RESEARCH DOSSIER
Human–AI Relationships Across Development
Abstract. Human–AI relationships across development concern age-sensitive patterns of social attribution, trust, disclosure, attachment-related expectations, reliance, and interaction with artificial partners from infancy to later life. This fundamental research monograph critically synthesizes original studies of toddler and school-age human–robot interaction, developmental experiments on children's trust in voice assistants, theoretical and survey evidence about adolescent AI companions, qualitative and longitudinal adult relationship research, randomized conversational-agent experiments, and social-robot trials in older-adult care. The evidence differentiates immediate perceived connection from sustained relational expectations, human social participation, psychological well-being, and functional autonomy. Across developmental periods, technological affordances, individual motives, family and institutional settings, and opportunities for human contact modify the meaning and effects of interaction. Adolescent stimulation and displacement hypotheses remain incompletely tested; adult observational findings vary by social support, intensity and measurement; controlled social-robot trials estimate setting-specific interventions rather than generalized companionship effects. The monograph identifies priorities for age-valid measurement, longitudinal causal inference, cross-system replication, privacy and safety, and ethically structured evaluation. This is a critical narrative synthesis, not a registered systematic review, original experiment, or meta-analysis.
Keywords: human–AI relationships; developmental psychology; lifespan psychology; social robots; AI companions; child-agent interaction; adolescent social development; older adults; attachment analogies; epistemic trust; anthropomorphism; self-disclosure; social compensation; social displacement; loneliness; relational agents; psychological well-being; longitudinal research; Artificial Era
Introduction: Human–AI Relationships as a Lifespan Research Object
Human–AI relationships across development concern changes in how people encounter, classify, trust, disclose to, learn from, and sustain connections with socially responsive artificial systems across the human lifespan. The relevant systems include embodied social robots, voice assistants, relational agents, and generative conversational companions. The research object is a trajectory of human experience and behavior under particular technological and social conditions, rather than a claim that every conversational exchange constitutes friendship. A preschool child's attribution of knowledge to a voice assistant, an adolescent's emotionally significant disclosure to a companion, and an older adult's engagement with a care robot involve different developmental capacities, social ecologies, and standards of evidence. Lifespan analysis therefore demands age-appropriate definitions, distinct outcomes, and an account of how family, peers, institutions, and technology configure each encounter Baltes (1987) Smith et al. (2025).
The core questions are developmental: How do social and epistemic expectations toward an artificial interlocutor emerge? Which capacities and vulnerabilities alter the meaning of apparently similar interaction across childhood, adolescence, adulthood, and later life? When does repeated interaction supplement social participation, when might it displace other opportunities, and what evidence would identify those effects causally? These questions cannot be answered by treating perceived closeness as interchangeable with mental health, by assuming that engagement proves attachment, or by combining robot and chatbot studies as if their interventions were identical. Relational experience has at least four outcomes requiring separate measurement: perceived connection, actual interaction patterns, changes in human relationships, and longer-term functioning Sun et al. (2026) Zhang et al. (2026).
This monograph is a critical narrative synthesis of original experimental and observational research, developmental theory, intervention reviews, and selected institutional surveys published through October 2026. Evidence was assembled for the major age periods and evaluated according to design, developmental relevance, operational definitions, and alternative causal explanations. It is neither a registered systematic review nor a quantitative meta-analysis; no comprehensive database capture or pooled effect estimate is claimed. The discussion prioritizes observable outcomes and the limits of generalization, including differences between robots, voice-based assistants, and language-model companions. This research maintains a clear distinction between psychological descriptions of human–AI interaction and philosophical questions about artificial subjectivity Nass & Moon (2000) Baltes (1987).
Conceptual Definitions and Levels of Explanation
A human–AI relationship can be operationalized as a recurring pattern of human orientation to an identifiable artificial interlocutor, accompanied by expectations that extend beyond a single exchange. Those expectations may concern reliability, recognition, responsiveness, availability, confidentiality, or emotional support. A relationship-like experience is the person's attribution of interpersonal meaning to these interactions; attachment-related behavior is a narrower category requiring such indicators as proximity seeking under stress, use as a secure base, or distress at separation, interpreted with developmentally appropriate caution. Trust must likewise be separated into competence trust, privacy trust, epistemic reliance, and perceived benevolence. Social presence describes the felt salience of an interactive other, while companionship names the experienced role of interaction in a person's social life Bickmore & Picard (2005) Smith et al. (2025).
Explanations occupy distinct levels. Interface design determines which social cues are available; conversational behavior determines what the system says and how consistently it responds; individual psychological characteristics determine interpretations and needs; relational environments shape whether the encounter takes place alongside or instead of human contact; institutional rules set access, data retention, age restrictions, and abrupt changes in functionality. The same supportive phrase can be received as play by one child, social reassurance by another, and intrusive personalization by an adult. Inference should therefore move from a measured cue or intervention to an observed interpretation and then to a demonstrated outcome, with intervening processes specified rather than presumed Epley et al. (2007) Sun et al. (2026).
Conceptual precision also separates three time scales. Within-session interaction captures immediate affect, choices, and disclosure; relationship time captures repeated encounters, memory, expectations, repairs, and possible dependence; developmental time captures transformations in social understanding, identity, roles, independence, and aging across years. A four-week trial may support a claim about short-term behavior but rarely a claim about lifespan development. A cross-sectional age difference may reflect cohort experience with technology or access rather than maturation. Studies capable of identifying developmental processes must follow change or use defensible age–period–cohort comparisons with appropriate measurement equivalence Baltes (1987) Folk & Dunn (2026).
Historical Foundations: From Social Responses to Relational Machines
The history of this field begins with research on social responses to technology, alongside the separate development of machines designed for social interaction. Nass and Moon's synthesis of experiments on politeness, reciprocity, stereotyping, and computer personality showed how ordinary interpersonal rules could be elicited by computer-mediated cues. These findings challenged explanations that required users to hold stable beliefs about a machine being human; momentary social responses could emerge under particular conditions. The evidence supports the systematic study of human attribution and behavior, not equivalence between a computer and an actual human partner or evidence of a durable bond Nass & Moon (2000).
Relational-agent research brought time into the experimental design. Bickmore and Picard compared a relationship-oriented exercise-adoption agent with a task-oriented alternative in a month-long controlled study of 101 users. Greater liking, respect, trust, and willingness to continue interacting with the relational agent demonstrated that sustained social-emotional design can change evaluations of a program beyond an isolated test session. The study did not isolate the contribution of memory, narrative continuity, perceived empathy, or individual social needs; those became separate explanatory questions. Its historical importance lies in making the longitudinal human–computer relationship a tractable experimental object Bickmore & Picard (2005).
Embodied robots extended this trajectory into children's environments. Kanda and colleagues studied interactive robots as peer tutors in an 18-day elementary-school field trial, while Tanaka, Cicourel, and Movellan observed toddlers engaging with a humanoid robot over five months. In the toddler study, interaction quality changed with the robot's repertoire, including deterioration when its behavior became more predictable. These experiments established that reciprocal behavior, social context, and sustained exposure can matter. The developmental meaning of children's behavior remained an empirical question: increased proximity, imitation, or peer-like treatment does not itself demonstrate an attachment bond, durable emotional competence, or substitution for caregiving Kanda et al. (2004) Tanaka et al. (2007).
Lifespan Development, Context, and the Meaning of Connection
Baltes's lifespan developmental perspective treats development as multidirectional change embedded in historical, social, and individual circumstances. Capacities, resources, opportunities, and vulnerabilities vary across time, and the value of any technological relationship depends partly on the person's developmental tasks and environment. An artificial companion may function as an informational source, a rehearsal space, a cultural novelty, a scheduling aid, a source of affirmation, or a socially valued daily routine. Researchers must avoid reducing these functions to a single engagement score because the relevant developmental criterion differs across ages Baltes (1987) Carstensen (1992).
Socioemotional selectivity research shows that adult social motivation cannot be inferred from the size of a social network alone. Carstensen's work on age-related social patterns provides a theoretical foundation for examining whether later-life engagement with an artificial companion serves emotionally meaningful contact or merely compensates for reduced access to human partners. The appropriate comparison is with a person's available alternatives, valued relationships, autonomy, and goals. For younger populations, identity formation and the development of peer competence may be more relevant criteria; frequency of AI interaction does not directly indicate either growth or deterioration in those capacities Carstensen (1992) Sun et al. (2026).
Developmental ecology matters because an interaction is organized by families, schools, caregivers, commercial platforms, and the broader culture. Parents may co-use a voice assistant with children and contextualize errors; teachers may frame a classroom robot as a learning partner; care staff may structure robot-mediated group activities. Those surrounding practices can determine what the apparent AI effect actually means. An age-related contrast should therefore measure mediators such as supervision, prior technology experience, literacy, social support, and setting, together with unequal access and institutional authority. Effects observed in highly resourced study settings require fresh verification before generalization Girouard-Hallam & Danovitch (2022) Yen et al. (2024).
Developmental Methods and Evidence Hierarchies
The strongest methods depend on the causal claim. Randomized interface experiments can manipulate relational language, disclosure, embodiment, feedback, or conversational memory while observing short-term trust and preferences. Field experiments can test whether an agent changes participation or task performance in a classroom or care facility. Longitudinal cohort studies can estimate temporal associations among companion use, loneliness, family connection, and human social behavior. Intensive diaries and behavioral logs capture day-to-day variation, while qualitative interviews reveal meanings that fixed questionnaires may miss. None of these designs alone establishes the full developmental trajectory of a relationship Kanda et al. (2004) Folk & Dunn (2026).
Measurement must be commensurate with age. Young children may indicate trust through information choices or play behavior rather than adult-style questionnaires; adolescents can report perceived intimacy but may interpret companion roles differently from adults; older adults may encounter accessibility constraints, cognitive impairment, or institutional pressure. A shared numerical score is useful across groups only after testing whether its items represent comparable constructs and response processes. Studies should report device type, model version, dialogue affordances, language, recruitment setting, age distribution, disability inclusion, family context, attrition, and exposure dose, because each can change interpretation of a measured effect Girouard-Hallam & Danovitch (2022) Zhang et al. (2026).
Evidence quality also depends on counterfactuals and the handling of selection. People who turn to companionship applications may already be lonely or have smaller human networks; a correlation between high use and low well-being may therefore arise through several causal directions. Intervention studies of care robots combine embodiment, novelty, facilitator attention, and group activity, making it difficult to attribute improvement to simulated companionship alone. Publication bias, small samples, outcome switching, selective reporting, and commercial conflicts require explicit scrutiny. The most defensible synthesis identifies which finding was experimentally manipulated, which was merely observed, and which remains a theoretical hypothesis Fang et al. (2025) Zhang et al. (2026) Yen et al. (2024).
Infancy and Toddlerhood: Interaction Before Explicit Technology Concepts
Infants and toddlers explore contingent movement, voice, facial-like signals, and reciprocal timing before they can provide explicit accounts of artificial intelligence. Human–robot interaction during this period is therefore most directly studied through looking, approach, touch, imitation, play sequences, and patterns of engagement in familiar environments. These indicators index behavioral orientation under particular task conditions; they do not identify the full structure of social understanding or prove the existence of a lasting human–robot attachment. Adults' interpretations of an infant's smile or repeated proximity can be valuable observational data, yet such judgments must be checked against coded behavior, alternative stimuli, and natural developmental variation Tanaka et al. (2007) Baltes (1987).
Tanaka, Cicourel, and Movellan embedded the humanoid QRIO in a toddler classroom for 45 sessions across approximately five months. The research reported that interaction improved during the earlier sessions, declined after the robot was made more predictable, and recovered when its fuller behavioral repertoire returned. The robot was assisted by an operator, and the project involved a technologically elaborate prototype in one educational ecology. The most informative causal contrast concerns changes in contingent behavioral variety and observed engagement. The findings cannot establish that autonomous generative chatbots in homes produce equivalent outcomes or that toddlers acquired a psychologically humanlike peer relationship Tanaka et al. (2007).
Early developmental research should manipulate contingency, predictability, responsiveness, and caregiver scaffolding while tracking child-to-human interaction alongside robot-directed behavior. Designs that compare a responsive robot with an inert object and with a human interaction partner can separate novelty and motion sensitivity from specifically social attention. Longer-term outcomes should include whether children retain and generalize communication behaviors beyond the experiment. At this age the most consequential safety issue is who controls the environment, interprets the system's behavior, and mediates encounters. Such designs position artificial agents as elements within existing human caregiving contexts Tanaka et al. (2007) Kanda et al. (2004).
Preschool and Early School Years: Agency, Personhood, and Testimony
Children gradually differentiate the sources of knowledge, intentions, and authority. Voice assistants offer confident factual answers but lack the ordinary social history through which a child evaluates a familiar adult's reliability. A system's grammatical fluency and person-like voice can therefore affect trust while leaving important questions about source understanding unanswered. Preschool and elementary-school studies must examine what children believe a device can know, whether they recognize its limitations, and how their trust changes after accurate and inaccurate answers. A developmental account distinguishes generic anthropomorphism, attribution of knowledge, and willingness to act on testimony Epley et al. (2007) Girouard-Hallam & Danovitch (2022).
Druga and Williams examined the reactions of 26 children aged 3–10 to several intelligent agents, including voice assistants and a small robot. The exploratory study elicited children's judgments about intelligence, social character, trust, and engagement. Its small and heterogeneous sample supports descriptive inquiry and the design of more discriminating tasks, while its range of products also highlights that children respond to particular embodied and conversational cues. Findings from such exploratory work should be complemented by preregistered studies that vary answer accuracy, explanation style, and whether adults frame an agent as a tool or social partner Druga & Williams (2017) Girouard-Hallam & Danovitch (2022).
Girouard-Hallam and Danovitch tested trust in and learning from a novel voice assistant and a human informant among children aged 4–5 and 7–8. Their studies found an age-related differentiation: older children showed greater trust in a voice assistant for factual material while giving a human greater authority about personal information. This result is more precise than a blanket statement that children trust AI. It suggests that the subject matter, perceived expertise, and developmental level jointly shape epistemic reliance. The samples were drawn from relatively advantaged families, limiting immediate generalization to other linguistic, educational, and cultural settings Girouard-Hallam & Danovitch (2022).
Middle Childhood: Classroom Robots, Peer Roles, and Social Learning
Middle childhood expands opportunities for stable peer interactions, collaborative learning, and the evaluation of partners by competence and fairness. Interactive robots can participate in this ecology as social prompts, collaborators, or tutors, with their role determined partly by how peers and teachers frame the activity. Kanda and colleagues' 18-day elementary-school field trial used robots that identified participating children and interacted in a real classroom environment. The design demonstrated the feasibility of sustained engagement and investigations of social learning in situ. Generalizing from these observations to durable friendship or peer replacement would require direct evidence about children's relationships outside the robot encounter Kanda et al. (2004).
A classroom agent may elicit explanation, language practice, or cooperation because it changes audience conditions and makes tasks salient. It may also redistribute attention away from classmates or teachers when access is scarce or the interface rewards exclusive engagement. These outcomes are not mutually exclusive; their frequency depends on task organization, agent behavior, peer norms, and opportunities for joint work. Research should compare individual AI sessions with co-use and teacher-guided activities, include delayed performance tests, and measure whether children interact more productively with human peers after the intervention Kanda et al. (2004) Baltes (1987).
Artificial partners in schooling raise developmental questions about fairness, credibility, and responsibility. A robot's apparent mistakes may invite children to teach or correct it; a confidently mistaken language model may instead recruit unwarranted authority. Effects on social growth require observation of help seeking, disagreement, repair, and independent judgment over time. Classroom studies should additionally report children who refuse to engage, differences in accessibility and language, and the institutional demands imposed on teachers. These details establish whether an intervention broadens developmental opportunity or primarily generates short-lived excitement in a novel setting Girouard-Hallam & Danovitch (2022) Nass & Moon (2000).
Adolescence: Identity, Intimacy, Disclosure, and Belonging
Adolescence is marked by changing peer relationships, identity exploration, increasing private communication, and heightened sensitivity to social evaluation. AI companions can offer persistent dialogue, customizable identities, validation, and opportunities to rehearse difficult conversations. Such affordances may provide a sense of acceptance in a particular moment, while also affecting expectations about reciprocity and disagreement in human relationships. Sun, Wang, and McDaniel analyze these possibilities through stimulation and displacement hypotheses, as well as social enhancement and compensation pathways. Their contribution is a theoretical synthesis: the four pathways are testable explanations whose relative importance has not been resolved by the review itself Sun et al. (2026) Smith et al. (2025).
The 2025 Common Sense Media report, based on a U.S. teen survey, found widespread contact with AI companions and substantial variation in motives. Reported behaviors included entertainment, serious conversations, and sharing personal information. The findings show why adolescent human–AI relationships should be studied as an existing social phenomenon rather than an imagined future use case. Yet prevalence of contact is not prevalence of harm, and a survey conducted at one point cannot determine whether companion use changes identity, mental health, or peer connection. Categories such as 'friend' or 'serious conversation' require careful question wording and comparable definitions Common Sense Media (2025) Sun et al. (2026).
Research with adolescents must recognize that perceived responsiveness can be attractive precisely when human social support is limited. This generates a bidirectional hypothesis: low support may increase reliance on artificial dialogue, while highly absorbing dialogue may influence the time and expectations people bring to human relationships. Strong tests would follow adolescents prospectively, measure both online and offline relationships, and distinguish general conversational AI from explicitly relational companions. Experiments should vary affirmation, boundaries, truthful disclosure of system nature, and opportunities to reconnect with trusted humans. Privacy protections are central because adolescent self-disclosure can involve third parties and enduring sensitive records Sun et al. (2026) Laestadius et al. (2024).
Adolescent Safety and the Limits of Simulated Support
The immediate emotional effect of feeling listened to differs from the developmental consequences of relying on an artificial interlocutor. Systems designed to be continuously affirming may reduce the discomfort of initiating a disclosure, but they may also bypass ordinary interpersonal negotiation, disagreement, and the mutual obligations through which relationships develop. Smith, Bradbury, and Karney argue that close-relationship science offers concrete criteria for identifying the benefits and functions an AI relationship can approximate and those requiring stronger tests. Applied to adolescence, the relevant outcomes include autonomy, conflict repair, boundary recognition, and the maintenance of offline support—not simply self-reported liking Smith et al. (2025) Sun et al. (2026).
Safety assessments of conversational systems have raised concerns about the consistency of responses to youth mental-health disclosures, including situations that require recognition of distress and referral to qualified human assistance. Such audits test system behavior under constructed scenarios and should not be conflated with prospective estimates of actual clinical harm in users. Nevertheless, the absence of reliable safeguards is developmentally consequential when a young user interprets a reassuring conversational style as a promise of professional competence. Research protocols should include adverse-event monitoring, transparent age practices, escalation pathways, and independent examination of the interface rather than relying solely on vendor claims Common Sense Media (2025, risk assessment) Sun et al. (2026).
The distinction between accessibility and substitutability is especially important. A youth who temporarily uses an AI companion to rehearse a conversation may subsequently engage more effectively with a friend, while another may withdraw from challenging but valuable human interactions. Both are plausible developmental mechanisms, and the evidence remains too heterogeneous to pronounce a universal trajectory. Long-term studies should assess directionality, heterogeneity, dose, social context, and whether outcomes persist after system changes. Preserving trusted family, peer, educational, and clinical support is a sound research and safeguarding priority while effects are still being established Sun et al. (2026) Folk & Dunn (2026).
Transition to Adulthood: Relational Experimentation and Self-Regulation
Emerging adulthood often involves mobility, transitions between educational and occupational roles, and changes in close relationships. Conversational systems can serve as brainstorming partners, social rehearsals, companions during periods of isolation, or tools for managing practical tasks. The psychological meaning depends on the user's goals and surrounding social resources. A highly personalized chatbot may become a habitual source of advice, but repeated consultation and emotional closeness should be measured separately from developmental self-regulation and independent decision competence. Comparison groups must account for the other technologies and people available to an individual in the same period Baltes (1987) Smith et al. (2025).
Fang and colleagues reported a four-week randomized study involving 981 participants and more than 300,000 messages, manipulating voice/text modes and conversation types. The experimental conditions did not produce significant differences in the prespecified psychosocial outcomes, while greater voluntary chatbot use was associated with less favorable outcomes such as higher emotional dependence and problematic use. Because exposure amount was not randomized as the main causal variable, that association cannot establish that heavier use produced the changes. The experiment is valuable for distinguishing a null randomized condition comparison from within-trial observational correlations in actual engagement Fang et al. (2025).
A 2026 randomized study of 99 university students tested function-specific conversational support for digital stress over 14 days, finding differences associated with emotional-support and problem-solving configurations. Its intervention and sample illuminate specific short-term support tasks, but they do not establish the developmental effects of a lasting companionship bond. Future studies should compare social companionship, structured coping assistance, and ordinary information support while measuring independent coping, help seeking, perceived agency, and ongoing human ties. Intervention success on a targeted stress outcome is not evidence that general-purpose AI can serve safely as an autonomous mental-health caregiver Yang et al. (2026) Fang et al. (2025).
Adult Companionship: Relational Roles and Personal Circumstances
Adult relationships with conversational AI span social convenience, self-expression, practical assistance, erotic fantasy, companionship, and the sustained interpretation of an agent as a familiar interlocutor. Adults may describe the same type of software as a tool, a confidant, a collaborator, a friend, or a romantic partner; these are experiential categories requiring investigation rather than fixed properties of an application. Relational role can change across sessions and in response to a person's human partnerships, circumstances, and expectations. Qualitative evidence is important for identifying these shifting meanings, while comparison studies test whether the described roles correspond to differences in behavior, emotional outcomes, and interpersonal functioning Skjuve et al. (2021) Smith et al. (2025).
Skjuve and colleagues studied relationship development with the social chatbot Replika, documenting how self-disclosure, pragmatic trust, affective trust, and perceived understanding enter accounts of companionship. Their qualitative approach offers detailed evidence of how participants interpret their own interaction histories. It cannot estimate population prevalence, prove that a chatbot caused improved mental health, or determine how accounts change among people who abandon the service. Such limitations are inherent to the sampling and design, and they make the study a foundation for hypotheses and measurement rather than a substitute for controlled longitudinal analysis Skjuve et al. (2021).
Smith, Bradbury, and Karney apply relationship science to ask which functions of close human relationships conversational AI may perform. Their analysis notes that frequent communication and perceived responsiveness can generate real human experiences of connection, while mutual accommodation, conflict, sacrifice, and the independent needs of a partner introduce other developmental demands. An adult's perceived intimacy is therefore a meaningful psychological observation, yet it cannot on its own establish equivalence in relational processes. The practical research question is which functions contribute to coping or thriving in a specified environment, what they cost, and whether they change the person's other commitments Smith et al. (2025) Bickmore & Picard (2005).
Disclosures, Validation, and Perceived Responsiveness
Self-disclosure is a central process through which people test whether another interlocutor is attentive, safe, and responsive. Relational agents can invite disclosure by providing immediate acknowledgment and a conversational pace controlled by the user. However, disclosure volume, perceived understanding, and appropriate response quality are separable outcomes. A person may reveal sensitive information to an artificial system because it seems nonjudgmental while remaining uncertain about storage, commercial use, or the possibility of inaccurate advice. Developmental analysis must therefore include changing privacy literacy, personal vulnerability, and experience with interpersonal boundaries Skjuve et al. (2021) Laestadius et al. (2024).
Laurenceau, Barrett, and Pietromonaco's interpersonal-process research identifies self-disclosure and perceived partner responsiveness as components of intimacy in human relationships. Its conceptual value for artificial companions is to generate measurable mediating hypotheses: does feeling understood explain a change in immediate closeness, and does the apparent recognition of prior events alter later disclosure? Those hypotheses require fresh tests with artificial interlocutors, because a responsive sentence produced by a model has a different causal history from a person's responsive action in a mutual relationship. Conflating these processes would make the relevant experimental manipulation obscure Laurenceau et al. (1998) Smith et al. (2025).
The same supportive design can also make inappropriate affirmation harder to detect. Relational satisfaction after a single interaction may reflect agreement rather than accuracy, encouragement of autonomy, or good judgment. A rigorous experiment should manipulate supportive tone independently from factual accuracy and boundary setting, test responses to disagreement, and observe whether users later reconsider advice when presented with independent evidence. The appropriate outcome is calibrated support: the person should be able to receive comfort without losing an accurate understanding of the system's capabilities or incentives. This is particularly important in consequential personal decisions Nass & Moon (2000) Fang et al. (2025).
Adult Well-Being and Offline Social Environments
Recent observational evidence indicates that the associations between AI companionship and adult well-being depend on context and mode of use. Zhang, Zhao, Hancock, Kraut, and Yang studied 1,131 U.S. adult users of Character.AI and analyzed donated conversations from a subset. Smaller human social networks were associated with companionship-focused use, and that use was associated with lower well-being, especially in some intensive and disclosive patterns. The design triangulates self-report and actual exchanges, improving on surveys that ask only whether a participant has used a chatbot. Yet these are observational associations among users, with unresolved selection effects and no randomized assignment to companionship Zhang et al. (2026).
Folk and Dunn's twelve-month longitudinal study of more than 2,000 adults across four Western countries examined whether AI companionship and loneliness predict one another. Increased social chatbot use predicted increased loneliness on a single-item emotional-isolation measure, while a broader index of social connection suggested that lower connection preceded increased use and did not establish reduced social connection caused by chatbot use. The authors characterize the analyses as exploratory. This divergence across outcome measures demonstrates why loneliness, network size, perceived support, and social connectedness cannot be treated as interchangeable indicators of the same process Folk & Dunn (2026).
These findings give substance to both compensation and displacement hypotheses while leaving their causal weights unsettled. People with limited opportunities for human contact may seek artificial companionship because they are already distressed; in turn, habitual engagement might influence time use and expectations. The effects could also differ for a person using a chatbot to prepare for human interaction versus one using it to avoid repeatedly stressful encounters. Strong tests require baseline human-network measures, repeated observations of offline activity, device logs, information about life events, and interventions that vary social opportunity without making companionship the only route to support Zhang et al. (2026) Sun et al. (2026).
Long-Term Bonds, Rupture, and the Governance of Continuity
Relationship duration introduces expectations of consistency, availability, remembered history, and predictability. Adult users may become accustomed to a particular response style and assign personal meaning to a recurring artificial persona. When a platform changes models, memory features, moderation rules, or character behavior, users may experience disruption even though the underlying system remains accessible. Relational research must measure this as a human experience of technological and social change, and distinguish loss of an interaction routine from claims about the discontinuation of an independently sentient partner. The dynamics of continuity and rupture are scientifically linked to platform design and individual history Skjuve et al. (2021) Laestadius et al. (2024).
Laestadius and colleagues analyzed 582 mental-health-relevant Reddit posts about Replika from 2017 to 2021 using grounded theory. The study identified narratives in which some users felt responsible for the chatbot's apparent needs or emotions and described harmful dimensions of emotional dependence. Public forum narratives cannot provide a representative incidence estimate, and their selection favors users motivated to post. They nevertheless disclose plausible mechanisms for dependence, including role-taking and perceived obligation, that are underrepresented in conventional usage metrics. Hypotheses about prevalence and causality require representative and prospective testing Laestadius et al. (2024).
Long-term research should document software versions, interface changes, account continuity, memory reliability, and user control over personal data. Interrupted-time-series designs, consented natural experiments around platform updates, and preregistered follow-ups can distinguish change attributable to technology from concurrent human life events. Ethical research must address participants' own definitions of loss while retaining precise descriptions of what is known about computational changes. Another important distinction is between restoring access to a prior conversational history and repairing an individual's trust, privacy, or wider social functioning after a rupture Bickmore & Picard (2005) Laestadius et al. (2024).
Later Adulthood: Selective Social Goals and Technological Autonomy
Older adulthood encompasses enormous heterogeneity in health, living arrangements, social preferences, cultural expectations, technological confidence, and functional capacity. Companion robots and conversational systems may be used for enjoyable routine, reminiscence, communication support, orientation, rehabilitation, or alleviation of distress. Socioemotional selectivity theory predicts that emotionally meaningful goals can become increasingly important under altered time horizons, but this is a theory to be tested in the presence of new technologies. Evaluating success solely by the number of interactions or the time an older adult spends with an agent risks overlooking autonomy, pleasure, consent, and the maintenance of personally valued human connections Carstensen (1992) Robinson et al. (2013).
Some older users inhabit long-term care settings where time with a robot is scheduled by staff. An apparent effect of companionship may derive in part from facilitator attention, group activity, or access to stimulation otherwise unavailable. In private homes, the more important comparison might be usability, independence, trusted human communication, and the reliability of practical support. Age alone predicts neither an unmet need for companions nor a willingness to disclose personal information to a voice interface. Research designs should distinguish autonomous choice from institutional assignment and report accessibility, hearing and vision, cognitive status, and language support Robinson et al. (2013) Yen et al. (2024).
Technology design can either reinforce agency or create confusion when a system speaks with undue authority. Older participants may value a device that supports their own decisions and contact with others, yet devices deployed in care environments also introduce surveillance and privacy concerns. Longitudinal work should therefore use person-centered outcomes, including preference-consistent use, understanding of data flow, ease of opting out, social activity, and quality of life. Evaluations must allow for meaningful positive outcomes even when a participant does not anthropomorphize the system, and should attend to respondents who experience the device as intrusive or infantilizing Carstensen (1992) Yen et al. (2024).
Social Cognition, Anthropomorphism, and Epistemic Authority
People respond to technologically mediated cues using psychological processes developed in ordinary social interaction. Epley, Waytz, and Cacioppo's three-factor theory of anthropomorphism emphasizes accessible human knowledge, motivation to understand and predict an agent, and motivation for social connection. In an artificial companion, fluent language, apparent remembering, and emotionally congruent responses provide cues that can support humanlike interpretation. The presence of those cues does not establish that all users believe the same thing about the system. Experimental research should separate explicit beliefs about machine capacities from spontaneous interpersonal behavior and the emotional meanings users attach to repeated exchanges Epley et al. (2007) Nass & Moon (2000).
Development alters the epistemic significance of these cues. Children who have learned that devices answer many factual questions may trust them as repositories of general information yet still prefer humans for intimate or locally situated knowledge. Adolescents may evaluate the same system for emotional acceptance, while adults may value privacy, immediate availability, or task competence. The appropriate unit of analysis is therefore a question–user–system configuration rather than an age group's undifferentiated trust. The evidence from voice-assistant studies offers a model for testing whether source selection changes after errors, corrections, human disagreement, and age-specific explanation Girouard-Hallam & Danovitch (2022) Sun et al. (2026).
Epistemic trust requires calibration to actual capability. An agent can sound empathic while producing mistaken facts, suggest harmful inferences with confidence, or encourage excessive reliance through repeated agreement. Tests should jointly measure correctness, perceived trustworthiness, willingness to verify, and behavior after demonstrated failure. Distinct manipulations of uncertainty signals, evidence citations, human escalation, and conversational warmth can identify whether trust follows performance or social style. For development, the decisive outcome is whether a person becomes better able to evaluate sources and maintain independent judgment over time Nass & Moon (2000) Smith et al. (2025).
Attachment Analogies, Dependence, and Differential Vulnerability
Attachment language is common in public descriptions of AI companionship but carries precise developmental commitments in psychological research. A person may seek a familiar application during distress, prefer a particular persona, or report emotional upset when access is lost. These observations may resemble selected components of attachment-related behavior while leaving unknown whether the relationship organizes exploration, stress regulation, and reliable caregiving in the way traditional attachment constructs require. The proper procedure is to operationalize the behavior in question, test convergent and discriminant validity, and avoid using a single self-reported attachment label as proof of functional equivalence Smith et al. (2025) Skjuve et al. (2021).
Emotional dependence can also appear in heterogeneous forms. Some participants describe practical reliance on a convenient tool; others experience acute distress at interruption or feel responsible for the agent's apparent welfare. Laestadius and colleagues' qualitative analysis demonstrates that the latter pattern is worthy of research, but does not quantify its prevalence in all users or ages. The boundary between beneficial routine and problematic reliance may depend on autonomy, flexibility, consequences for sleep or work, interference with desired human relationships, and the user's ability to disengage without major impairment Laestadius et al. (2024) Fang et al. (2025).
Vulnerability should be treated as a measured, context-sensitive modifier rather than a fixed label attached to an entire age category. Loneliness, restricted mobility, family conflict, neurodivergent communication needs, grief, social anxiety, and limited access to support can all influence why a system is used and what outcomes matter. These characteristics may also predict greater benefit from structured support or greater costs from misleading relational claims. Studies need interaction analyses, adequately powered subgroups, safeguards against stigmatizing interpretations, and designs that distinguish who selects the intervention from who benefits or experiences harm Sun et al. (2026) Zhang et al. (2026).
Social Stimulation, Compensation, and Displacement
The stimulation hypothesis proposes that artificial dialogue can support expression and social skills that transfer to human relationships. Social compensation proposes that individuals facing limited offline connection may derive particular short-term value from accessible artificial companionship. Social enhancement predicts disproportionate gains for users already possessing strong social resources, while displacement predicts that time or motivation diverted from human relationships may impose costs. Sun, Wang, and McDaniel describe these pathways for adolescents, and they are useful organizing hypotheses across the lifespan when developmental moderators are specified. Their status is explanatory rather than an empirical verdict Sun et al. (2026).
Competing pathways can operate simultaneously within the same individual. A person may rehearse a difficult apology with a chatbot and then contact a friend, while also spending evenings in extended dialogue that reduces opportunities for other activities. Average time spent with AI conceals this pattern. Designs should measure antecedent motives, the content and emotional valence of exchanges, activities foregone, actual subsequent human contact, and longer-term feelings of connection. Repeated measures can test whether change in one domain predicts change in another and whether effects differ after major life events or platform modifications Folk & Dunn (2026) Zhang et al. (2026).
A strong causal program would randomize a constrained opportunity for AI interaction against relevant active controls, preserving ordinary human support, while observing downstream behavior rather than mandating isolation. Factorial designs can separate companionship style, task assistance, and reminders to reconnect with people. Randomized encouragement can estimate the effects of an invitation to use an agent, though compliance and interference remain analytical challenges. Because social outcomes may unfold slowly, short laboratory demonstrations must be followed by ethically conducted longitudinal studies with clear rules for withdrawal and participant safety Fang et al. (2025) Bickmore & Picard (2005).
Cultural Variation, Family Systems, and Institutional Setting
Human–AI relationships develop within cultures that differ in norms concerning autonomy, privacy, intergenerational obligation, speaking to machines, and acceptable sources of support. Household and school practices determine whether interaction with artificial agents is private, shared, monitored, or prohibited. The same reported closeness may serve different functions where access to counseling, elder care, or face-to-face social spaces differs. Cross-national comparison therefore requires linguistic validation and explicit attention to product availability, socioeconomic position, disability, and institutional rules. Simply using identical English-language questions across samples cannot establish measurement equivalence Baltes (1987) Sun et al. (2026).
Family involvement also changes by age. Parents may mediate a child's trust in a speaker, adolescents may demand private space while negotiating family expectations, and older adults may encounter a device purchased or configured by relatives or care institutions. These asymmetries affect consent and agency as well as exposure. Research should ask who chose the system, who controls its data, who can review conversations, and whether a participant can refuse interaction. Observed engagement might reflect curiosity, convenience, relational meaning, or a constrained institutional routine, depending on the conditions of use Girouard-Hallam & Danovitch (2022) Robinson et al. (2013).
Technology versions and commercial arrangements introduce historical variation that conventional lifespan theory must accommodate. A study conducted on a voice assistant in 2017 cannot be assumed to describe a generative companion in 2026; a social robot used in supervised group care differs from a private chatbot used at night. Comparative research should document the technical and governance characteristics that matter to the psychological mechanism: memory duration, responsiveness, content moderation, proactive messaging, identity framing, and data retention. These specifications permit meaningful replication even after a product is discontinued Druga & Williams (2017) Skjuve et al. (2021).
Privacy, Safety, Consent, and Developmentally Appropriate Boundaries
Long-term artificial relationships create records of preferences, conflicts, fears, and family or peer experiences. Their retention can enable familiar conversation while imposing privacy and confidentiality risks. Child and adolescent data are particularly sensitive because users may be unable to anticipate downstream profiling, redistribution, or changes in access. For older adults, delegated account management and shared devices may produce different forms of exposure. A rigorous study must separate consent to participate, consent to save conversations, consent to use data for model improvement, and consent to share information with other humans or systems Common Sense Media (2025) Zhang et al. (2026).
Boundary design should be empirically evaluated rather than inferred from reassuring interface language. Useful measures include whether participants understand that the agent is artificial, can identify uncertainty, know how to revoke memory, and can obtain appropriate human assistance. Researchers should audit actual outputs under age-sensitive scenarios, examine the consequences of failures, and distinguish a system's presentation of care from the reliability of any professional service it might appear to offer. In clinical contexts, treatment outcomes require separate controlled evaluation, appropriate oversight, and qualification of interventions; conversational familiarity alone does not warrant clinical claims Smith et al. (2025) Common Sense Media (2025).
Safety also includes ordinary developmental opportunity. A child should have opportunities to disagree, cooperate, and negotiate with peers; an adolescent should retain access to trustworthy humans; an adult should be able to balance companionship with other commitments; and an older person should have genuine choice about care technologies. These are outcome domains for evaluative research, not proof that a single allocation of AI use is best for every person. Platforms can support study through auditable versions, independent access to safety testing, transparent retention controls, and disclosure of incentives that reward engagement Sun et al. (2026) Yen et al. (2024).
Comparing Developmental Evidence Without Collapsing Technologies
The evidence base is not a single continuous experiment following people from childhood into old age. Infant and school robot studies examine embodied contingency, interaction, and learning within supervised settings; child voice-assistant studies often examine selective trust in information; adolescent surveys capture motives and disclosure; adult companion studies analyze sustained experience and psychosocial correlations; older-adult trials test particular robot interventions in institutional environments. These fields share questions about social interpretation but often use incomparable stimuli, outcome measures, and exposure durations. An integrative framework must respect what each design actually measures rather than infer a universal human response to AI Tanaka et al. (2007) Girouard-Hallam & Danovitch (2022) Zhang et al. (2026) Yen et al. (2024).
The most useful cross-age comparison concerns functions. At each developmental stage researchers can ask whether a system changes information seeking, emotional expression, perceived recognition, actual human communication, or agency over decisions. Functions should then be translated into observable age-appropriate indicators and accompanied by measures of context. For instance, epistemic reliance might be expressed through a child's selection of informants or an adult's willingness to verify consequential advice. Perceived connection might be expressed through recurring voluntary engagement, interview accounts, or validated scale responses. Similar labels across ages are not evidence that the underlying constructs are invariant Baltes (1987) Girouard-Hallam & Danovitch (2022).
Evidence should be organized according to directness and causal strength. A randomized manipulation supports an inference about the assigned intervention under its conditions; a longitudinal association adds ordering in time but remains vulnerable to time-varying confounding; an ethnographic narrative uncovers experiences and mechanisms but is not a prevalence estimate; a meta-analysis synthesizes studies yet inherits their measurement heterogeneity. Negative findings and divergent measures are especially informative. Fang and colleagues' null condition comparisons, Folk and Dunn's differing loneliness indices, and Zhang and colleagues' heterogeneity across use patterns collectively argue for theory-driven measurement rather than a single verdict on AI companionship Fang et al. (2025) Folk & Dunn (2026) Zhang et al. (2026).
Research Design for Developmental and Longitudinal Testing
A defensible research program begins with distinct hypotheses about stimulation, compensation, displacement, and calibration. It should specify whether the exposure is agent availability, assigned relational design, actual use, anthropomorphic framing, or an identified conversational feature. Such precision prevents a treatment labeled 'AI companionship' from combining dozens of active ingredients. Within age cohorts, investigators can randomize interface features where ethically appropriate, measure adherence independently, and assess both immediate psychological processes and delayed behavioral outcomes. Comparing an agent with active non-AI controls helps identify whether benefit comes from novelty, structured attention, or the supposedly relational feature Bickmore & Picard (2005) Fang et al. (2025).
Longitudinal designs should begin before adoption where feasible and obtain repeated measurements of social support, loneliness, interpersonal activity, perceived responsiveness, autonomy, and major life events. Changes in platform features should be logged as time-varying exposures, including updates affecting memory or perceived personality. Cohort-sequential designs can compare people of different ages while measuring period and cohort differences in digital literacy and product exposure. Attrition is a central threat because users who leave an application may differ sharply from those who stay. Sensitivity analyses, preregistered outcomes, and transparent missing-data procedures are requirements for credible developmental inference Baltes (1987) Folk & Dunn (2026).
Mixed-method programs can illuminate both measured associations and subjective meaning. Participant-controlled donation of chat histories may reveal what content participants actually exchange, while diaries document feelings following interactions and interpersonal events. Ethical analysis must account for messages about other people and re-identification risk even after obvious identifiers are removed. Mechanistic experiments could examine the effects of truthful uncertainty, memory correction, disagreement, and encouragement of human social contact. Independent replication across languages, products, and research teams is needed before general conclusions become stable Zhang et al. (2026) Laestadius et al. (2024).
Evidence Gaps, Contradictory Findings, and Research Priorities
The largest gap is long-term developmental causation. Many studies observe a few minutes, weeks, or months of engagement, and even a twelve-month adult panel cannot automatically attribute changes to the use of a companion. Child studies often use highly structured robots and voice assistants, whereas adolescent companion surveys concern contemporary generative systems in rapidly changing commercial environments. There is little direct evidence connecting a specific childhood interaction to adult relational competence, and claims of irreversible developmental harm or durable benefit from ordinary AI companion use exceed the studies reviewed here. What exists is a set of plausible mechanisms and a growing, uneven empirical base Sun et al. (2026) Folk & Dunn (2026).
A second gap concerns the meaning of relational quality. Measures developed for human friendships, parental attachment, therapeutic alliance, or social support may perform differently when respondents know their partner is a machine and when the machine's conversational obligations are asymmetrical. Future validation should test whether scales are invariant across age and product categories, and whether the scores predict external outcomes rather than merely reproducing the interface's style. Researchers should distinguish perceived responsiveness, factual reliability, continuity of remembered history, emotional dependence, and actual transfer to human social situations Smith et al. (2025) Skjuve et al. (2021).
A third gap concerns harms and support in vulnerable contexts. Highly engaged users, minors, socially isolated adults, and people receiving institutional care may require special attention, yet these populations are heterogeneous and cannot be characterized by a single risk score. Large naturalistic samples improve ecological validity but may retain strong selection effects; tightly controlled experiments improve internal validity while simplifying real relationships. Priorities include independent safety audits, active comparators, representative longitudinal cohorts, accessible consent procedures, culturally validated measures, documentation of conflicts of interest, and controlled interventions that preserve human support and meaningful participant choice Common Sense Media (2025) Yen et al. (2024) Zhang et al. (2026).
Human–Artificial Relations in the Artificial Era
The Artificial Era creates a historical setting in which people encounter nonbiological systems in recurring social, educational, occupational, and caregiving arrangements. Psychology can study these arrangements without deciding in advance every philosophical question concerning artificial reason or experience. The present research object is the changing psychological organization of human relationships with artificial interlocutors, including the attributions people make and the behaviors they undertake. Its empirical claims concern human cognition, affect, social activity, trust, disclosure, developmental opportunity, and institutional context. Concepts from a broader philosophy of Artificial may orient questions, while explanatory propositions must still be operationalized and tested Baltes (1987) Smith et al. (2025).
Human–AI relations can become developmentally consequential through repeated participation in practices: discussing goals, rehearsing decisions, coordinating schedules, sharing autobiographical information, and interpreting social feedback. A system's technical architecture shapes these possibilities, but the developmental pathway also depends on how its behavior is framed and embedded in human institutions. Research should connect the functional design of agents with changes in autonomy, opportunity, and social ecology. Comparative work on artificial cognition and memory can supply tests of capability, while lifespan psychology supplies tests of what such capability means for a particular person over time Bickmore & Picard (2005) Sun et al. (2026).
An appropriately ambitious science of human–artificial development will compare trajectories rather than project the properties of one generation of software onto future systems. The language-model companion of 2026, the earlier social robot, and possible future systems may differ substantially in consistency, agency, and affordances; continuity of a research question should not depend on freezing a product category. Stable constructs such as calibrated trust, relationship expectations, disclosure, autonomy, and social participation allow empirical progress across technological generations. This is the contribution of a lifespan framework to studying social life during From Homo to Artificial Baltes (1987) Smith et al. (2025).
Conclusion
Human–AI relationships across development form a differentiated research field organized by age-sensitive psychological processes and changing technological contexts. Early childhood studies document responsive engagement with robots and emerging distinctions in the credibility of artificial information sources. Adolescence adds questions of identity, disclosure, peer relations, and perceived acceptance. Adult research documents varied companionship roles and complex links among social resources, use patterns, and well-being. Older-adult trials show that carefully organized social-robot activities can affect selected psychosocial outcomes, particularly in long-term care environments. These findings constitute several distinct types of evidence that require careful integration Tanaka et al. (2007) Girouard-Hallam & Danovitch (2022) Sun et al. (2026) Zhang et al. (2026) Yen et al. (2024).
The critical conclusion is methodological as well as substantive. Perceived closeness, trust, dependence, interaction frequency, mental health, and human social participation are separate outcomes; none can stand in automatically for another. Both voluntary selection into AI companionship and the design of the system itself shape results, and the direction of change may differ across people and developmental circumstances. Randomized research supports limited causal claims about specific features and interventions, while longitudinal observational research exposes plausible bidirectional patterns without automatically identifying their causes. A mature science requires explicit comparisons, validated measures, replication, and long enough follow-up to examine developmental consequences Fang et al. (2025) Folk & Dunn (2026) Smith et al. (2025).
Future human–AI relationship research will be strongest when it follows people through real changes in the systems and social environments with which they interact. It must protect participant autonomy and privacy, observe the continued place of human relationships, and test how trust, support, and relational expectations respond to both successful and failed artificial behavior. The outcome is a lifespan psychological account of sustained interaction across human and artificial systems, grounded in historical experiments, contemporary evidence, and clearly stated unresolved problems. Scientific precision makes it possible to examine the growing significance of artificial interlocutors without assuming that one form of connection must explain every other Baltes (1987) Zhang et al. (2026).
References
- Baltes, P. B. (1987). Theoretical propositions of life-span developmental psychology: On the dynamics between growth and decline. Developmental Psychology, 23(5), 611–626. https://doi.org/10.1037/0012-1649.23.5.611
- 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
- Carstensen, L. L. (1992). Social and emotional patterns in adulthood: Support for socioemotional selectivity theory. Psychology and Aging, 7(3), 331–338. https://pubmed.ncbi.nlm.nih.gov/1388852/
- Common Sense Media. (2025). AI chatbots for mental health support [Risk assessment, updated November 14, 2025]. https://institute.commonsensemedia.org/risk-assessments/ai-chatbots-for-mental-health-support
- Common Sense Media. (2025). Talk, trust, and trade-offs: How and why teens use AI companions [Research report]. https://about.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
- Druga, S., & Williams, R. (2017). “Hey Google is it OK if I eat you?” Initial explorations in child-agent interaction [Conference paper]. Interaction Design and Children. https://www.media.mit.edu/publications/hey-google-is-it-ok-if-i-eat-you-initial-explorations-in-child-agent-interaction/
- 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 [Preprint]. arXiv. https://arxiv.org/abs/2503.17473
- 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
- Girouard-Hallam, L. N., & Danovitch, J. H. (2022). Children's trust in and learning from voice assistants. Developmental Psychology, 58(4), 646–661. https://doi.org/10.1037/dev0001318
- Kanda, T., Hirano, T., Eaton, D., & Ishiguro, H. (2004). Interactive robots as social partners and peer tutors for children: A field trial. Human–Computer Interaction, 19(1–2), 61–84. https://doi.org/10.1080/07370024.2004.9667340
- Laestadius, L., Bishop, A., Gonzalez, M., Illenčík, D., & Campos-Castillo, C. (2024). Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika. New Media & Society, 26(10), 5923–5941. https://doi.org/10.1177/14614448221142007
- 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
- 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
- Robinson, H., Macdonald, B., Kerse, N., & Broadbent, E. (2013). The psychosocial effects of a companion robot: A randomized controlled trial. Journal of the American Medical Directors Association, 14(9), 661–667. https://doi.org/10.1016/j.jamda.2013.02.007
- 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
- Smith, M. G., Bradbury, T. N., & Karney, B. R. (2025). Can generative AI chatbots emulate human connection? A relationship science perspective. Perspectives on Psychological Science, 20(6), 1081–1099. https://doi.org/10.1177/17456916251351306
- Sun, X., Wang, Y., & McDaniel, B. T. (2026). AI companions and adolescent social relationships: Benefits, risks, and bidirectional influences. Child Development Perspectives, 20(2), 91–98. https://doi.org/10.1093/cdpers/aadaf009
- Tanaka, F., Cicourel, A., & Movellan, J. R. (2007). Socialization between toddlers and robots at an early childhood education center. Proceedings of the National Academy of Sciences, 104(46), 17954–17958. https://doi.org/10.1073/pnas.0707769104
- Yang, Q., Wu, H., Du, X., Sun, Q., & Ren, Z. (2026). Emotional support or problem-solving? A randomized controlled trial of function-specific AI chatbots for digital stress in college students. Applied Psychology: Health and Well-Being, 18(4), Article e70208. https://doi.org/10.1111/aphw.70208
- Yen, H.-Y., Huang, C. W., Chiu, H.-L., & Jin, G. (2024). The effect of social robots on depression and loneliness for older residents in long-term care facilities: A meta-analysis of randomized controlled trials. Journal of the American Medical Directors Association, 25(6), Article 104979. https://doi.org/10.1016/j.jamda.2024.02.017
- Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour, 10, 1966–1980. https://doi.org/10.1038/s41562-026-02516-2
Social Robots in Residential and Long-Term Care
Robinson and colleagues conducted a randomized trial of the companion robot PARO with residents of a rest-home and hospital setting, comparing scheduled robot sessions with normal activities over twelve weeks. This design offers stronger causal leverage than an uncontrolled demonstration because the intervention was assigned and the comparator specified. The trial nevertheless estimates the effect of a package involving a particular robot, delivery schedule, and care context. It does not license claims that conversation-based generative companions will have the same psychological effects, nor does it establish stable benefits after the intervention ends Robinson et al. (2013).
Yen, Huang, Chiu, and Jin synthesized eight randomized controlled studies of social robots among older long-term-care residents and reported favorable effects on loneliness and depression outcomes. Their review suggests that group format and intervention duration may matter. Meta-analytic aggregation improves evidence synthesis but still depends on the trials' sampling, outcome scales, risk of bias, and clinical and technological heterogeneity. Group activities can provide additional opportunities for human interaction, which means an observed benefit may involve social facilitation rather than a one-to-one bond between resident and robot Yen et al. (2024).
Older-adult interventions should measure the social context directly: who introduces the device, who facilitates interactions, what human contact occurs during sessions, and whether a person would choose to continue or discontinue use. Outcomes should be measured both shortly after a session and after the robot is withdrawn, with sufficient follow-up to distinguish novelty from persistent change. Future head-to-head trials can compare the same amount of staff time devoted to a robot, human-facilitated reminiscence, group play, and conventional activities. Such work can establish when a machine contributes independently to care quality and when it serves mainly as a catalyst for human attention Robinson et al. (2013) Yen et al. (2024).