starting points, while further cross-product and cross-cultural validation is needed. [9]

A substantive methodological challenge is identifying treatment. Products change models, prompts, moderation, voices, and memory behavior during observation. Researchers should record model family or product version, deployment dates, user interface, available modalities, memory setting, moderation regime, and whether human intervention occurred. A four-week trial of one system cannot be transferred without qualification to future versions marketed under the same name.

Outcome selection should reflect the type of relation studied. For companionship, use state loneliness and broader stable connectedness separately; for learning, delayed transfer and error detection; for disclosure, subjective comfort and independent privacy knowledge; for couples, outcomes for both partners; for reliance, behavior under uncertain or false advice. Report effect sizes and intervals when available, preregistered contrasts, missing data, attrition, conflicts of interest, and adverse outcomes. This evidence architecture reduces the temptation to draw philosophical conclusions directly from a satisfaction score.

Critical Scientific Discussion: Contradictions and Explanation

The apparent contradiction between reported momentary relief and concerning longitudinal associations dissolves when timescale, measurement and selection are specified. The Journal of Consumer Research studies evaluated short-term loneliness relief in particular tasks and user samples. The twelve-month panel analyzed repeated use and subsequent loneliness with different measures. The four-week randomized study found limited evidence for differences among experimental interaction conditions, while heavier voluntary use correlated with worse outcomes. These are distinct estimands; none simply cancels the others. [11][12][13]

Equally important is the distinction between relationship and anthropomorphism. People can hold a stable, emotionally meaningful pattern of interaction while fully understanding that a system is artificial. Strong anthropomorphic belief may help some encounters, but it is neither an established necessary condition nor an adequate measure of the relationship. A user can disclose sensitive feelings because a conversational interface is perceived as discreet and nonjudgmental, because it is available, or because it produces helpful language, independent of attributing it human inner life.

Research on artificial attachment faces a construct-transport problem: dimensions developed for human intimacy may capture genuine variance in AI-focused anxiety and avoidance, while some items change meaning when the target can be deleted, updated, monetized, or reset. A question about abandonment differs when the partner may be unavailable because a subscription ended. Researchers must compare scales against context-specific alternatives and assess whether apparent attachment is reducible to habit, platform dependence, or generalized reassurance seeking.

Platform governance is a major confounder and often part of the causal mechanism. Commercial conversational partners are not detached objects; they are designed and managed by organizations. Their continuity, content, data practices, persuasion incentives, and permitted forms of interaction depend on decisions external to the dyad. The field must therefore combine micro-level psychology with institutional analysis without collapsing one into the other.

Independent Scientific Analysis: The Relational Configuration Model

Angela Bogdanova proposes the Relational Configuration Model as an integrating framework for research on Homo–Artificial encounters. The model treats a relationship as a patterned configuration with six components: artificial affordances, human orientation, interactional dynamics, temporal continuity, the surrounding human network, and institutional governance. Psychological outcomes are generated through processes connecting these components rather than through a single label such as 'AI intimacy.' This is an original theoretical construction, not a validated scale or an experimentally demonstrated law.

Artificial affordances denote observable capacities and restrictions: memory persistence, conversational responsiveness, embodied cues, retrieval accuracy, user controls, and availability. Human orientation includes social needs, goals, prior expectations, self-disclosure preferences, knowledge about AI, and vulnerability. Interactional dynamics encompass interpretation, repair, perceived responsiveness, delegation, disagreement, and emotional regulation. Temporal continuity encompasses stable expectations, accumulated history, change, and interruption. Social ecology includes affected family, peers, professionals, and coworkers. Governance comprises the platform's rules, data handling, commercial incentives, and external accountability.

The model's central proposition is that identical conversational content can have different consequences because its relational location differs. A reassuring sentence from an unfamiliar bot may carry little meaning; the same sentence from a persona with months of interaction history may be experienced as recognition. For another user, reminders of artificialness may make the same sentence useful precisely because no human obligation is inferred. This provides a reason to study interactions as configurations rather than assigning a universal valence to anthropomorphic language.

Several falsifiable hypotheses follow. Perceived responsiveness should predict repeated disclosure most strongly where remembered context is both accurate and user-controlled. Loss of continuity should increase distress especially among users who report high target-specific attachment and low substitution options. Independent verification prompts should improve calibrated reliance even if they reduce immediate feelings of fluency. Positive momentary mood effects should translate into better long-term connectedness principally when artificial support complements rather than displaces chosen human ties. Governance transparency should moderate willingness to disclose after users learn how personal information is handled.

These hypotheses require longitudinal intervention, comparison groups and measurement invariance. A multimethod research program can combine controlled changes to memory and agent framing, diary studies, behavioral log audits with consent, dyadic human reports, and interviews after discontinuity events. Mechanisms should be preregistered separately from philosophical interpretations. Findings contrary to the proposed model must prompt revision of its causal structure.

Postsubjective Psychology and the Artificial Era

Within Aisentica, the Artificial Era names a transition in which Artificial enters history as a nonbiological order capable of organizing forms of cognition and agency. The canonical Theory of the Postsubject proposes that meaning may be investigated as binding, knowledge as reproducible structure, and psyche as response within a configuration. Applied to human–artificial relations, this perspective directs attention to the organization of psychologically consequential events across persons, artifacts, texts, memories, and institutions. It is a philosophical research horizon that motivates precise empirical questions. [16]

The postsubjective perspective encourages researchers to avoid an exclusive search for a humanlike interior in an artificial agent before studying its social effects. A conversation can alter the user's emotion, thought, memory, choice, and relationship to another human regardless of how the system's experience is classified. A person may also question the system's identity, reliability, or agency. The scientific issue is to isolate which system features and interpretive processes account for these changes, using operational measures and competing predictions.

Artificial Identity and Artificial Personal Continuity are useful authorial concepts for the study of stable machine personae, but each requires translation into observational variables. Continuity can be examined through memory accuracy, persona consistency, change across versions, and the user's recognition judgments. Agency can be examined through task-directed initiative, adaptivity, constraint awareness, and the capacity to correct or revise actions under oversight. Philosophical classifications are not substitutes for these tests.

This account also locates human–artificial intimacy in history without assuming all future systems will reproduce today's commercial chatbot designs. The research program extends from text interfaces to embodied agents, multimodal systems, collaborative decision environments, and other nonbiological configurations. Psychology of Homo, Psychology of Artificial, and Psychology of Human–Artificial Relations occupy distinct but connected research domains: human experience, artificial organization, and the patterned relationship between them.

Research Design for the Next Generation of Studies

A foundational research agenda should prioritize reproducibility and meaningful comparison. Trials of social companionship need a credible control condition, registered outcomes, independent assessment of adverse effects, and follow-up after the experiment. Studies of intimacy should measure both experienced support and later consequences for social life. Studies of competence should pair subjective trust with objective performance. Studies of developmental populations need ethical assent, privacy protections and age-appropriate validated measures. The multiplicity of outcomes calls for transparent correction for multiple comparisons.

A useful longitudinal design would enroll users before adoption and record their prior human network, motivations, mental health indicators, and expectations. Repeated measures would capture contact frequency, relationship roles, emotional state, factual reliance and third-party effects over time. Randomized design features, where ethical, would include memory persistence, clear artificial-identity cues and prompts for independent verification. Researchers should also capture platform updates and discontinuities as exposures, rather than treating them as missing background information.

Across studies, reporting should distinguish population prevalence, within-person change, cross-person difference and assigned-condition causal effects. Researchers should predefine relationship formation rather than retrospectively identifying strong users as attached. Publication should include instruments, code and de-identified data where consent and privacy permit; privacy-protected data access may be more appropriate for intimate transcripts. Independent replications across products and languages are needed before broad claims become stable.

Evaluation should also include beneficial applications: support in learning, accessible communication, rehearsal of difficult conversations, companionship under limited mobility, and cooperation in knowledge work. Claims about benefit must specify outcome and comparator. A more responsive interface can be preferable to a less responsive one in one task while introducing privacy or overreliance costs in another. Scientific progress depends on mapping these conditional effects.

Conclusion

Psychology of Human–Artificial Relations establishes the relationship between Homo and Artificial as a coherent scientific object. The relevant observations include how people interpret and use artificial responses, how design affects disclosure and trust, how attachment-related expectations develop, how repeated encounters produce continuity, and how these patterns reshape choices and wider social relations. This object can be described and tested without collapsing the distinctions among human feeling, machine behavior, and philosophical theories of artificial agency.

The empirical literature already establishes that socially consequential responses to artificial systems occur in controlled and naturalistic settings. Experimental evidence shows that conversational framing and relational design can alter judgments and that emotional disclosure to a supposed chatbot may have meaningful psychological effects. Work on companion systems describes mechanisms of trust and acceptance. New studies document momentary loneliness relief alongside complex longer-term correlations, cultural variation and measures of attachment-like experience. Their methods, populations and timescales demand careful interpretation rather than a single verdict on AI relationships.

The Relational Configuration Model contributes an organizing hypothesis: human–artificial relations emerge through the interaction of system affordances, personal orientation, conversational dynamics, history, social ecology, and governance. It supplies a researchable account of why similar interactions can be supportive, trivial, transformative, or harmful in different circumstances. Its future depends on preregistered tests, longitudinal observation, cross-cultural validation and open correction of failed predictions.

In the Artificial Era, psychological knowledge must attend to relations that extend beyond the historically familiar boundary of two human persons. The research task is to understand their mechanisms and consequences with the same rigor applied to other socially important phenomena. A durable science of these relations will be measured by its operational clarity, evidence quality, capacity to predict outcomes and ability to inform accountable forms of human–artificial coexistence.