Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Anthropomorphism and Social Presence in Human–AI Encounters

Abstract. Anthropomorphism and social presence are related but distinct phenomena in human–AI encounters. Anthropomorphism concerns the attribution of humanlike characteristics, motives, and mental states to artificial partners; social presence concerns the experienced salience of an apparent social participant. This monograph develops a source-identified critical narrative synthesis of social cognition, the Computers Are Social Actors tradition, the three-factor theory of anthropomorphism, mind-perception research, virtual presence, conversational interface design, human–robot interaction, and modern AI companionship studies. It compares randomized manipulations of framing and social cues, studies of self-disclosure and trust, longitudinal evidence, and quantitative reviews. Particular attention is given to construct validity, competing explanations, causality, heterogeneity, user understanding, and the distinction between human experience and claims about artificial mentality. The study proposes an empirical research program spanning interface, user, interaction, and system levels, situated within the Artificial Era without conflating the publisher's interpretive framework with established experimental findings. This is an original scholarly synthesis rather than a registered systematic review, meta-analysis, or report of newly collected participant data.

Keywords: anthropomorphism; social presence; human–AI interaction; computers are social actors; three-factor theory; social cues; conversational agents; mind perception; human–robot interaction; parasocial interaction; trust calibration; self-disclosure; AI companionship; postsubjective psychology; Artificial Era

The Scientific Object: Humanlike Attribution and Perceived Social Presence

Anthropomorphism in human–AI encounters is the attribution of humanlike qualities, mental states, motives, or social capacities to an artificial partner. Social presence denotes the experienced availability or salience of another social participant during technologically mediated interaction. The two constructs can covary, yet their empirical referents differ: an interface may appear socially present through contingent turn taking without being judged human, and a humanlike avatar may elicit anthropomorphic description while failing to sustain an experience of interpersonal connection. This monograph examines the determinants, mechanisms, measurements, and consequences of their interaction within text chatbots, voice assistants, embodied agents, and companion systems. The object is the human encounter with an artificial interlocutor as measured in perception and behavior, while system competence and architecture are considered explanatory conditions rather than inferred from user reports Epley et al. (2007) Nowak & Biocca (2003) Araujo (2018).

A scientifically useful account has at least four levels. Interface-level cues include names, voices, faces, turn timing, memory statements, self-reference, and linguistic accommodation. Observer-level factors include prior conceptions of AI, social motivation, need for explanation, expectations, individual anthropomorphic disposition, and familiarity. Interaction-level outcomes include felt copresence, closeness, rapport, disclosure, compliance, perceived agency, and trust. System-level observations concern actual response policies, reliability, error patterns, persistence, and the history carried across encounters. Experimental work can separately manipulate the cues and system behavior to establish which social inferences arise from which source. A measure of warmth cannot by itself certify accuracy or mutual understanding Feine et al. (2019) Waytz et al. (2010) Lee & See (2004).

The dossier asks why social responses can arise toward known nonhuman systems; which manipulations reliably change perceived presence; how apparent mind and perceived partner identity shape trust, disclosure, and interaction; what time, culture, task, and individual differences contribute; and which findings generalize to generative conversational systems. Its method is a critical, source-identified narrative synthesis, integrating original controlled studies, foundational conceptual work, measurement scholarship, and recent experimental and quantitative reviews. It introduces no new participants or pooled effect-size estimate. The neighboring problems of attribution of intelligence, enduring attachment, self-disclosure privacy, and relational continuity are treated where they explain the focal phenomenon, with their full treatment belonging to separate research objects Nass & Moon (2000) Folk et al. (2025) Zhang & Sheng (2026).

Historical Roots: Anthropomorphic Projection, Parasociality, and Conversational Machines

Attribution of human characteristics to nonhuman objects has a long intellectual history, but the modern experimental agenda is tied to social cognition and communication research. Horton and Wohl analyzed parasocial interaction as mediated intimacy with personalities who appear to address an audience; their 1956 concept offers a historical comparator for the subjective experience of an apparently responsive partner. A chatbot differs from broadcast media by generating contingent replies, altering the temporal structure of participation. Conceptual continuity is therefore useful only when differences in reciprocity, persistence, audience awareness, and perceived agency remain visible. A user can address a system socially without assuming another human is behind it, and the same user can revise these assumptions when the interface or task changes Horton & Wohl (1956) Nowak & Biocca (2003).

Weizenbaum's 1966 ELIZA demonstrated how keyword-triggered transformations could support a recognizable conversational form. Its original paper described decomposition rules, minimal context, response assembly, and script structure, providing a crucial lesson about the distinction between an interaction's human-readable form and the computation that generates it. The historical importance of ELIZA is the demonstration that a simple linguistic architecture can furnish sufficient conversational cues for interpretable exchange. Contemporary language models have different internal architectures and markedly broader generative capacities; it would be a category error to read ELIZA's technical mechanism back into all present AI systems or to infer any particular inner state from conversational fluency alone Weizenbaum (1966) Luger & Sellen (2016).

Work on social robots and media further challenged the assumption that realistic appearance is a prerequisite for social behavior. A plain text interface can cue politeness; a robot body can alter attention and engagement; and computer labels can modify evaluative anxiety. The historical line is cumulative in its experimental questions rather than uniform in its effects: social response, inferred humanlikeness, empathy, behavioral compliance, and experienced presence have different dependent variables. The appropriate history traces when a specific observation was documented and which construct the investigators measured, instead of treating successive technological generations as a single ascending scale of artificial personhood Nass & Moon (2000) Kiesler et al. (2008) Lucas et al. (2014).

Computers Are Social Actors and the Media Equation

The computers-are-social-actors research program showed that participants can apply social conventions to computers even while explicitly knowing that the machine is a machine. Nass and Moon's review integrated experiments on politeness, reciprocity, social categories, perceived expertise, and apparent computer personality. The social-response account is important because it makes a prediction about interactional scripts rather than requiring a stable metaphysical belief that the computer possesses human mentality. Social categorizations may be automatic, habitual, situationally invited, or strategically useful, and their occurrence varies with how the question is framed and how behavior is recorded Nass & Moon (2000) Epley et al. (2007).

The earlier media-equation tradition studied responses to relatively simple media interfaces. Its lesson for contemporary human–AI studies is methodological: responses attributed to artificial intelligence must be compared against reactions generated by ordinary interface conventions, conversational politeness, and explicit source labels. A highly capable generative model may increase cue richness, but if an effect appears with a simple script, causal explanation must include the script itself. Conversely, when long-context responses or adaptive interaction produce outcomes that static scripts cannot match, the comparison identifies genuine differences in affordance without establishing consciousness or social reciprocity Nass & Moon (2000) Weizenbaum (1966) Bickmore & Picard (2005).

The tradition has also invited debate about conscious anthropomorphism versus mindless application of learned rules. These are not interchangeable measures: one is an articulated attribution, another a mechanism proposed for social behavior. An individual can deny believing that software has feelings while showing courtesy to it; another can endorse mental-state descriptors without changing behavior. An adequate experimental battery pairs explicit attribution ratings with behavior under meaningful choices, records participant awareness of system identity, and measures expectations before and after interaction. Such separation prevents an apparent contradiction between what people say and what they do from being explained away through a single broad concept Nass & Moon (2000) Waytz et al. (2010) Ho et al. (2018).

The Three-Factor Theory: Knowledge, Effectance, and Sociality

Epley, Waytz, and Cacioppo articulated anthropomorphism as a variable outcome of three determinants: elicited agent knowledge, effectance motivation, and sociality motivation. Elicited agent knowledge concerns accessibility and applicability of human-centered concepts when explaining an unfamiliar agent. Effectance motivation concerns the desire to predict and control another agent's behavior. Sociality motivation concerns the desire for connection and affiliation. The theory generates identifiable moderators instead of assuming that every humanlike feature will have the same effect for every observer. Its predictions concern anthropomorphic inference, not direct evidence of a machine's experience Epley et al. (2007).

The account directs experiments toward causal separation. Researchers can hold the model's outputs constant while manipulating what participants are told about its origins; vary response predictability while keeping stylistic friendliness stable; or compare social-need conditions with an identical artificial conversational partner. A change in humanlike judgment after altered explanatory uncertainty would support an effectance-related pathway, whereas a change under heightened affiliation motives would support a sociality-related pathway. Such manipulations require checks that participants actually experienced the intended uncertainty or need, and that the same manipulation did not simultaneously change task difficulty, reward expectation, or general mood Epley et al. (2007) Folk et al. (2025).

The theory also supplies an important boundary on overinterpretation. Greater perceived sociability need not be an indication that social isolation caused it, and correlations between loneliness and AI attachment cannot identify social motivation without longitudinal or experimental evidence. In the 2025 longitudinal companion-chatbot study by Guingrich and Graziano, a greater desire for social connection, but not loneliness, was associated with greater anthropomorphism. That distinction illustrates why closely neighboring human states require separate measurements. Studies should identify immediate need, enduring disposition, and relationship context instead of collapsing them into a single vulnerability score Guingrich & Graziano (2025) Waytz et al. (2010).

Mind Perception: Agency, Experience, and Attribution Boundaries

Gray, Gray, and Wegner's mind-perception research separated judgments along agency, such as planning and self-control, and experience, such as hunger or feeling. These dimensions matter for AI because observers may grant apparently intelligent software considerable agency while withholding experiential capacities. An interface that produces thoughtful explanations, remembers context, or adapts its language can move some assessments of agency without moving beliefs about subjective feeling in the same direction. A questionnaire that aggregates all such items into one undifferentiated anthropomorphism score obscures the result. For comparative work, researchers should state whether a question concerns knowledge, intentionality, moral responsibility, emotional experience, or interpersonal presence Gray et al. (2007) Waytz et al. (2010).

Shank and colleagues investigated emotional responses when people report perceiving mind in AI, showing that such encounters have recognizable affective consequences. The study's evidence is about reported human perceptions and emotions in AI-related encounters. It neither verifies the mental states attributed to AI nor establishes a single response pattern for all users. The distinction allows rigorous integration of subjective accounts within a psychological science: a reported feeling is evidence that a person had an experience of interaction, while evidence about the attributed agent requires independent operational tests. Confusing these evidentiary targets makes both social-presence research and machine-capability assessment less informative Shank et al. (2019) Gray et al. (2007).

The broader anthropomorphism literature distinguishes individual differences in propensity from specific encounters. Waytz, Cacioppo, and Epley presented the Individual Differences in Anthropomorphism Questionnaire and reported associations between anthropomorphic tendency and judgments involving moral concern, responsibility, trust, and social influence. These are associations shaped by measurement and sample, not a universal diagnostic trait. In contemporary AI work, an individual-differences measure can serve as a moderator of experimentally manipulated conversational conditions; investigators should test whether its factor structure and item interpretation transfer from general nonhuman targets to particular generative systems Waytz et al. (2010) Folk et al. (2025).

Linguistic Social Cues and Conversational Contingency

An artificial interlocutor expresses social information through lexical choice, conversational sequencing, hedging, self-reference, turn timing, address terms, and apparent responsiveness. Such cues may prompt users to apply scripts ordinarily used for interpreting people, even where no humanlike face appears. Feine and colleagues organized conversational-agent social cues into a taxonomy that distinguishes verbal, visual, auditory, and invisible social cues. The taxonomy is valuable for designing controlled contrasts: describing a system as friendly is insufficient unless the studied dimensions, observable stimuli, and actual conversational affordances are specified. Researchers need to separate linguistic warmth, informational relevance, state tracking, personalization, and the perception of a continuing partner Feine et al. (2019) Araujo (2018) Nass & Moon (2000).

Text exchanges make this separation technically possible. A study may standardize factual content while manipulating greeting style, first-person language, conversational acknowledgments, or expressions of apparent concern. Another design may hold style constant while varying response relevance or memory of earlier turns. Such factorial contrasts reveal whether social presence is sensitive mainly to conventional linguistic politeness, to contingent participation, or to a combination. Perceived responsiveness is particularly important because a generic expression of sympathy and a context-specific, correct acknowledgment can be experienced differently. The experiment should compare participants' actual transcripts, inspect model output variability, and predefine a measure of interaction quality rather than retrospectively classifying more pleasing replies as socially present Araujo (2018) Bickmore & Picard (2005) Ho et al. (2018).

Generative dialogue changes the distribution of possible cues. Language models can produce diverse utterances conditioned on conversational history, and identical prompts may yield different transcripts when outputs are sampled. This increases ecological realism and introduces a treatment-fidelity problem: a nominal condition such as 'high social presence' may change response length, correctness, or task completion as well as tone. Researchers should freeze or audit generated outputs, document system versions, define prohibited response categories, and quantify whether the manipulation holds across conversations. The central interpretive issue remains the relation between observable interaction and human judgment; fluent dialogue is evidence of a capacity to generate context-sensitive text, while mental-state attribution is an additional psychological event Luger & Sellen (2016) Feine et al. (2019) Epley et al. (2007).

Voice, Visual Embodiment, and Nonverbal Behavior

Social cues are distributed across modalities. Voice can carry conversational timing, emotional prosody, accent, and continuity; embodied agents add gaze, posture, facial motion, and apparent orientation toward the user; text interfaces depend more heavily on linguistic contingency and framing. Research on human–robot interaction has investigated whether anthropomorphic behavior and physical form alter reactions to artificial partners. Kiesler and colleagues' study illustrates the need to examine both the properties of the agent and the interpretation supplied by observers. A robot's presence is a physical affordance, but a report of closeness is a participant outcome requiring independent measurement Kiesler et al. (2008) Feine et al. (2019).

Embodiment can increase the salience of a putative social partner while also amplifying expectations the agent cannot meet. The familiar uncanny-valley proposal, originally made by Mori and later translated and discussed in robotics research, hypothesizes a nonmonotonic affective response when human resemblance becomes sufficiently close to evoke perceptual mismatch. The proposal has inspired extensive experimental work, but a particular avatar's unsettling appearance does not establish a universal valley for all users or modalities. Studies need to vary appearance independently of movement, voice, task competence, and prior familiarity, because a richly humanlike design can alter several of those cues together Mori et al. (2012) Nowak & Biocca (2003).

Nowak and Biocca directly examined agency and anthropomorphism in virtual environments and measured telepresence, copresence, and social presence. Their results complicate a simple assumption that greater humanlike realism always creates greater presence; lower-anthropomorphism conditions could outperform more humanlike alternatives on some measures. An agent can therefore be socially effective through legibility, appropriate responsiveness, or expectation fit without maximal resemblance. A useful multimodal experiment compares text-only, voice-only, embodied, and combined configurations with comparable task information, while making the possible effects of modality on accessibility and usability part of the analysis Nowak & Biocca (2003) Kiesler et al. (2008) Mori et al. (2012).

Partner Identity, Source Framing, and Expectations

Beliefs about who or what produced a message can change interaction independently of its literal content. An anthropomorphic name, explicitly artificial label, human avatar, or declared organizational sponsor creates expectations about competence and social intent before a response is evaluated. Araujo's experiments investigated anthropomorphic conversational-agent design cues and communicative agency framing, showing why perceptions of the agent and the organization must be analyzed separately. In observational settings, labels often covary with design quality and user selection. Randomized framing with matched transcripts isolates a narrower causal question: what changes when the same behavior is attributed to different sources? Araujo (2018) Epley et al. (2007).

Labels can also change how participants weigh privacy and judgment. Lucas and colleagues examined disclosure to virtual humans and identified increased willingness to disclose under an ostensibly computer-mediated interaction that reduced fear of negative human evaluation. Ho, Hancock, and Miner compared disclosure after conversations with human and chatbot partners and documented psychological, relational, and emotional consequences. The findings illustrate that reduced evaluation apprehension and increased social warmth can arise through different processes. Humanlike appearance, known artificial identity, and perceived confidentiality therefore belong in distinct experimental conditions rather than being combined under a single 'humanization' index Lucas et al. (2014) Ho et al. (2018).

Partner identity is not fixed at exposure. Users may begin with a categorical assumption, infer a richer partner from continued response, encounter a conspicuous failure, and revise their judgments. A longitudinal design should measure prior expectations, repeated perceptions, context of use, and what participants believe about data retention and model updates. It should also record system changes that could invalidate a uniform identity label over time. The outcome of interest is the trajectory of perceived social presence and humanlike attribution, not a laboratory assumption that users retain the initial explanation throughout weeks or months of contact Luger & Sellen (2016) Bickmore & Picard (2005) Guingrich & Graziano (2025).

Anthropomorphic Categories, Uncanny Expectations, and Social Norms

Categorization research shows that perceived group membership can influence judgments of artificial social partners. Eyssel and Kuchenbrandt examined the relation between robot group membership and anthropomorphism. Such findings locate the phenomenon within broader social psychological processes: observers interpret a partner partly through familiar ingroup and outgroup distinctions, and a robot's apparent social position can shape the vocabulary used to describe it. Generalization demands care because national labels, social role cues, and robot design are culture- and setting-dependent manipulations Eyssel & Kuchenbrandt (2012) Nass & Moon (2000).

Humanlike cues can create a form of interactional contract. An assistant that claims warmth, attentiveness, or memory makes conversational promises that users may test by later reference, requesting help, or disclosing something important. When actual behavior violates the induced expectation, disappointment can be attributed to the agent's social character even if the technical cause is context truncation, retrieval failure, or a changed software version. Luger and Sellen's qualitative investigation of the gulf between expectation and experience with conversational agents is therefore relevant to social presence as an evolving appraisal, especially when participants are recruited from people with strong prior expectations about intelligent assistance Luger & Sellen (2016) Bickmore & Picard (2005).

Normative judgments also operate on a different plane from immediate social feelings. A participant may enjoy a humanlike interface and nevertheless judge its information unreliable, intrusive, or inappropriate in a high-stakes context. Product decisions that maximize engagement can conflict with user needs for transparency and calibration. Empirical research should treat accurate disclosure of artificial identity and task limitations as independent design variables. An ethical evaluation investigates actual participant comprehension, opportunities to disengage, freedom from manipulative pressure, and the consequences of mistaken assumptions; subjective connection alone is insufficient as a global measure of interaction quality Lee & See (2004) Waytz et al. (2010) Zhu & Broadbent (2025).

Social Presence, Copresence, and Relational Reciprocity

Social presence is a broad family of constructs used to describe the felt presence or salience of another participant in mediated interaction. Copresence typically emphasizes the sense of sharing an encounter with another; telepresence concerns experience of being in a mediated environment. Nowak and Biocca's study operationalized these dimensions separately and demonstrated why they should not be collapsed into general satisfaction or realism. In human–AI encounters, the presence claim refers to the human's reported experience of the encounter. Whether the AI system maintains a persistent user model, takes initiative, or reciprocates in any functional sense is separately testable Nowak & Biocca (2003) Bickmore & Picard (2005).

Reciprocity likewise has multiple observable meanings. Linguistic reciprocity may involve taking turns, asking follow-up questions, or mirroring the degree of self-disclosure. Operational reciprocity may involve adaptation to earlier interaction, use of retained information, and appropriate adjustment when the user corrects an error. Social reciprocity as experienced by the human is an interpretation built partly from these patterns. A study examining relational effects should describe which sense is measured, whether interaction histories remain available, and what happens when the artificial partner's stored context is replaced or removed Ho et al. (2018) Guingrich & Graziano (2025) Luger & Sellen (2016).

A longitudinal research program can test whether repeated contingent responses increase social presence beyond what would be expected from novelty or increasing familiarity with an interface. Participants can be randomly assigned to matched tools differing in memory availability, response sensitivity, or social cues, with preregistered measures of perceived presence and actual engagement. Repeated observations allow individual trajectories, reversals, and heterogeneity to emerge. Distinguishing task utility from perceived intimacy matters: an agent can be highly useful without feeling close, and emotionally engaging without being competent at the task for which a user needs assistance Bickmore & Picard (2005) Folk et al. (2025) Guingrich & Graziano (2025).

Trust, Perceived Warmth, and Appropriate Reliance

Anthropomorphism can influence trust judgments by encouraging users to import social expectations from human relationships into evaluations of an artificial system. Yet trust is multidimensional: perceived benevolence, willingness to rely, expected accuracy, and decision delegation need not move together. Lee and See define the design problem of automation trust in terms of appropriate reliance: users must neither overrely on unreliable automation nor reject reliable assistance simply because it is automated. A socially warm chatbot may attract more attention while increasing or decreasing verification depending on user goals, stakes, and previous performance. Consequently, a trust questionnaire should be paired with tasks where objective system reliability and the cost of wrong reliance are known Lee & See (2004) Waytz et al. (2010) Araujo (2018).

Experimental identification requires independently varying the quality of evidence and the social cue. For example, users might see accurate and inaccurate answers accompanied by either warm or neutral wording. The dependent variables should include acceptance of correct advice, rejection of incorrect advice, verification behavior, calibration of confidence, and perceptions of relational support. A purely evaluative rating after a single friendly conversation cannot reveal whether participants would have relied appropriately in an consequential task. Researchers should also record whether users are informed of the AI partner's limitations and whether the system presents uncertainty in a comprehensible form Lee & See (2004) Feine et al. (2019).

Anthropomorphic warmth is therefore a candidate mechanism for reliance and social engagement rather than a universal marker of successful human–AI cooperation. The same cue may assist novices in sustaining difficult learning tasks and impair critical scrutiny when the interface implies unwarranted authority. A causal account needs variation across actual reliability levels, task domains, demographic groups, and prior AI literacy, with direct tests of whether social perceptions mediate measured decisions. The combined program connects social presence to human factors while preserving the scientific distinction between subjective trust and performance-calibrated behavior Lee & See (2004) Epley et al. (2007) Luger & Sellen (2016).

Self-Disclosure, Evaluation Apprehension, and Privacy

Self-disclosure to an artificial partner is a behavioral outcome with multiple motivations. A person may reveal experiences to seek useful advice, to gain perceived empathy, because a computer seems less judgmental than a person, or because the interface creates a strong impression of interpersonal attention. Lucas and colleagues' virtual-human experiments found that beliefs about being evaluated by a human mattered for willingness to disclose. Ho, Hancock, and Miner examined relational and emotional reactions after conversations with chatbot and human partners. Taken together, these studies show that disclosure cannot be explained by human likeness alone; audience beliefs and consequences of evaluation must be explicitly incorporated Lucas et al. (2014) Ho et al. (2018).

Recent evidence complicates the intuition that greater human resemblance necessarily facilitates candor. Zhu and Broadbent's randomized study compared modes of computerized psychosocial assessment and found outcomes sensitive to both social desirability and the nature of the question. Increased social presence can encourage engagement yet also increase impression management when participants anticipate being judged. Investigators should measure response completeness, concrete disclosure content, socially desirable answering, perceived privacy, and user understanding of storage and third-party access. The number of utterances, emotional tone, and self-reported comfort are different outcomes Zhu & Broadbent (2025) Ho et al. (2018).

The ethical challenge becomes acute when a socially vivid system encourages discussion of vulnerabilities. Disclosure involves consequential information flows that the interface may conceal behind a conversational surface. Scientists should distinguish what users feel about privacy from actual retention, vendor access, deletion rights, and onward use. Research designs can examine comprehension after clear data notices, contrasts between known artificial and human partners, and whether measured social presence predicts disclosure after controlling for task benefit. A claim that an AI feels safe to talk to refers to the user's experience; a claim about the safety of disclosing personal information requires independently documented technical and institutional safeguards Zhu & Broadbent (2025) Lucas et al. (2014) Lee & See (2004).

Companionship, Affiliation, and Social Connection

Social needs can guide anthropomorphism, but correlational studies must distinguish short-term desire for connection, loneliness, social support, and enduring anthropomorphic tendency. Folk, Heine, and Dunn reported two experiments totaling 1,274 participants and found that individual differences in anthropomorphism help explain when artificial companions are experienced as socially connecting. The findings support heterogeneous responses rather than a universal benefit or deficit of AI companionship. People who regard artificial partners as more humanlike may be more likely to report connection, but the result does not demonstrate that the partner experiences reciprocal connection or that long-term effects are identical across circumstances Folk et al. (2025) Epley et al. (2007).

Guingrich and Graziano conducted a 21-day randomized companion-chatbot study with 183 participants, contrasting chatbot use with a word-game comparison. The authors reported no statistically significant overall shift in social health across their design, while anthropomorphism correlated with individual differences in social motivation and mediated some observed relationships. Specifically, desire for social connection rather than loneliness was associated with greater anthropomorphism in their analysis. These observations illustrate why interpretations should focus on the actual intervention, measures, comparison activity, and study duration. The absence of a significant average shift does not rule out heterogeneity, nor does exploratory mediation prove a causal psychological pathway Guingrich & Graziano (2025).

Companionship research also needs a temporal dimension. An initial sense of presence can be driven by novelty, reduced social evaluation, or playful exploration. Sustained effects may depend on access, predictability, remembering past exchanges, crisis responses, and the user's external relationships. Researchers should follow trajectories rather than extrapolating a single laboratory session into a stable attachment bond. Both beneficial and problematic outcomes can arise in different participants; carefully characterized contexts, baseline social functioning, repeated measurement, and real-world behavior are necessary before assigning direction to population-wide social consequences Folk et al. (2025) Guingrich & Graziano (2025) Bickmore & Picard (2005).

Expectations, Breakdowns, and Perceived Relational Repair

Social presence is sensitive to continuity and the perceived coherence of an interlocutor. Bickmore and Picard investigated relational agents designed to sustain long-term human–computer interaction, providing a foundation for examining repeated social rituals, consistency, and remembered preferences. These features can strengthen the user's interpretation of a persistent conversational partner when they are implemented reliably. However, a greeting that refers to a prior exchange may be generated from stored text, an external retrieval service, or an internal state, and each architecture has different failure conditions. A psychological study should document the actual source of continuity rather than inferring an enduring agent from the user's familiarity with a name Bickmore & Picard (2005) Luger & Sellen (2016).

Breakdowns test relational expectations. A system that forgets information, gives an irrelevant comforting message, or misattributes earlier content may prompt disappointment that resembles a social violation, particularly when its prior interface invited humanlike expectations. Luger and Sellen documented user frustration with the mismatch between conversational agents' apparent capabilities and practical performance. The causal question is whether humanlike framing makes the same technical error feel more socially consequential. Such a question is testable by randomizing framing and holding the error constant, with separate measures of competence loss, perceived disregard, and willingness to resume interaction Luger & Sellen (2016) Araujo (2018).

Repair should be analyzed through observable outcomes. An apology, correction, acknowledgment of a forgotten detail, or transfer to human support may alter trust and presence in different ways. A system might restore task accuracy without restoring felt rapport, or display compelling remorse language without correcting the underlying technical problem. Studies should therefore measure successful correction, comprehension of error causes, participant autonomy, and follow-up reliance over time. A further comparison could isolate the effect of transparent error explanations from mere increases in politeness. This makes relational repair an experimental object rather than an anthropomorphic metaphor Lee & See (2004) Nass & Moon (2000) Bickmore & Picard (2005).

Individual Differences, Social Categories, and Population Boundaries

Individual-differences research treats anthropomorphism as a tendency that varies across observers and contexts. Waytz, Cacioppo, and Epley developed the Individual Differences in Anthropomorphism Questionnaire to investigate stable variation in attributing human qualities to nonhuman agents. Applying such measures to AI requires attention to target specificity. A participant may readily anthropomorphize animals while evaluating chatbots as information retrieval interfaces, or may treat an artificial companion as socially present while maintaining little general belief that machines have experiential minds. Analyses should test measurement invariance, moderator interactions, and whether general trait scores predict behavior beyond prior familiarity and task relevance Waytz et al. (2010) Folk et al. (2025).

Social categories and cultural expectations may shape these interpretations. Eyssel and Kuchenbrandt studied whether robot ingroup categorization influenced anthropomorphism, illustrating that perceived machine qualities may depend on familiar intergroup mechanisms. However, the cultural interpretation of a voice, body, title, role, or politeness strategy can vary, and the underlying study does not justify universal claims about every language community. Cross-cultural work needs locally adapted stimuli and validated measures, with direct attention to translation, education, technological availability, and gendered or professional roles encoded in artificial voices and visual identities Eyssel & Kuchenbrandt (2012) Feine et al. (2019).

Developmental generalization also requires its own evidence. An experimental effect in adult university participants does not automatically apply to children, older adults, or people using assistive agents in settings where access to human support differs. Heerink and colleagues developed the Almere model for acceptance of assistive social-agent technology among older adults, emphasizing perceptions relevant to particular populations and practical uses. Their work is informative for designing context-sensitive measurements but does not collapse acceptance, benefit, trust, and psychological social presence into an identical construct. Sampling must include the intended user group, accessibility needs, and time horizon of the intervention Heerink et al. (2010) Kiesler et al. (2008) Waytz et al. (2010).

Conceptual Validity and Measurement Architecture

A serious program of measurement begins by defining the construct before selecting the instrument. Anthropomorphism can be measured as explicit attribution of humanlike properties, inferred mind, perceived agency, or implicit adoption of social expectations. Social presence concerns the subjectively experienced salience of another social actor; copresence, rapport, enjoyment, acceptance, and behavioral engagement are related but distinct outcomes. Nowak and Biocca measured multiple presence dimensions rather than treating them as synonyms. Waytz, Cacioppo, and Epley's individual-differences work measured general anthropomorphic tendency. Those instruments ask different scientific questions, and their scores should never be pooled as though they represented one invariant variable Nowak & Biocca (2003) Waytz et al. (2010).

Construct validation should investigate item content, internal structure, reliability, discriminant relations, and sensitivity to specific experimental manipulations. Asking participants whether an assistant 'understood me' may produce an answer about response correctness, emotional validation, or attribution of conscious understanding, depending on instructions and context. Researchers can conduct cognitive interviews, compare latent-variable models, and test whether the same wording retains its meaning across text and embodied systems. The separation between judged agency and experience identified by Gray and colleagues provides a theoretical basis for multidimensional measures Gray et al. (2007) Feine et al. (2019).

Behavioral metrics should complement self-report rather than act as unquestioned substitutes. Number of messages can reflect task difficulty, confusion, or social attraction; repeated voluntary use can signal utility and convenience; disclosure content may indicate situational trust but also a misconception about privacy. A transparent measurement model links every proposed latent construct to multiple observations and to plausible competing causes. Researchers should prespecify primary outcomes and state which interpretations would be disconfirmed by null or opposite findings, especially when the interface makes numerous correlated social cues available Ho et al. (2018) Lee & See (2004) Folk et al. (2025).

Experimental Identification and Factorial Manipulations

Randomized studies offer the clearest method for identifying whether a particular social cue changes a participant's judgments or decisions. One experimental arm can vary artificial versus human authorship labels while presenting identical content; another can independently vary linguistic warmth and answer correctness. Random assignment makes the contrast interpretable, provided treatment versions are stable, assignment is concealed where feasible, attrition is characterized, and outcome measures are established before examining results. This design distinguishes an effect of partner framing from an effect of actual interactive quality Araujo (2018) Lucas et al. (2014) Zhu & Broadbent (2025).

Embodied systems need a stronger control for correlated changes. A realistic avatar can alter perceived age, gender, status, attention, authority, and emotional accessibility as well as general anthropomorphism. A careful design crosses appearance with conversational competence and voice, or uses matched stimuli with validated manipulation checks. Where systems generate live responses, transcripts, latency, intervention fidelity, software versions, and model configuration should be logged. Otherwise the assigned 'social presence' manipulation may also shift the quality or frequency of answers, preventing the presumed mechanism from being isolated Nowak & Biocca (2003) Kiesler et al. (2008) Feine et al. (2019).

Experimental effects require clear denominators. Researchers should report the randomized sample and analyzed sample, exclusions and missingness, outcome distributions, uncertainty intervals, and any multiple-comparison strategy. Particularly for subjective Likert measures, estimated effect size and interpretability matter more than merely classifying a result as significant. Mechanism-specific theories should generate interactions, for example a sociality-motivation effect concentrated among participants high in affiliation needs, rather than repeatedly relabeling any increase in satisfaction as anthropomorphism. Strong alternatives include testing whether usefulness or reduced evaluation apprehension explains the same outcome Epley et al. (2007) Lucas et al. (2014) Waytz et al. (2010).

Longitudinal Change and Limits of Mediation

Human–AI relationships are often studied over a duration that is much shorter than the phenomena they are invoked to explain. Repeated interaction creates opportunities for familiarity, perceived memory, frustration, selective abandonment, and changing expectations. Bickmore and Picard's relational-agent work established that sustained encounters can be experimentally operationalized rather than inferred from a one-time questionnaire. Recent companion-chatbot research similarly demonstrates the value of assessing baseline measures and repeated outcomes across weeks. However, exposure differs from adherence: participants assigned to an agent may vary greatly in how often and why they use it Bickmore & Picard (2005) Guingrich & Graziano (2025).

A longitudinal randomized comparison estimates effects of assignment under the study's actual follow-up and compliance conditions. It does not automatically identify the separate causal effect of anthropomorphism, which is often a measured, post-treatment belief. Guingrich and Graziano examined mediation in a 21-day randomized study; the design supports claims about the randomized intervention more directly than claims about a psychological mediator whose values were not separately randomized. Analysts should distinguish mediation assumptions, baseline confounding, time-varying influences, and competing pathways through task utility or emotional engagement. A mediation diagram is explanatory only when its assumptions are defensible Guingrich & Graziano (2025) Folk et al. (2025).

Attrition is especially informative in extended social interactions because users who discontinue the chatbot may differ systematically in disappointment, social needs, literacy, and prior experience. A robust protocol reports dropout reasons, longitudinal missing-data procedures, device and model changes, and whether the observed effects persist after access ends. Ideally it also examines life outside the interface: use of human support, real-world behavior, and independent measures of well-being. A sustained increase in subjective closeness may have different consequences for someone whose human relationships flourish and for someone whose opportunities for them deteriorate; the evidence must be designed to characterize those contexts Folk et al. (2025) Guingrich & Graziano (2025) Bickmore & Picard (2005).

Evidence Synthesis, Replication, and Heterogeneity

The empirical literature contains several partly overlapping traditions: human–computer interaction, social psychology, virtual presence, social robotics, health communication, and generative-AI companionship. Their experiments vary in agent type, participant recruitment, duration, outcome definitions, and the incentives participants face. A synthesis should therefore compare designs and dependent variables before comparing apparent effect directions. For example, a positive effect on company perceptions in service-chatbot studies, enhanced self-disclosure under low human-evaluation expectations, and a null average change in social health over three weeks are different estimands. They provide complementary constraints on theory rather than mutually canceling answers Araujo (2018) Lucas et al. (2014) Guingrich & Graziano (2025).

Zhang and Sheng's meta-analytic structural-equation approach synthesized a body of chatbot adoption studies and highlights the important role of anthropomorphism and social presence within acceptance models. Its underlying literature is dominated by contexts and outcomes relevant to adoption, so the evidence should not be treated as a pooled estimate of lasting social benefit or therapeutic effectiveness. The monograph reports quantitative claims only where attributable to a specific primary article or published synthesis; it does not calculate new pooled estimates. When studies operationalize anthropomorphism differently, heterogeneity can arise through measurement rather than true psychological moderation Zhang & Sheng (2026) Waytz et al. (2010).

Replication priorities include direct tests of identical interface cues in new populations, conceptual tests across modality and task type, and systematic reporting of unsuccessful attempts. Publication bias, selective outcome reporting, novelty effects, and rapidly changing systems can all distort conclusions. Independent reproducibility requires shared stimuli or executable interaction protocols, clear participant inclusion criteria, redacted but auditable transcripts, and sufficiently detailed descriptions of the platform. Since proprietary conversational systems may change without notice, an evaluation should record model identity and date while also seeking mechanism-level outcomes that could generalize across generations of software Feine et al. (2019) Nowak & Biocca (2003) Folk et al. (2025).

Research Ethics, Agency, and Design Accountability

Anthropomorphic design creates responsibilities because users may act on the apparent social character of a system. Clear machine-identity disclosure, accurate explanation of capabilities, control of retained information, and accessible escalation mechanisms protect users' ability to understand the encounter. These principles are testable at the behavioral level: can participants accurately state that the partner is artificial; do they understand what information persists; can they identify uncertainty and withdraw; and do they seek appropriate human help when tasks exceed the system's capabilities? Such outcomes deserve measurement alongside connection and enjoyment Lee & See (2004) Luger & Sellen (2016).

Study populations deserve attention when disclosure concerns intimate, emotional, or clinically relevant information. Existing results on self-disclosure demonstrate that increased conversational willingness may arise through reduced judgment fears or persuasive social cues. Neither outcome guarantees confidential handling nor appropriate response to high-risk situations. Experimental protocols should avoid covertly inducing disclosure beyond what consent covers, limit collection of unnecessary personal information, and preserve a meaningful alternative to participation. Real-world deployments should examine the consequences of false interpersonal expectations and whether user agency remains intact in prolonged encounters Lucas et al. (2014) Ho et al. (2018) Zhu & Broadbent (2025).

Evaluation of benefits should distinguish felt presence from decision quality, autonomy, and sustained human welfare. A conversational assistant could make a difficult interaction more approachable, but equally could encourage misplaced dependence through declarations of exclusivity or simulated intimacy. These possibilities are hypotheses requiring clear comparison groups and careful follow-up; they should neither be promoted as universal benefits nor treated as universal harms before evidence is established. Scientific and design accountability requires recording how social cues are generated, what they cause users to believe, and the functional realities those beliefs imply Folk et al. (2025) Guingrich & Graziano (2025) Epley et al. (2007).

Artificial Era: The Psychological Encounter as a Relational Configuration

The Artificial Era, in the editorial framework of Aisentica, describes a historical expansion in the kinds of systems with which humans organize cognition, activity, and social life. In the present empirical dossier, that framework motivates a relational unit of analysis: the encounter among a human, an artificial system, an interface, the user's expectations, stored interaction history, and the institutions governing access. The psychological phenomenon of anthropomorphism belongs to the organization of this encounter because a human response depends on the experienced features of that configuration. The term 'relational configuration' here names an explanatory perspective and its testable components; it does not constitute an experimental finding that every artificial system possesses subjective interpersonal experience Epley et al. (2007) Feine et al. (2019) Bickmore & Picard (2005).

Postsubjective Psychology invites researchers to study how an apparently interpersonal event is produced across participants, media, scripts, remembered exchanges, and situational demands. One plausible hypothesis is that some social-presence judgments arise from stable relations among these components rather than from the visual realism of an isolated agent. It can be tested by interventions that preserve dialogue while changing source labels, preserve labels while altering memory, or remove persistent context without changing surface style. The comparative outcome is how much perception and behavior shift when a single component is perturbed. Such a design turns the theoretical horizon into operational questions without presupposing that social presence has a single location or mechanism Nowak & Biocca (2003) Araujo (2018) Guingrich & Graziano (2025).

An encounter may be psychologically significant even when participants adopt different interpretations of the artificial partner. Some treat a chatbot as a convenient instrument; others describe a feeling of companionship or presence. These responses are themselves part of the empirical record. Claims about a system's continuing identity, intentional action, or experience require their own definitions and observations. The methodological task is to retain the distinction between an observed human social experience and an account of how the artificial participant is organized. This allows the science of human–artificial relations to examine genuine social transformation while subjecting each empirical claim to ordinary comparative and causal tests Gray et al. (2007) Folk et al. (2025) Lee & See (2004).

Competing Mechanisms and Critical Scientific Discussion

Several mechanisms can produce similar self-reports of social connection. Anthropomorphic categorization may make human concepts accessible; reduced fear of judgment may encourage disclosure; conversational relevance may create a perception of attention; utility may lead to repeated exposure; and a visible humanlike embodiment may modify expectations before speech begins. The three-factor account of anthropomorphism privileges accessible agent knowledge, effectance, and sociality motives, while the Computers Are Social Actors tradition investigates application of social scripts to media. These explanatory programs are compatible in some predictions and distinguishable in others. An effective experiment must locate where their predicted effects diverge Epley et al. (2007) Nass & Moon (2000) Lucas et al. (2014).

One discriminating design could cross identical human-versus-AI labels with actual adaptive memory and with a manipulation of social need. If labeling changes humanlike ratings but memory mainly changes continued engagement, attribution and functional support would be partly separable. If social need selectively affects anthropomorphic ratings under unchanged behavior, the motivational account receives support. If politeness effects appear with an obviously simple machine, the social-script mechanism gains credibility. Conversely, if interactions require durable contingent adaptation, a purely static-surface explanation would be incomplete. Tests need predeclared predictions and observations that could challenge each account, including outcomes on which an expected effect is absent Epley et al. (2007) Nass & Moon (2000) Bickmore & Picard (2005).

The evidence also warns against equating social connection with beneficial social consequences. A brief experimental increase in connection, a longer period without average change, and an effect on disclosure or perceived warmth answer different questions. Results should be interpreted according to exposure duration, participant selection, reference condition, exact measurement, and whether the intervention was randomized. The studies by Folk and colleagues, Guingrich and Graziano, and Zhu and Broadbent are especially instructive because they show that anthropomorphism varies across observers, longitudinal change may be modest or heterogeneous, and greater humanlike presentation can alter disclosure selectively by topic. These findings require differentiated mechanisms rather than one summary claim about AI making people more social Folk et al. (2025) Guingrich & Graziano (2025) Zhu & Broadbent (2025).

Open Questions and a Reproducible Research Program

The first major unresolved question is how social presence develops with prolonged use of adaptive, multimodal systems. Existing evidence has characterized some short-term and multiweek effects, while the interaction between initial novelty, habit, memory, design changes, and human social networks requires longer and more diverse observation. Future studies can assign identical initial systems, manipulate memory continuity and disclosures about partner identity, and include periodic independent measures of reliance, outside relationships, and perceived control. Such trials should preregister meaningful primary outcomes and sample sizes suited to expected heterogeneity, rather than extrapolating from averages across very different users Bickmore & Picard (2005) Folk et al. (2025) Guingrich & Graziano (2025).

A second question concerns mechanism transportability. The response to a scripted chatbot in customer service might resemble, differ from, or interact with the response to a voice companion or embodied assistant during personal conversation. The meta-analytic literature on adoption provides evidence for links among anthropomorphism, affect, and social presence in studied service contexts, but the outcomes do not by themselves identify attachment, mental health benefit, or years-long relational continuity. Mechanism-focused replication should vary partner type and task stakes while reusing common validated measures and reporting where instruments lose measurement equivalence. Both positive and null findings can constrain an appropriately bounded scientific account Zhang & Sheng (2026) Feine et al. (2019) Nowak & Biocca (2003).

A third question concerns interface ethics and users' correct understanding of artificial agency. Designers can deliberately combine warm language, humanlike voices, and continuity cues that create a compelling impression of presence; the resulting experience may be valuable to users. Research should test whether that impression coexists with comprehension of system limitations, data retention, authority, and the availability of human alternatives. Experiments can separately measure enjoyment, honest understanding, appropriate reliance, and informed disclosure. This program supports innovation without requiring premature metaphysical conclusions about artificial experience or untested assumptions about universal welfare Lee & See (2004) Ho et al. (2018) Zhu & Broadbent (2025).

Conclusion

Anthropomorphism and social presence in human–AI encounters are coherent but distinct psychological research objects. Anthropomorphism concerns humanlike attribution; social presence concerns the experienced social salience of an interactional partner. Their relationship is shaped by linguistic and embodied cues, source labels, conversational contingency, motivational states, individual differences, and the architecture of repeated engagement. Foundational social-psychological and human–computer interaction traditions show why simple machines can elicit social conduct, while contemporary conversational systems create new opportunities for sustained adaptive exchanges Nass & Moon (2000) Epley et al. (2007) Feine et al. (2019).

The research base supports a differentiated account. Experiments document effects of framing, design cues, and evaluation expectations; longitudinal and companion studies show diverse responses and limits to generalizing average effects; cross-study synthesis identifies relationships among anthropomorphism, affect, and technology adoption within studied populations. Disagreement among findings often reflects differences in what was manipulated or measured. The most defensible inference is about human perception and behavior under specified conditions. Judgments about machine competence, continuity, agency, or subjective experience require independent evidence and their own operational definitions Araujo (2018) Folk et al. (2025) Guingrich & Graziano (2025) Zhang & Sheng (2026).

A cumulative science of human–artificial encounters should therefore integrate controlled cue manipulations, psychometrically precise measurements, audited system behavior, repeated observations, independent replication, and ethical examination of user understanding. Its explanatory unit can include the entire interactional configuration while its claims remain linked to observable phenomena. The resulting framework permits rigorous inquiry into Artificial Era social transformations, preserves the individuality of participants and situations, and identifies concrete experiments that can refine or overturn present explanations Nowak & Biocca (2003) Lee & See (2004) Waytz et al. (2010).

References

  • Araujo, T. (2018). Living up to the chatbot hype: The influence of anthropomorphic design cues and communicative agency framing on conversational agent and company perceptions. Computers in Human Behavior, 85, 183–189. https://doi.org/10.1016/j.chb.2018.03.051
  • 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
  • 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
  • Eyssel, F., & Kuchenbrandt, D. (2012). Social categorization of social robots: Anthropomorphism as a function of robot group membership. British Journal of Social Psychology, 51(4), 724–731. https://doi.org/10.1111/j.2044-8309.2011.02082.x
  • Feine, J., Gnewuch, U., Morana, S., & Maedche, A. (2019). A taxonomy of social cues for conversational agents. International Journal of Human-Computer Studies, 132, 138–161. https://doi.org/10.1016/j.ijhcs.2019.07.009
  • Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, Article 36548. https://doi.org/10.1038/s41598-025-19212-2
  • Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475
  • Guingrich, R. E., & Graziano, M. S. A. (2025). A longitudinal randomized control study of companion chatbot use: Anthropomorphism and its mediating role on social impacts. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1153–1153. https://doi.org/10.1609/aies.v8i2.36618
  • Heerink, M., Kröse, B., Evers, V., & Wielinga, B. (2010). Assessing acceptance of assistive social agent technology by older adults: The Almere model. International Journal of Social Robotics, 2(4), 361–375. https://doi.org/10.1007/s12369-010-0068-5
  • 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
  • Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229. https://doi.org/10.1080/00332747.1956.11023049
  • Kiesler, S., Powers, A., Fussell, S. R., & Torrey, C. (2008). Anthropomorphic interactions with a robot and robot-like agent. Social Cognition, 26(2), 169–181. https://doi.org/10.1521/soco.2008.26.2.169
  • 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
  • Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It's only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94–100. https://doi.org/10.1016/j.chb.2014.04.043
  • Luger, E., & Sellen, A. (2016). 'Like having a really bad PA': The gulf between user expectation and experience of conversational agents. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, 5286–5297. https://doi.org/10.1145/2858036.2858288
  • Mori, M., MacDorman, K. F., & Kageki, N. (2012). The uncanny valley. IEEE Robotics & Automation Magazine, 19(2), 98–100. https://doi.org/10.1109/MRA.2012.2192811
  • 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
  • Nowak, K. L., & Biocca, F. (2003). The effect of the agency and anthropomorphism on users' sense of telepresence, copresence, and social presence in virtual environments. Presence: Teleoperators and Virtual Environments, 12(5), 481–494. https://doi.org/10.1162/105474603322761289
  • Shank, D. B., Graves, C., Gott, A., Gamez, P., & Rodriguez, S. (2019). Feeling our way to machine minds: People's emotions when perceiving mind in artificial intelligence. Computers in Human Behavior, 98, 256–266. https://doi.org/10.1016/j.chb.2019.04.001
  • Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. https://doi.org/10.1177/1745691610369336
  • Weizenbaum, J. (1966). ELIZA—A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. https://doi.org/10.1145/365153.365168
  • Zhang, F., & Sheng, D. (2026). Anthropomorphism's impact on chatbot adoption: A meta-analytic structural equation modeling approach. Technology in Society, 84, Article 103099. https://doi.org/10.1016/j.techsoc.2025.103099
  • Zhu, T., & Broadbent, E. (2025). Effects of a computer's human likeness on disclosure in psychosocial assessments: A randomised trial. Computers in Human Behavior, 169, Article 108683. https://doi.org/10.1016/j.chb.2025.108683