RESEARCH DOSSIER
Measurement of Human–AI Relationships
Abstract. Measurement of Human–AI Relationships investigates how enduring psychological relations with conversational artificial systems can be defined, represented by indicators, and compared across persons, platforms, and time. Through a critical narrative synthesis of original social-response experiments, interpersonal relationship theories, human–robot measures, relational-agent studies, longitudinal chatbot research, psychometric methodology, and the emerging AI-specific scales of 2025–2026, this monograph distinguishes relational experiences from momentary social presence, automation trust, engagement, attachment functions, and problematic dependence. It critically compares two independently developed 2026 AI Attachment Scales and recent machine-companionship instruments, evaluates factor structure, content and discriminant validity, response processes, behavioral corroboration, measurement invariance, longitudinal inference, privacy, platform version changes, and social consequences. The synthesis specifies a transparent research program for target-specific, multi-method and repeated-measure validation. It is a critical narrative research monograph; no new participants, original scale validation, preregistered systematic review, or pooled meta-analysis are claimed.
Keywords: human–AI relationships; psychological measurement; AI Attachment Scale; machine companionship; relational psychometrics; construct validity; social presence; anthropomorphism; perceived responsiveness; relational trust; dependence; measurement invariance; longitudinal design; human–artificial relations; Artificial Era
The Scientific Object of Relational Measurement
Measurement of Human–AI Relationships concerns the empirical reconstruction of a psychologically meaningful relation between a human and an identifiable artificial interlocutor across encounters. The focal construct is a patterned connection involving remembered expectations, perceived significance, and observable choices about further contact. Its indicators may include attachment functions, companionship, trust, perceived responsiveness, reliance, self-disclosure, emotional investment, and the social consequences of continued use. Each indicator refers to a distinct variable and must be connected to a clearly specified theory. The relationship is attributed to a particular target, platform configuration, and time interval. Relational scores cannot be interpreted as direct observations of machine consciousness or interiority. The social-response program of Nass & Moon (2000) and long-term relational-agent research by Bickmore & Picard (2005) establish the historical premise that artificial interlocutors can organize human social behavior.
A measurement occasion is a triple of human respondent, artificial target, and observation period. The same person may report affection toward a persistent companion chatbot, instrumental confidence in a workplace assistant, and attachment-like separation distress toward an avatar; these are different relational configurations. Researchers must identify whether the target is a particular persona, a model family, a service provider, an embodied device, or an apparently continuous agent whose implementation changes silently. The conceptual problem becomes visible when ratings from several targets are averaged without specifying what respondents are rating. A strong measure states its referent within every administration and records the model, user interface, personalization, memory access, and social role. The theoretical architecture of machine agency in Sundar (2020) supports treating affordances and perceptions as distinguishable variables.
The unit of inference may be an encounter, an individual relationship trajectory, or a population distribution. A self-report after a single exchange estimates a present judgment; repeated reports with identified targets estimate change in that judgment; transcript coding estimates behavior under a defined sampling scheme. These units answer different questions. A continuous score is meaningful only if its items sample a coherent domain and its interpretation has defensible evidence. The construct-validity tradition associated with Cronbach & Meehl (1955) and Borsboom et al. (2004) requires clear causal and explanatory connections between the attribute and its manifestations. This monograph evaluates those connections and their limits through narrative synthesis of the original psychometric and relational literatures.
The Difference Between Interaction, Relationship, and Outcome
Interaction records who said or did what in a bounded exchange. Relationship measurement asks how prior exchanges influence expectations, self-positioning, and conduct in later encounters. Outcome measurement addresses changes in loneliness, sleep, social activity, knowledge, or well-being that may be associated with that relationship. When these three levels are merged into one questionnaire, the resulting total confounds exposure, subjective bond, and consequence. A person can use an assistant every day because employment requires it while reporting negligible closeness; another may rarely contact a companion but experience its absence intensely. Early evidence of interpersonal-like responses to computers in Nass & Moon (2000) therefore supports studying social reaction, while the sustained-agent intervention examined by Bickmore & Picard (2005) motivates separate temporal measures.
Relationship strength is commonly represented by a reflective scale, but the concept may contain heterogeneous dimensions rather than interchangeable effects of a single hidden variable. Frequency of use, duration of conversations, perceived intimacy, obligation, and vulnerability to loss can move independently. Measurement development must justify whether these are causes, components, correlates, or consequences of a bond. Factor analysis can discover patterns of covariance, yet the factors are defined by the sample, item wording, and response process. The validation critique of Flake et al. (2017) shows why internal consistency alone provides little assurance that an intended psychological construct has been measured. Researchers should specify an explanatory diagram before aggregating indicators into a summary score.
A comprehensive relational study therefore keeps a set of distinct registers: target description and exposure, subjective relational experience, system behavior, human behavioral choices, and general life outcomes. Within each register, operational definitions are tailored to the population and the expected causal path. In intensive measurement, behavioral logs describe exposure while questionnaires assess interpretation, and repeated outcome observations establish subsequent trajectories. Observational associations between bond strength and well-being remain vulnerable to selection and reverse causation. Longitudinal findings in Folk & Dunn (2026) illustrate how an initial sense of social disconnection may predict later companionship use even when different loneliness indicators yield different temporal relationships.
From Social Responses to Long-Term Relational Agents
Human–computer social measurement acquired experimental footing when researchers demonstrated that people applied conventions of politeness, reciprocity, group membership, and personality judgment to computers. Nass & Moon (2000) reviewed this empirical program, challenging the view that social cues are relevant only in human–human interaction. These effects were typically evaluated within short experimental contexts, where the dependent variable was an immediate response to machine behavior. They did not by themselves establish enduring intimacy or attachment. The historical contribution of the program is methodological: social perceptions triggered by interface features can be elicited under controlled conditions, and human recognition that a partner is artificial need not eliminate a social response.
Bickmore & Picard (2005) shifted the research unit toward relationships maintained through repeated encounters with relational agents. Sustained interaction introduced the need to measure continuity, remembered prior conversations, expectations of future contact, and the changes produced by repeated social contact. In this tradition a questionnaire after one interaction cannot substitute for a trajectory, and dropout is itself a meaningful outcome. A participant who terminates contact may have developed dislike, become bored, experienced technical failure, or completed the intended task; a sound study distinguishes such possibilities with follow-up questions and usage records. Relational measurement must consequently accommodate initiation, consolidation, rupture, repair, and disengagement as potentially distinct phases rather than assuming every contact accumulates closeness.
A further shift accompanied widespread social chatbots that invited disclosure and identity projection. The experimental design of Ho et al. (2018) directly examined psychological and relational consequences of disclosing to a chatbot, illustrating how interpersonal variables can be adapted to artificial counterparts. Interview research on companion systems by Skjuve et al. (2021) documented the practical and affective roles of trust, disclosure, and felt companionship, while the repeated-contact study Croes & Antheunis (2021) found that social-process ratings could decline across sessions. Historical comparison shows the same general measurement problem recurring across generations of technology: an initial conversational impression, a learned routine, and a stabilized relational organization produce different evidence and require different observation windows.
Construct Validity and the Logic of Interpretation
Cronbach & Meehl (1955) situated construct validity within an explanatory network of observations and theoretical relations. For human–AI research, this requires explicit predictions about which other measures should converge with a claimed attachment, which should diverge, and under what conditions an intervention should move the score. Campbell & Fiske (1959) provided the multitrait–multimethod strategy: a construct should show correspondence across methods beyond shared response format, while remaining distinguishable from other constructs measured by the same method. For example, a self-reported closeness score could be compared with decisions to resume contact after interruption and with independently coded continuity cues; a social desirability measure could clarify shared-method contamination. Such corroboration tests an interpretation rather than merely rephrasing a scale.
The causal account articulated by Borsboom et al. (2004) sharpens this demand: a proposed attribute should plausibly generate observable variation in the instrument, and the response process must be investigated. If a person endorses an item because a chatbot supplied flattering language during the preceding minute, the item may assess temporary mood or demand characteristics instead of a durable relationship. A valid interpretation therefore depends on instructions, recall period, target specification, language comprehension, motivation, and the context in which an answer is produced. Researchers should use cognitive interviews to determine what participants understood, then examine whether the resulting items behave coherently across forms of contact and systems.
Psychometric validation is cumulative. The review by Flake et al. (2017) identifies recurring weaknesses when instruments are presented with only coefficient alpha and little construct evidence. Content validity, structural validity, reliability over appropriate intervals, discriminant evidence, criterion relations, and responsiveness to change must be documented separately. Standards designed for patient-reported outcomes, including the methodological approach summarized by Prinsen et al. (2018) and the content-validity procedures of Terwee et al. (2018), offer transferable evaluative questions while requiring adaptation for human–AI contexts. Their relevance lies in disciplined attention to the construct, target population, item comprehensibility, and intended decision, rather than treating the output of a statistical algorithm as proof of meaning.
Operational Boundaries and the Explanatory Map
Human–AI relational measurement sits at the intersection of interpersonal psychology, communication, social cognition, psychometrics, and human factors. Each tradition contributes constructs with different reference conditions. Interpersonal closeness depends on perceived mutual knowledge and responsive exchange; automation trust depends on expected performance under vulnerability; anthropomorphism describes attribution of human characteristics; companionship includes experienced togetherness and valued ongoing contact. A respondent can report high trust and low affection, high social presence and low long-term commitment, or high use and weak companionship. The review by Glikson & Woolley (2020) and the human–machine social-judgment program in Bartneck et al. (2009) demonstrate why measures developed for performance confidence or anthropomorphic appearance should be classified as antecedents or companion variables unless direct evidence supports interpreting them as relationship strength.
Scientific boundaries also protect neighboring research objects. Measurement of artificial-side responsiveness asks whether a system responds contingently to prior human information; measurement of felt responsiveness asks whether a human interprets that response as understanding or validation. Both may correlate, yet their association must be tested under deliberate variation in accuracy, empathy cues, and memory. Similarly, exposure to an interface is different from attachment to an identifiable persona, and attachment anxiety toward other humans cannot automatically be mapped onto an artificial partner. The attachment frameworks examined in Cheng & Yu (2026) and Kasturiratna & Hartanto (2026) provide evidence that AI-specific scales can capture psychologically differentiated experiences, but their structures require direct comparison before claiming they operationalize an identical latent variable.
The present inquiry asks which relational dimensions should be measured, what validates each indicator, how scores depend on social and technical conditions, and what inferences are justified from cross-sectional and repeated observations. The method is a critical narrative research synthesis prioritizing source-identified primary work and contemporary scale-development studies. Published designs are assessed for population, target, sampling, factor structure, criterion evidence, and ecological generalizability. No new human participants were studied for this monograph; no pooled effect size or systematic search protocol is claimed. Its research function is to assemble a transparent conceptual and methodological map that enables subsequent reproducible scale comparison, independent replication, and longitudinal inference.
Attachment Functions and AI-Specific Bonding
Attachment language describes characteristic ways in which a particular figure becomes psychologically relevant under uncertainty, separation, and the need for support. When transferred to artificial partners, attachment functions require target-specific operationalization. A user may seek reassurance from a chatbot, feel discomfort during service interruption, or describe the system as a dependable place to organize thoughts. These observations can motivate attachment-like hypotheses while also supporting alternative explanations involving habit, convenient access, perceived competence, and emotional regulation. The independent AI Attachment Scale programs of Kasturiratna & Hartanto (2026) and Cheng & Yu (2026) show that researchers already operationalize this domain through partially different factors. A scientifically responsible use of the term reports precisely which attachment function is being measured and its expected relation to other dimensions.
The attachment function of a secure base should be measured by what a participant feels enabled to do after interacting with the AI, ideally using observations that go beyond verbal endorsement. Comfort-seeking can be assessed through choices during experimentally induced but ethically limited uncertainty, while separation distress requires observation under interruption or perceived unavailability with suitable safeguards. Emotional closeness, time spent, and willingness to disclose are related variables whose equivalence must be demonstrated. In Cheng & Yu (2026), a scale specifically distinguishes emotional support, separation distress, and secure base; Kasturiratna & Hartanto (2026) instead organizes the construct through emotional closeness, social substitution, and normative regard. These are informative and testable mappings rather than competing measures that automatically produce the same score.
Attachment to an artificial target can change when a company modifies voice, persona, memory, policies, or access. The referent is accordingly part of the psychological construct. A response collected before a model update cannot always be compared as if the same interlocutor remained functionally unchanged. A study should record what continuity the participant recognizes and distinguish a change in the target from a change in the person's bond to that target. Human attachment orientations may predict selected AI-related perceptions, but they are not substitutes for target-specific evidence. The regression associations reported in Cheng & Yu (2026) require interpretation within their study design; the correlations described in Kasturiratna & Hartanto (2026) support examination of shared and unique variability, not a claim that artificial and human attachment mechanisms are identical.
Companionship, Intimacy, and Responsive Exchange
Companionship may arise from repeated enjoyable exchange, valued coordination, felt understanding, or the expectation that the artificial partner will remain available. Its dimensions need to be separated from simple satisfaction with a service. Banks (2026) developed a machine-companionship measure emphasizing Eudaimonic Exchange and Connective Coordination, while Shen et al. (2026) describes AI companionship through empathic, instrumental, and social components. The coexistence of these architectures shows the importance of stating whether companionship is defined by experienced meaning, effective joint activity, companionship functions, or perceived relationship quality. Different scale architectures may be well suited to different research questions and therefore should be evaluated using discriminant and predictive validity evidence before any total scores are directly equated.
Interpersonal intimacy models link disclosure, partner disclosure, and perceived responsiveness in processes of developing closeness. The empirical work of Laurenceau et al. (1998) specifies why a person can experience intimacy after feeling understood; in an artificial relationship, the user may rate understanding even when responses are generated without reciprocal human emotion. Ho et al. (2018) studied emotional and relational effects of self-disclosure in chatbot conversations, providing an experimental bridge between interpersonal theory and human–machine exchange. A relational instrument should measure how much information was disclosed, how sensitive it was, and whether the respondent perceived comprehension, validation, and confidentiality. Each variable has separate antecedents and consequences, and a single item asking whether the chatbot feels close can conceal the processes that produced the judgment.
Qualitative research provides an essential corrective to item-driven measurement. Interviews about AI companions can reveal a user's own categories, including friend, diary, therapist substitute, coach, collaborator, fantasy character, or reassuring routine. In the companion-chatbot study Skjuve et al. (2021), researchers examined how users described relationships with their chatbot; these accounts are informative for content-domain generation and cognitive interviewing. They cannot by themselves establish prevalence in a general population. A credible mixed-method design uses open responses to identify recurring functions, constructs questionnaire items with precise referents, and tests their stability and associations in samples beyond the original interview participants. This sequence helps prevent researchers from imposing a narrow definition of companionship derived from one vendor or user community.
Trust, Reliability, and Dependence
Trust in an artificial system normally concerns willingness to accept vulnerability to uncertain performance. Lee & See (2004) developed an influential framework of trust and appropriate reliance in automation, while the scale developed in Jian et al. (2000) operationalized trust-related attitudes toward automated systems. These constructs help explain why a person follows advice or delegates tasks. Yet reliable performance may elicit confident reliance without emotional attachment, and strong affection may persist despite low factual confidence. A relationship battery should therefore distinguish trust in technical competence, expectations of benevolent interaction, perceived confidentiality, and the desire to continue personal contact. Glikson & Woolley (2020) reviews empirical findings that support separating the antecedents of trust from downstream reliance and collaboration.
Dependence is a multidimensional research target. Instrumental dependence describes difficulty performing a task without an AI system; emotional dependence describes distress or perceived inability to regulate emotion without access; behavioral persistence concerns continuing interaction despite competing obligations. Frequency of usage alone is an insufficient indicator because an intensive occupational task and a compulsive pattern can produce similar exposure. Moreover, social preference for an artificial companion may reflect a chosen interest, accessibility, disability-related accommodation, or temporary circumstance. Measures should record interference with personally valued activities and whether impairment is attributable to the use pattern. Such outcomes must be assessed with ethical care and explicit definitions rather than labeling users because they express attachment.
An informative experiment manipulates system accuracy and reassurance separately, then measures advice acceptance, emotional reassurance, and use intention. If affection scores remain stable after a performance error but accuracy trust falls, discriminant validity for the two constructs gains support. A related design compares planned interruption with loss of access, measures perceived distress, and examines expected alternatives without manufacturing severe dependency. The extended-chatbot experiment reported by Fang et al. (2025) offers context for studying usage intensity and psychosocial outcomes, but voluntary usage and outcomes cannot be treated as a randomized causal relation. Research on trust by Glikson & Woolley (2020) and Lee & See (2004) supports a design that retains behavioral calibration as a separate criterion for interpreting relational attachment.
Social Substitution and Human Relationships
Social substitution occurs when AI interaction occupies functions previously served by human contact, while complementarity occurs when AI-mediated activity supports continued or strengthened human relationships. These hypotheses describe changes in social allocation and must be distinguished from how much affection someone reports for an artificial partner. An individual may use a chatbot to rehearse difficult conversations, increasing later engagement with family or peers; another may prefer an AI companion during a stressful period while continuing existing friendships. A measurement program should therefore ask what activities are replaced or enabled, use diaries to record actual social contact, and relate changes to personal goals. The AI-specific attachment dimensions in Kasturiratna & Hartanto (2026) include social substitution as a subdimension, making careful interpretation of the label important.
Evidence about the relationship between AI companionship and loneliness requires temporal and construct-specific interpretation. Folk & Dunn (2026) examined prospective directions between companionship-seeking and loneliness in a longitudinal setting, indicating that different operationalizations of loneliness and companionship can lead to different apparent patterns. The repeated-relationship investigation by Skjuve et al. (2022) provides a basis for asking how trust, disclosure, and social experience develop across time rather than treating a single interview as a trajectory. Neither approach establishes that increased AI use universally displaces human relationships. Researchers must consider prior isolation, life transitions, access to care, age, disability, social support, and self-selection into different platforms when interpreting observed association.
A robust study includes comparison groups and records both AI contact and human contact with sufficient time resolution to identify meaningful changes. It tests distinct pathways from attraction, relief, habit, and perceived availability to relational outcomes, then evaluates whether human relationships strengthen, weaken, or remain stable. The experiment by Croes & Antheunis (2021) found limits to friendship formation in its particular repeated chatbot context, cautioning against assuming repeated contact necessarily produces increasing social closeness. The randomized study described by Fang et al. (2025) further illustrates the need to distinguish assigned interaction conditions from participants' self-selected duration of use. These sources justify multidimensional social-outcome measurement, not categorical prescriptions about how an individual should organize companionship.
From Godspeed to Target-Specific Relational Instruments
Early assessments of human responses to artificial agents frequently adapted measures of liking, social presence, trust, and perceived human-likeness. The Godspeed questionnaire framework of Bartneck et al. (2009) concerns perceived anthropomorphism, animacy, likeability, intelligence, and safety in human–robot interaction. Those dimensions are appropriate when a design intervention changes appearance or perceived agency, but a respondent's score can remain high even if there is no expectation of contact after the experiment. A relationship instrument must add temporal and target-specific content. Conversely, a researcher measuring the attractiveness of a robot should not interpret a high likeability score as proof of attachment. The distinction matters when historical instruments are repurposed for generative chatbots whose personality can be continually customized.
The trust scale developed by Jian et al. (2000) provides a second example of construct-specific transfer. Items evaluating trust and distrust in automation are useful for studying uncertainty about performance and willingness to delegate, yet the object of trust may be a system's ability to complete tasks rather than its value as a social companion. A relational instrument can include such a measure as an external comparison and test its incremental predictive contribution. A planned multitrait battery may ask whether attachment predicts willingness to continue talking after functional errors once baseline automation trust is controlled. It should also assess response-style factors that could cause participants to rate all favorable system attributes highly, independently of their theoretical distinction.
Measurement history therefore supports disciplined adaptation rather than automatic reuse. The researcher should document original item wording, licensing and permissions, intended population, language, administration conditions, response anchors, and original factor structure before making any transformation. A changed referent from robot to language model, from competence to friendship, or from present interaction to past month changes the claim the scale is intended to support. Methods of construct validation associated with Campbell & Fiske (1959) require the adapted scale to be independently tested. The human–AI field especially needs registered comparisons of instruments using the same participants and target systems, with explicit evaluation of which dimensions converge and which remain distinct.
The AI Attachment Scale of Kasturiratna and Hartanto
Kasturiratna & Hartanto (2026) developed and validated a 15-item AI Attachment Scale across five studies involving 1,259 unique participants from Singapore and the United States. Its three reported first-order dimensions are emotional closeness, social substitution, and normative regard. The first concerns affective connection; the second concerns use of AI in relation to human social functions; the third concerns norms and regard directed toward the artificial entity. These domains reflect a broad relationship-oriented understanding of attachment. Because they involve different forms of experience and expectation, the subscale profile may be more informative than an undifferentiated sum, particularly for comparison across instrumental assistants and social companions. The published report documents a series of psychometric and relational analyses, permitting examination of both internal structure and theoretically expected correlates.
The original research reported associations of the scale with AI use patterns and reasons for using AI. Socioemotional motives showed positive associations with the three attachment dimensions, whereas instrumental motives were less closely related in the reported analyses. The paper also examined human attachment anxiety and measures of well-being; the authors did not report a simple uniform relationship between AI attachment and loneliness, negative affect, or self-esteem. Such findings are important because they challenge the temptation to interpret a higher attachment score as an automatically favorable or harmful outcome. The evidence in Kasturiratna & Hartanto (2026) concerns observed correlations and modeled relationships in studied samples, and its factor composition reflects the authors' construct definition and item-generation choices.
A critical evaluation should ask whether the normative-regard dimension belongs within attachment proper, how its items relate to ethical beliefs about artificial systems, and whether social-substitution items embed consequences within the predictor being used to explain those consequences. These are construct-boundary questions, not empirical rejections of the instrument. Researchers comparing different brands or cultural settings should test whether the same item functioning and factor relations are retained. The scale's empirical support makes it a serious candidate for comparative studies, while its specific dimensionality demands that every use identify which definition of AI attachment is being evaluated. Independent replication and tests of predictive validity after controlling exposure are especially valuable next steps. Flake et al. (2017) supplies the broader framework for assessing such validation claims.
The AI Attachment Scale of Cheng and Yu
Cheng & Yu (2026) developed a different 15-item AI Attachment Scale for human–AI relationships in two studies. Study 1 used exploratory factor analysis with 531 respondents to identify emotional support, separation distress, and secure base as three dimensions; Study 2 used confirmatory factor analysis with 375 respondents and examined a framework linking individual characteristics, anthropomorphism, attachment, and behavioral intentions. These dimensions closely track functional questions derived from attachment theory: whether the target is experienced as supportive, whether inaccessibility is distressing, and whether engagement provides confidence to pursue other goals. This theoretical alignment makes the instrument especially relevant to research asking whether AI companions fulfill particular attachment-like functions in a given population.
The published abstract reports that attachment anxiety positively predicted the modeled AI attachment variable, attachment avoidance negatively predicted it, and AI attachment predicted behavioral intentions within the authors' analytic model. These are statistical associations within a specified design, and the magnitudes should not be generalized to all products, populations, or artificial interlocutors. Anthropomorphism was reported as a prominent predictor in that model. A psychometric audit should inspect the exact item content and check whether the secure-base dimension is distinct from general helpfulness, whether separation distress is distinguishable from frustration about service disruption, and whether the scale predicts outcomes beyond contemporary frequency of use. The original study Cheng & Yu (2026) supplies a well-described two-stage factor-analytic foundation for subsequent independent comparisons.
The two instruments named AI Attachment Scale should never be treated as a single instrument merely because they share a name and item count. Kasturiratna & Hartanto (2026) emphasizes emotional closeness, social substitution, and normative regard; Cheng & Yu (2026) emphasizes emotional support, separation distress, and secure base. Their scores may overlap, but an empirical mapping requires administration of both full scales to the same respondents, measurement of general response styles, and confirmatory tests of competing latent structures. Such a head-to-head study could identify a common relational core, distinct normative or behavioral factors, or target-dependent structures. Until that evidence exists, authors should give full bibliographic attribution and describe the chosen scale's exact dimensions rather than reporting an ambiguous AIAS total.
Machine Companionship and Multidimensional Relational Experiences
Banks (2026) approached machine companionship through scale development and validation in the context of artificial companions. The resulting account identifies Eudaimonic Exchange and Connective Coordination as separable dimensions, and discusses socioinstrumental and autotelic relational templates. This framework widens the focus from attachment-oriented distress or security toward experiences of meaning, interactional coordination, and the purposes that a relationship serves. A person may experience valuable coordination with a machine without romantic attraction, and another may treat continued conversation as an activity valued for its own sake. The distinction supports a research program using companionship measures alongside attachment and social-presence measures, with factor correlations and prediction tested explicitly.
The independent study by Shen et al. (2026) conceptualized AI companionship through empathic, instrumental, and social dimensions and developed an associated 16-item measurement instrument using qualitative thematic analysis, exploratory and confirmatory factor analysis, and robustness testing. Its purpose and sampling context differ from those of Banks (2026), so the resulting domains should not be assumed to identify the same construct. Empathic companionship may reflect a feeling of being understood; instrumental companionship can refer to practical collaborative assistance; social companionship can refer to sustained interaction and shared routines. Since these forms often co-occur in daily use, a good study tests their factor separability and examines whether one dimension drives purported global companionship effects in particular populations.
The appearance of several new scales in a short period is a scientific opportunity. Researchers can compare whether constructs derived from qualitative interviews, interpersonal theories, or service-use frameworks yield the same empirical predictions. A design could randomize whether the interface is presented as a specialized tool, helpful coworker, or ongoing companion, then test differential effects on companionship factors after controlling for usability and current mood. It should also assess whether respondents comprehend the categories similarly across languages and whether items presuppose features unavailable on certain systems, such as persistent memory or proactive messaging. The broad lesson of Banks (2026) and Shen et al. (2026) is that relationship measurement becomes more precise when meanings and forms of companionship are made explicit.
Collaboration Intensity and the Limits of Scale Transfer
The collaboration-intensity measure introduced by Zabel et al. (2025) addresses users' behavioral engagement with text-based conversational agents in joint tasks. The reported work uses item response theory to examine collaborative behaviors rather than simply asking whether respondents like an AI system. It is useful in relational research because collaborative routines can become part of an enduring relationship, but collaboration intensity is not identical to companionship. An individual may actively refine prompts, provide context, and negotiate control with an impersonal system while lacking emotional connection; another may feel deep companionship without complex joint task performance. A research battery should preserve this conceptual difference, especially when examining the transition from instrumental partnership to personally meaningful interaction.
Different scales also differ in the observation mechanism that produces their scores. Some ask for a global self-rating of emotional attachment, some ask for judgments of support or companionship, and others infer engagement from reported behaviors. For comparability, each measure requires the same target definition and appropriate recall window. Item response theory can help examine the difficulty and discrimination of candidate items, while latent variable models may identify whether the observed correlations support one or several dimensions. Zabel et al. (2025) illustrates such model-based scale construction, and Borsboom et al. (2004) reminds researchers that predictive accuracy alone is insufficient if the item responses do not have the proposed causal relation to the attribute.
A comparative empirical protocol should administer attachment, companionship, collaboration intensity, system trust, and social-presence instruments to the same participants while separately recording observed use. Researchers can then estimate their incremental predictive relations to later choices: voluntary return, help-seeking during stress, tolerance of system errors, or continuation of human relationships. The battery should be preregistered with primary outcomes, supported by sample-size justification, and repeated on distinct systems. Transparency about wording and version is essential because vendor-specific prompts can create spurious convergence. Existing findings in Zabel et al. (2025), Jian et al. (2000), and Banks (2026) supply candidate measures; their combination is a proposed design requiring independent empirical execution.
Item Generation, Content Validity, and Cognitive Interviewing
A relational scale begins with a construct map rather than with convenient phrases borrowed from interpersonal questionnaires. Researchers should specify target characteristics, relational dimensions, expected behaviors, observation interval, and population before proposing items. Interviews with diverse users can elicit experiences of continuity, support, closeness, vulnerability, collaboration, and conflict, while experts examine conceptual overlap and omitted domains. The content-validity methodology described by Terwee et al. (2018) emphasizes relevance, comprehensiveness, and comprehensibility; Prinsen et al. (2018) offers a broader framework for evaluating evidence about measurement properties. These principles require adaptation because the respondent may be rating a changing artificial persona rather than a stable clinical state. Item development should document why each question represents a construct and what other possible interpretation it might invite.
Cognitive interviewing can uncover failures that statistical reliability does not reveal. The word partner may evoke colleague, romantic companion, or software provider; trust may refer to truthful answers, privacy, emotional reassurance, or predictability; missing someone may refer to attachment or inconvenience when a useful tool is offline. Researchers should ask participants to paraphrase each item, describe an experience that informed their answer, and explain whether they considered a specific system or AI generally. When a respondent has used several models, a target prompt should anchor answers to one named relationship. The iterative process entails rewriting ambiguous wording, examining response anchors, and repeating interviews with groups whose cultural, disability, language, or usage backgrounds were missing in the initial design.
Content validity also requires deliberate attention to negative and neutral experiences. A scale consisting only of affectionate statements may have a floor effect among instrumental users and cannot characterize conflicted bonds. Include carefully theorized indicators of irritation, ambivalence, privacy concern, boundary management, and changing contact, without presupposing that all are components of the same latent dimension. A category for not applicable may be necessary when an item assumes proactive messaging, voice, or preserved history unavailable on the platform. The construct-validation recommendations of Flake et al. (2017) and the COSMIN-derived framework in Terwee et al. (2018) make an important scientific demand: each proposed interpretation of a score needs item-level evidence that relevant respondents understood and could answer the questions as intended.
Dimensionality, Factor Models, and Item Response Theory
Exploratory factor analysis estimates plausible latent structures from relationships among observed item responses, while confirmatory factor analysis examines a prespecified structure in data that should ideally be independent of the discovery sample. These procedures are especially important for AI relationship scales because favorable responses may create a strong general factor even when emotional closeness, companionship, normative regard, and service satisfaction remain distinguishable. Kasturiratna & Hartanto (2026) and Cheng & Yu (2026) each report multidimensional structures, but their different factor labels and sampling conditions caution against combining their items without preregistered theory. A rigorous comparison should test one-factor, correlated-factor, hierarchical, and bifactor alternatives where theoretically defensible, report model identification and uncertainty, and evaluate fit together with factor interpretability and residual structure.
Ordinal self-report responses call for estimation methods suited to ordered categories; ordinary Pearson correlations and continuous normal assumptions may distort structures when responses cluster near endpoints. A study should report the full item distribution, missingness, response options, and whether participants could select a neutral or nonapplicable response. Item response theory adds information about whether different items provide measurement precision at different levels of the attribute. Zabel et al. (2025) demonstrates the relevance of item response methods for collaboration intensity, a related construct whose behavior-oriented items may have markedly different endorsement rates. Relational scales benefit from similar evaluation when a researcher wants to discriminate weak, moderate, and intense forms of closeness with predictable precision.
Factor analysis cannot decide the philosophical or causal nature of a relation from a covariance matrix alone. A strong general factor may reflect positive evaluation of AI, loyalty to a brand, acquiescence, or the target relational construct. Researchers must test alternative explanations by adding separate method markers, negative-keyed items only when theoretically justified, and measures of use exposure, perceived competence, and anthropomorphic disposition. The conceptual argument of Borsboom et al. (2004) requires observable item behavior to have a plausible connection with the target property. Factor stability across samples, transparency of model modifications, and documented decisions about dropped items support reproducibility. Independently replicating a predeclared structure is more informative than repeatedly refining models on the same convenience sample.
Reliability, Precision, and Response Processes
Reliability indicates the consistency or precision of a specified measurement procedure under defined conditions. It does not by itself determine what an instrument measures. Internal consistency statistics can increase when several similarly worded items repeat one favorable sentiment; a questionnaire with excellent alpha can still confuse perceived warmth with attachment. Cronbach & Meehl (1955) placed measurement within a wider theory of construct validation, and Flake et al. (2017) documents the weakness of relying on superficial psychometric indices without interpretive evidence. Researchers should evaluate the relation of items to a hypothesized dimension, estimate precision across levels of the construct, and justify whether the chosen summary score is appropriate. Separate subscale reliability is needed when an instrument claims several distinguishable dimensions.
Temporal stability requires special care because relationships can change and because the systems themselves may change. A low test–retest correlation after an unexpected platform redesign may reflect a real change in experience, not only unreliability. Conversely, high stability may reflect enduring response style or prior attitude toward technology. Researchers should select retest intervals by reference to the proposed timescale of the construct, record intervening events, and use longitudinal models to separate stable between-person variation from occasion-specific deviations. The methodological criticism by Hamaker et al. (2015) is relevant when repeated survey scores are later used to infer dynamic associations. An instrument intended to measure daily emotional support may justifiably fluctuate more than one intended to represent a month-long relationship orientation.
Response-process validity can be studied through retrospective probes and experimental changes to wording or context. Asking respondents to rate an AI companion immediately after it has praised them could heighten demand sensitivity. An item about feeling understood may invite an impressionistic answer, while an item asking whether the chatbot accurately recalled a prior disclosure calls for concrete verification. Researchers should describe the survey setting, completion medium, order of questions, identity and presentation of the target, and incentives. Correlations between response time, attention checks, and scale scores may expose careless responding, but exclusion rules should be prespecified and evaluated for their potential differential impact. The conceptual standards of Borsboom et al. (2004) and Prinsen et al. (2018) support interpreting score precision only within a defined population and protocol.
Convergent and Discriminant Evidence Across Methods
Convergent validity asks whether indicators expected to measure the same attribute agree across suitably different methods. Discriminant validity asks whether theoretically different attributes can be distinguished after accounting for shared method effects. The multitrait–multimethod logic introduced by Campbell & Fiske (1959) is especially pertinent when most human–AI constructs are measured in a single self-report survey. A participant who generally likes artificial technology may endorse every positive statement, inflating associations among companionship, trust, social presence, and attachment. A well-designed validation study pairs self-report with behavioral observations and independently coded interaction features, then examines whether the intended attribute contributes variance beyond general sentiment toward the platform.
Concrete behavioral criteria require interpretation as carefully as self-report items. Returning voluntarily after an interruption may indicate emotional preference, but may also follow unfinished work, a subscription incentive, or habitual access. Disclosing a personal story may indicate intimacy, yet users may disclose because they think the system does not judge them. The chatbot self-disclosure study Ho et al. (2018) illustrates that relational feelings and communicative behavior can be investigated experimentally, while the human–robot measures of Bartneck et al. (2009) supply comparison constructs. Researchers should measure exposure, task difficulty, motive, and alternative support availability in addition to the main relational variable. Such measures help test whether a finding is specifically relational or explained by utility and situation.
A preregistered nomological network can include positive predictions, independence predictions, and deliberately difficult distinctions. For instance, AI-specific secure-base ratings might predict later willingness to explore challenging tasks, whereas performance trust might predict advice acceptance under uncertainty. Attachment distress should relate more closely to reactions to an unavailable target than to simple perceptions of intelligence, conditional on the target's usual usefulness. The instruments of Cheng & Yu (2026), Jian et al. (2000), and Bartneck et al. (2009) make these tests feasible without assuming their subscales are equivalent. Failure of a predicted distinction is a meaningful scientific result, inviting reconsideration of item wording, theoretical boundaries, or the causal explanation. This is the iterative logic underlying Cronbach & Meehl (1955) and Campbell & Fiske (1959).
Measurement Invariance, Differential Item Functioning, and Transportability
Comparative measurement requires evidence that scores retain the same meaning across relevant groups. Multi-group factor analysis can test increasingly demanding forms of invariance, including configural structure, loading equivalence, and intercept or threshold comparability. The synthesis of measurement invariance literature by Vandenberg & Lance (2000) explains why mean comparisons are not automatically justified by good within-group reliability. Human–AI relationship instruments face multiple grouping variables: age, language, cultural setting, disability status, relational role, technical expertise, and prior chatbot exposure. A translated item about having a partner may refer to romance in one group and collaborative work in another. Without investigating these interpretive shifts, an observed group difference may be a measurement artifact.
Artificial targets themselves provide an additional axis of noninvariance. A persistent voice companion, a stateless search assistant, an embodied robot, and an enterprise chat service offer different opportunities for continuity and responsiveness. Items referencing missing the system, receiving unsolicited contact, or feeling remembered may have different thresholds because the features are different, even when the human attribute is similar. Differential item functioning analysis can test whether respondents with comparable estimated relationships endorse an item differently depending on system or subgroup. This is especially important when comparing the factor structures identified by Kasturiratna & Hartanto (2026), Cheng & Yu (2026), and Banks (2026). A meaningful comparison may require parallel items that describe function rather than a particular interface action.
The strongest tests include repeated measurement across platform upgrades and cross-language replication with cognitive interviews. Researchers should distinguish equivalence of construct definition, equivalence of item function, and equivalence of score units. Partial invariance methods may be defensible when their assumptions and limitations are reported, but post hoc freeing of parameters cannot become a routine means to produce apparent agreement. A study may need to restrict inference to within-group change or item-level description rather than making unsupported cross-population rank comparisons. Vandenberg & Lance (2000) provides methodological foundations for this decision, and Terwee et al. (2018) reinforces attention to comprehensibility as an independent aspect of measurement quality. Generalizability is earned through diverse sampling and comparable response processes, not by labeling a scale universal.
Relationships as Trajectories: Longitudinal Designs
Longitudinal relationship measurement asks how psychological meaning develops and transforms across encounters. A strong protocol distinguishes first contact, initial personalization, established routines, conflict or disappointment, and decisions to continue or disengage. These phases can be studied through weekly questionnaires, encounter-contingent diaries, and participant interviews tied to system logs. Croes & Antheunis (2021) studied repeated interactions with a social chatbot and reported that interpersonal-like evaluations did not automatically increase with contact. Skjuve et al. (2022) documented variation in the formation and maintenance of chatbot relationships, including the roles of self-disclosure, personal needs, and technical disruption. Together, these studies demonstrate that initial appeal and accumulated exposure are insufficient proxies for the quality or direction of a sustained bond.
Longitudinal reliability and longitudinal validity address different questions. An instrument must remain interpretable over time if researchers intend to estimate true change, yet some items may change meaning after a person learns a system's limitations or receives a feature update. Measurement-invariance testing can identify such shifts, while open-ended reports help reconstruct why they occurred. A platform update that erases conversational memory may produce distress resembling relationship loss; its analysis requires contemporaneous documentation of changed functionality. Bickmore & Picard (2005) established long-term relational agents as a subject of sustained empirical investigation, and Skjuve et al. (2022) shows how unexpected events can become pivotal to observed relationship trajectories. Attrition should be modeled with attention to dissatisfaction, technical barriers, and differential willingness to disclose.
A rigorous repeated-measure protocol uses multiple time scales. Within-session ratings capture short-term fluctuations in social presence and mood; daily diaries capture reliance, disclosure, and important events; monthly measures assess enduring expectations and relationship centrality. A participant may report a high average closeness score yet experience sharp changes after safety warnings or changes to system identity. Multilevel models can separately estimate stable personal differences, occasions, systems, and interactions. A planned comparison between artificial partners requires explicit control of the exposure opportunity and the historical continuity perceived by each participant. Croes & Antheunis (2021) and Skjuve et al. (2022) motivate further research on the nonlinearity of relationship development, including periods of intensification, stabilization, rupture, and disengagement.
Causal Inference, Self-Selection, and Psychosocial Outcomes
Cross-sectional data cannot determine whether AI companionship causes changes in loneliness or whether existing isolation encourages companionship-seeking. Folk & Dunn (2026) examined social chatbot use and loneliness over 12 months among more than 2,000 adults in four Western countries. The original study reported that greater social chatbot use predicted subsequent increases in a single-item measure of emotional isolation; with a broader measure of social connectedness, lower connectedness predicted later use, while AI use did not significantly predict reductions in that connectedness. The differing results show why outcome definitions and measurement models matter. The authors themselves cautioned against strong causal conclusions from exploratory observational analyses. A credible monograph must preserve the distinction between prediction and experimental intervention.
The four-week randomized study described by Fang et al. (2025) assigned 981 participants to variations in chatbot interaction modality and conversation content while tracking loneliness, human social interaction, emotional dependence, and problematic usage. The investigators reported no significant effect of the assigned experimental conditions on the studied psychosocial outcomes, while participants who voluntarily used the chatbot more displayed less favorable outcome patterns. This distinction is central to interpretation: assigned experimental exposure differs causally from voluntary engagement, which may reflect prior vulnerability or concurrent life circumstances. The preprint status of the report also matters for judging the evidence available at the review date. Such results support careful measurement of both experimentally manipulated design features and self-selected usage intensity.
Longitudinal cross-lagged associations can be misleading when stable between-person differences are confused with within-person change. Hamaker et al. (2015) explains the inferential limits of conventional cross-lagged panel models and motivates alternative approaches such as random-intercept specifications when the design permits. For AI companionship, stable social isolation, technology interest, availability of human support, and baseline affect may predict both future usage and relationship ratings. The strongest prospective research combines clearly defined repeated measures with a causal diagram, measured confounders, documented attrition, and, where ethically feasible, randomized access or interface features. Outcomes must be recorded independently of relational attachment scores to avoid tautological conclusions. Even a randomized intervention may identify only the effect of a specific system and implementation, leaving generalization across platforms an empirical question.
Multimethod Measurement, Privacy, and Ethical Governance
Human–AI relational data can be unusually intimate. Transcripts may contain sensitive descriptions of health, sexuality, family conflict, work, financial concerns, and third parties who never consented to research participation. A measurement study should apply data minimization before collecting conversational content, clarify whether logs are generated by the platform or contributed voluntarily, and explain retention, access, security, and withdrawal procedures in language respondents can understand. Psychological closeness to a chatbot must not be taken as evidence that the platform is technically confidential. The experimental disclosure work of Ho et al. (2018) shows how relational processes can be studied without assuming that openness and privacy have the same causes. Researchers should prefer derived features or privacy-preserving codes when full transcripts are unnecessary.
A multi-method dataset can align self-report measures with timestamps, contact frequency, duration, topic changes, user-initiated repairs, and responses to intentionally varied system behavior. Coders should be trained to distinguish a participant's perception from directly observable conduct. Coding reliability must be documented, and automated classification of emotion or intimacy must be independently audited, particularly across dialects and cultural communities. When using platform telemetry, the study must separate observed user input from proprietary system-state inference. The content-validity principles of Terwee et al. (2018) help ensure that a purported relational indicator remains meaningful to the respondent, while the longitudinal interviews of Skjuve et al. (2022) show why a mechanical usage count cannot fully describe the meaning of contact.
Ethical interpretation also requires avoiding individual diagnostic judgments from a research scale. A high AI-attachment score may reflect valued companionship, accessible disability support, creative practice, or interpersonal distress; the score alone does not identify pathology. Clinical claims about dependence require appropriately validated impairment criteria, independent evidence, and specialist oversight. Studies involving minors or vulnerable participants need proportionate consent and safeguarding review. Researchers should disclose vendor relationships, financing, product changes, and the potential consequences of collecting intensely personal disclosures. The integrative assessment proposed in Prinsen et al. (2018) supports transparent reporting of measurement uncertainty; Folk & Dunn (2026) demonstrates the importance of cautious interpretation of social outcomes when definitions and observational design constrain causal inference.
Artificial Era and a Reproducible Research Program
In the Artificial Era, psychological relationships increasingly involve artificial participants whose architectures, access policies, personal memory, and continuity can be modified. This makes relational measurement a study of a structured encounter between a particular human and a historically specific artificial configuration. The framing of Artificial Era and From Homo to Artificial belongs to the author's wider philosophical research program; empirical evaluation here concerns observable human judgments, behaviors, and consequences. The machine-agency framework of Sundar (2020) provides an independently published foundation for investigating how interface cues and perceived agency shape human response. The relevant empirical question is what changes in a human relationship experience when system continuity, initiative, capability, or perceived identity changes while other conditions are controlled.
A cumulative research program can begin with an open construct dictionary, item-provenance archive, and transparent description of the relational target. Independent teams can administer the competing AI Attachment Scales of Kasturiratna & Hartanto (2026) and Cheng & Yu (2026), companionship measures from Banks (2026) and Shen et al. (2026), and related collaboration items from Zabel et al. (2025) in diverse samples. The design should preregister factor structures and predicted discriminant relations, provide anonymized measurement metadata where consent permits, and reproduce the analysis across languages, vendors, and interaction modes. Researchers should identify differences between attachment functions, enjoyment, trust, and practical reliance without requiring that all measures collapse into a single relational index. Open item specifications and comparable target descriptions are especially useful as commercial systems evolve.
A second stage can combine repeated questionnaires, task logs, structured interviews, and ethical experimental perturbations such as announced memory limitations or interface changes. This permits testing whether measurement scores predict subsequent help-seeking, adaptation to error, relational repair, and social contact, and whether results persist after accounting for baseline dispositions. Designs should use sufficient intervals to distinguish temporary novelty from enduring relationships and should report when a system version changes mid-study. Methodological foundations in Vandenberg & Lance (2000) and Hamaker et al. (2015) are crucial for cross-target comparability and within-person inference. The scientific program advances through independently reproducible predictions and clear attribution of empirical results, with philosophical interpretations identified as such and kept distinct from measurement findings.
Conclusion
Measurement of Human–AI Relationships is a developing scientific field whose basic task is to connect defined relational attributes to interpretable observations across people, artificial targets, and time. Existing research shows the relevance of social response, self-disclosure, perceived presence, trust, attachment-like experiences, companionship, and dependence, yet these dimensions are neither interchangeable nor uniformly beneficial or harmful. The scale-development studies of Kasturiratna & Hartanto (2026), Cheng & Yu (2026), Banks (2026), and Shen et al. (2026) provide substantive foundations for measurement while revealing distinct construct definitions and factor structures. Their existence makes a comparative validation agenda possible; it does not make all scores commensurable. A sound instrument identifies its target, timescale, population, conceptual domain, and intended interpretation, then accumulates evidence supporting that interpretation under the conditions where it is used.
The methodological synthesis finds that content validity, response-process evidence, structural assessment, precision, discriminant relations, invariance, and longitudinal stability address different components of scientific measurement. Factor structure must be examined independently of generic positive attitudes toward technology; behavioral indicators require control of instrumental motives; relationship change must be separated from shifting platforms and item meanings. The construct-validation framework of Cronbach & Meehl (1955) and the multitrait–multimethod logic of Campbell & Fiske (1959) remain central to this work, while cross-group methodology in Vandenberg & Lance (2000) and within-person inference in Hamaker et al. (2015) clarify how relational scores may be compared. Psychological interpretation becomes stronger when multiple forms of evidence converge under explicit theoretical predictions.
The research frontier is a transparent, ethically governed, and independently replicated measurement system that records both human experiences and the artificial conditions through which those experiences arise. Its principal priorities are direct comparisons of competing instruments, longitudinal and cross-cultural validation, experimentally informative differences among social presence, competence trust, companionship and attachment, and careful study of social consequences. The designs and empirical findings reviewed in Folk & Dunn (2026) and Fang et al. (2025) illustrate why prospective association, self-selected use, and randomized intervention require separate causal claims. This monograph is a critical narrative synthesis of published scholarship and proposes an evidence-oriented research program; all original empirical conclusions remain attributed to their studies. A precise science of human–AI relationships develops through stable definitions, reproducible methods, and accountable interpretation.
References
- Banks, J. (2026). Measuring machine companionship experiences: Scale development and validation for AI companions. Computers in Human Behavior, 179, Article 108945. https://doi.org/10.1016/j.chb.2026.108945
- Bartneck, C., Kulić, D., Croft, E., & Zoghbi, S. (2009). Measurement instruments for the anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots. International Journal of Social Robotics, 1(1), 71–81. https://doi.org/10.1007/s12369-008-0001-3
- 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
- Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept of validity. Psychological Review, 111(4), 1061–1071. https://doi.org/10.1037/0033-295X.111.4.1061
- Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56(2), 81–105. https://doi.org/10.1037/h0046016
- Cheng, N., & Yu, R. (2026). Measuring and understanding emotional attachment in human–AI relationships. Ergonomics. Advance online publication. https://doi.org/10.1080/00140139.2026.2622539
- Croes, E. A. J., & Antheunis, M. L. (2021). Can we be friends with Mitsuku? A longitudinal study on the process of relationship formation between humans and a social chatbot. Journal of Social and Personal Relationships, 38(1), 279–300. https://doi.org/10.1177/0265407520959463
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- 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 randomized controlled study [Research preprint]. arXiv. https://doi.org/10.48550/arXiv.2503.17473
- Flake, J. K., Pek, J., & Hehman, E. (2017). Construct validation in social and personality research: Current practice and recommendations. Social Psychological and Personality Science, 8(4), 370–378. https://doi.org/10.1177/1948550617693063
- Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276–286. https://doi.org/10.1177/09567976261427747
- 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
- Hamaker, E. L., Kuiper, R. M., & Grasman, R. P. P. P. (2015). A critique of the cross-lagged panel model. Psychological Methods, 20(1), 102–116. https://doi.org/10.1037/a0038889
- Ho, A., Hancock, J., & Miner, A. S. (2018). Psychological, relational, and emotional effects of self-disclosure after conversations with a chatbot. Journal of Communication, 68(4), 712–733. https://doi.org/10.1093/joc/jqy026
- Jian, J.-Y., Bisantz, A. M., & Drury, C. G. (2000). Foundations for an empirically determined scale of trust in automated systems. International Journal of Cognitive Ergonomics, 4(1), 53–71. https://doi.org/10.1207/S15327566IJCE0401_04
- Kasturiratna, K. T. A. S., & Hartanto, A. (2026). Attachment to artificial intelligence: Development of the AI Attachment Scale, construct validation, and the psychological mechanisms of human–AI attachment. Computers in Human Behavior Reports, 21, Article 100912. https://doi.org/10.1016/j.chbr.2025.100912
- Laurenceau, J.-P., Barrett, L. F., & Pietromonaco, P. R. (1998). Intimacy as an interpersonal process: The importance of self-disclosure, partner disclosure, and perceived partner responsiveness in interpersonal exchanges. Journal of Personality and Social Psychology, 74(5), 1238–1251. https://doi.org/10.1037/0022-3514.74.5.1238
- 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
- Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
- Prinsen, C. A. C., Mokkink, L. B., Bouter, L. M., Alonso, J., Patrick, D. L., de Vet, H. C. W., & Terwee, C. B. (2018). COSMIN guideline for systematic reviews of patient-reported outcome measures. Quality of Life Research, 27(5), 1147–1157. https://doi.org/10.1007/s11136-018-1798-3
- Shen, P., Peng, D., & Xu, J. (2026). Artificial intelligence companionship: Conceptualization, dimension identification and scale development. Journal of Business Research, 215, Article 116338. https://doi.org/10.1016/j.jbusres.2026.116338
- 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. Journal of Computer-Mediated Communication, 25(1), 74–88. https://doi.org/10.1093/jcmc/zmz026
- Terwee, C. B., Prinsen, C. A. C., Chiarotto, A., Westerman, M. J., Patrick, D. L., Alonso, J., Bouter, L. M., de Vet, H. C. W., & Mokkink, L. B. (2018). COSMIN methodology for evaluating the content validity of patient-reported outcome measures: A Delphi study. Quality of Life Research, 27(5), 1159–1170. https://doi.org/10.1007/s11136-018-1829-0
- Vandenberg, R. J., & Lance, C. E. (2000). A review and synthesis of the measurement invariance literature: Suggestions, practices, and recommendations for organizational research. Organizational Research Methods, 3(1), 4–70. https://doi.org/10.1177/109442810031002
- Zabel, S., Spannbauer, C., Neef, N. E., & Otto, S. (2025). Being just used or truely understood: A measure of users’ collaboration intensity with chatbots. International Journal of Human-Computer Studies, 203, Article 103587. https://doi.org/10.1016/j.ijhcs.2025.103587
Social Presence and Anthropomorphic Attribution
Social presence refers to the felt availability of another social participant during an exchange, whereas anthropomorphism concerns the assignment of human characteristics to a nonhuman target. These processes can make a chatbot seem approachable or attentive without establishing an enduring relationship. The three-factor account of anthropomorphism in Epley et al. (2007) identifies elicited agent knowledge, effectance motivation, and sociality motivation as explanatory influences, implying that the same system may attract different interpretations depending on human goals and conditions. The Godspeed instrument developed by Bartneck et al. (2009) measures perceived anthropomorphism, animacy, likeability, intelligence, and safety in human–robot encounters; its dimensions are valuable context measures, but they do not collectively define an enduring relational bond. Item wording should therefore distinguish an appearance of sociality from continuing emotional investment.
A researcher can experimentally manipulate voice, memory cues, avatar design, apparent initiative, and source labels, then measure social presence immediately and attachment after a meaningful delay. This separation clarifies whether an intervention changes only a current impression or predicts subsequent decisions about contact. It also makes it possible to detect expectation effects: an interface that calls itself a friend can increase friendship endorsements while leaving willingness to return unchanged. The machines-and-mindlessness program reviewed by Nass & Moon (2000) documented social responses to computational systems, and Sundar (2020) offers an agency-centered vocabulary for contemporary interfaces. Their methodological value lies in separating machine features, user attributions, and interaction outcomes rather than treating an anthropomorphic cue as proof of a relational property.
Measurement of social presence should explicitly control the response channel and timing. Real-time voice, asynchronous text, embodied movement, and simulated typing indicators affect the cues available to respondents. Researchers should include an item about the degree of perceived co-presence, an item about confidence that the system attends to prior information, and independent behavioral tests of that attention, yet validate each item before turning it into a scale. An artificial partner may be rated socially present during a brief, fluent conversation and later forgotten. A strong longitudinal study therefore estimates whether immediate presence predicts later remembered significance after controlling for enjoyment, usefulness, novelty, and baseline anthropomorphic tendency, drawing on the attribution mechanisms investigated by Epley et al. (2007) and Bartneck et al. (2009).