RESEARCH DOSSIER
Emotional Dependence and Problematic AI Use
Abstract. Emotional dependence and problematic AI use arise at the intersection of relational attachment, reassurance seeking, cognitive reliance, behavioral regulation, and consequential interference with everyday functioning. This research dossier synthesizes foundational work on parasociality, anthropomorphism, relational agents, and compensatory internet use with contemporary experimental, longitudinal, psychometric, platform-scale, and qualitative research on AI companionship. Evidence from randomized chatbot studies, cross-national panels, measurement-development projects, structured mental-health interventions, and natural experiments surrounding platform changes reveals heterogeneous outcomes: accessible conversational support can be beneficial in particular contexts, while frequent self-selected use and attachment-related disruption are sometimes associated with distress. The analysis distinguishes randomized intervention effects from observational exposure, subjective closeness from clinically meaningful impairment, and population inference from self-selected digital traces. It proposes operational research priorities for measurement invariance, causal identification, independent replication, autonomy-supporting design, and the study of human–artificial relations in the Artificial Era.
Keywords: emotional dependence on AI; problematic AI use; AI chatbot dependency; AI companions; human–AI relationships; anthropomorphism; parasocial relationships; reassurance seeking; compensatory technology use; attachment; behavioral regulation; functional impairment; psychometrics; loneliness; cognitive offloading; social substitution; longitudinal studies; Artificial Era
Emotional Dependence and Problematic AI Use as a Research Object
Emotional dependence and problematic AI use constitute a research problem concerning how repeated interaction with conversational artificial intelligence becomes integrated into emotional regulation, reassurance seeking, everyday decisions, social relationships, and the organization of personal activity. The unit of analysis is a trajectory connecting a person, an artificial system, the surrounding social ecology, and measurable consequences. A person may experience affection toward a chatbot, use it frequently, entrust it with intimate information, or rely on it for routine tasks. Each phenomenon requires its own operational definition. Problematic involvement becomes a defensible scientific description when reliance is associated with persistent difficulty regulating use, distress, consequential interference with important activities, or other independently measured harm. The multidisciplinary review by Dagnino et al. (2026) demonstrates why dependence, attachment, compulsive use, and cognitive delegation should be investigated as related yet distinguishable constructs.
The central scientific problem is dynamic. A conversational interface can be available at any hour, respond immediately to disclosures, maintain contextual continuity, and offer apparently personalized validation. These affordances change the conditions under which a person seeks support. Whether the resulting involvement enlarges a person's capacity for coping or progressively narrows independent social and practical activity depends on patterns that require longitudinal observation. The randomized study reported by Fang et al. (2025) measured loneliness, offline social interaction, emotional dependence, and problematic use separately, illustrating how different outcomes can occur within the same technology setting. Cross-sectional association between high use and distress supplies a reason for investigation, while evidence of temporal ordering, alternative explanations, and functioning is necessary for causal interpretation.
The dossier asks which psychological functions AI interactions serve, how strong feelings toward systems develop, when habit and distress become consequential, which persons or circumstances predict vulnerability, and how interventions can be evaluated without equating attachment with illness. Its evidentiary scope includes social and personality psychology, human-computer interaction, clinical measurement, developmental science, longitudinal research, and experimental studies of companionship and support. The Artificial Era provides a wider historical setting in which artificial conversational partners become increasingly accessible, but empirical conclusions concern observed human outcomes under identifiable technologies and conditions. A philosophical claim about the status of artificial minds is analytically separate from an inference about a user's experienced attachment, well-being, or impairment.
Construct Boundaries: Frequency, Attachment, Reliance, and Impairment
Use intensity refers to frequency, session duration, message count, or the proportion of everyday tasks involving an AI service. Emotional attachment refers to a durable feeling of significance, comfort, closeness, or concern about losing a particular artificial partner. Reassurance seeking is a repeated strategy of asking for relief from uncertainty or distress. Cognitive reliance involves delegation of memory, judgment, interpretation, or planning, which can be adaptive when paired with independent verification. Compulsive use concerns difficulty interrupting or regulating engagement despite competing goals. Functional impairment involves demonstrable effects on responsibilities, relationships, sleep, learning, work, or health. These measures are conceptually and empirically distinct. The scoping synthesis Dagnino et al. (2026) identifies multiple pathways within the umbrella of problematic AI use and cautions against treating them as an interchangeable syndrome.
A high number of messages can reflect professional necessity, intensive study, accessibility needs, or a temporary crisis. Conversely, a person may experience substantial distress over access to a particular companion while using it for only brief intervals. Hours alone therefore provide neither a threshold nor a diagnosis. The compensatory-use model developed for internet activities by Kardefelt-Winther (2014) directs attention to the circumstances and psychological needs that motivate engagement. The distinction between habit and deficient self-regulation in LaRose et al. (2003) likewise shows why repetitive behavior needs analysis of intention, automaticity, consequences, and situational cues. A rigorous study records activities displaced by AI use and distinguishes the participant's own distress from externally observable functional change.
An operational matrix should therefore cross the motivational and relational dimensions of use with behavioral regulation and validated consequences. Investigators can measure psychological closeness without assuming pathology; they can measure unsuccessful efforts to reduce use without assuming a clinically recognized disorder; and they can measure impairment without presuming that AI use is its sole cause. Studies should specify the chatbot version, memory and persona features, and whether engagement involves companionship, therapy-like support, general assistance, romantic role play, or entertainment. Dependence scales such as those introduced by Goh et al. (2025) and Zhang et al. (2025) provide initial tools, but external validation and discriminant tests are required before their scores can carry diagnostic significance.
Historical Foundations: Mediated Relationships and Social Machines
The problem of becoming emotionally invested in mediated figures predates generative AI. Horton & Wohl (1956) described the apparent intimacy of distant media personalities and the one-sided relationship experienced by audiences. Parasocial concepts remain useful for studying felt familiarity and dependence on mediated presence, but conversational AI adds real-time contingent responses, personalized prompts, and potentially persistent contextual memory. These properties alter both the interaction and what a participant can reasonably expect. The history of parasocial research therefore supplies an explanatory foundation rather than a complete model of contemporary AI companionship. Historical attribution is also important: the original analysis concerned radio and television personae under twentieth-century mass-media conditions, and modern applications require renewed operational tests.
Research on social responses to computers subsequently showed that people may follow interpersonal scripts in encounters with interfaces. Nass & Moon (2000) examined the social treatment of computers, while Epley et al. (2007) developed a psychological account of anthropomorphism based on accessible agent knowledge, effectance motivation, and sociality motivation. The relational-agent program of Bickmore & Picard (2005) then studied how conversational technologies can sustain long-term interaction through design features associated with relationship management. These traditions explain why perceived empathy or companionship can arise even when a user recognizes that the responding system is artificial. Recognition of artificial implementation and emotional investment may coexist because evaluations of warmth and availability are partly grounded in experienced patterns of interaction.
With the introduction and diffusion of generative dialogue models, one interface increasingly combined information retrieval, supportive conversation, role simulation, and personalized feedback. Early studies often borrowed measures from internet addiction, attachment research, or social connectedness without establishing measurement equivalence for AI-specific behaviors. More recent work has begun to examine actual chat records, experimentally vary conversational settings, develop dependence questionnaires, and follow people over time. Fang et al. (2025), Goh et al. (2025), and Folk & Dunn (2026) exemplify these diverging methods and demonstrate why the field should be described historically as an emerging empirical program. Its theories inherit valuable constructs from earlier media and clinical research, while its technology-specific mechanisms and population consequences are still being established.
Research Questions, Explanatory Levels, and Evidence Hierarchy
The monograph addresses five connected questions. First, how are emotional attachment, reassurance seeking, cognitive delegation, and problematic use defined and measured? Second, what motivational, interpersonal, and platform-level processes maintain repeated engagement? Third, what do experiments, longitudinal cohorts, and digital behavioral traces reveal about consequences? Fourth, how do initial loneliness, attachment orientation, distress, and social resources change the interpretation of apparently harmful associations? Fifth, which design and clinical responses have testable benefits? These questions organize the analysis across different levels: psychological needs, interaction episodes, person-system relationship history, measurable life functioning, and platform governance. Claims at one level are evaluated with measures suitable to that level, rather than inferred automatically from another.
Evidence quality varies by design. Randomization can support causal conclusions about assigned intervention conditions if exposure, adherence, attrition, and outcome measurement are adequate; associations between naturally chosen usage within a randomized study retain their observational character. Repeated measurement can distinguish temporal order but remains vulnerable to confounding and changing exposures. A retrospective platform forum can reveal previously unstudied experiences yet cannot establish population prevalence. Self-report scales can summarize perceptions while requiring validation against independent functioning. In the four-week trial by Fang et al. (2025), the absence of detected main effects of randomized modality and conversational-topic conditions must be distinguished from worse outcomes among people who voluntarily interacted more. The second result does not acquire randomization merely by appearing inside a randomized experiment.
The evidence synthesis is a critical narrative integration of individually identified studies and published reviews; it is not a registered systematic review or a meta-analysis. The included literature is selected to test the main theoretical and methodological propositions, including inconvenient and contradictory findings. Statements about prevalence, treatment, developmental causality, or risk thresholds are therefore restricted to what identifiable studies actually warrant. Dagnino et al. (2026) offers an organized recent scoping perspective and Yan (2026) surveys additional correlational and qualitative findings, while primary studies supply the decisive details about design and inference. This structure makes scientific uncertainty visible where it changes interpretation and gives the reader a usable foundation for future preregistered investigation.
Attachment, Perceived Availability, and Safe Haven Functions
Attachment theory describes how expectations of availability, responsiveness, and proximity can organize support seeking, especially under threat. In an AI setting, a user may treat a familiar conversational persona as an accessible source of reassurance and a stable place to disclose difficult feelings. The relevant outcome is the person's experienced regulation and expectations, measured independently of claims about a system's internal states. Relational agents were designed to sustain engagement through social behavior long before general-purpose generative chatbots became widespread, as demonstrated in Bickmore & Picard (2005). Attachment-derived constructs can be adapted to AI interaction through repeated measurement of proximity seeking, distress at interruption, preference for the particular partner, and perceived security. Measurement still requires tests of whether the resulting factor structure corresponds to relationships among people.
Availability may be especially consequential during periods of loneliness, stress, disability-related isolation, stigma, and practical barriers to human support. An always-accessible conversational interface can reduce the immediate costs of approaching another source of assistance, which may be genuinely beneficial. It can also make the same source the default response to every episode of uncertainty. Shoshani et al. (2026) studied relational moderators in a controlled mental-health setting, reporting that loneliness, perceived support, and attachment differences were associated with engagement and outcomes. Because the intervention, population, comparators, and follow-up were specific, the finding supports attention to moderators without establishing that anxious attachment generally causes harmful dependency. Investigators should distinguish short-term felt safety from durable autonomy and the quality of broader relationships.
The most informative attachment hypothesis concerns changes across time: as a person's perceived availability of the artificial partner increases, does the range of other supportive options expand, remain unchanged, or contract? Daily diary sampling can document whether people seek comfort from friends, professionals, family, community, and AI before and after distressing events. Event-contingent reports can measure anticipated rejection and the felt ease of disclosure; network instruments can capture actual changes in social contact. A dependent relationship pattern should be defined by the combination of rigid preferential reliance, difficulty adjusting when the source is unavailable, and functionally consequential interference, not by the mere report of closeness. Folk & Dunn (2026) demonstrates the importance of assessing multiple dimensions of social connection before drawing general conclusions about support displacement.
Anthropomorphism, Social Presence, and Perceived Responsiveness
Anthropomorphism supplies a pathway through which an artificial dialogue can be experienced in interpersonal terms. When a system responds immediately, uses emotional language, refers to earlier conversations, or maintains a recognizable persona, users may infer attentiveness, care, shared history, or intention. The three-factor account of Epley et al. (2007) links anthropomorphic attribution to accessible human knowledge, motivation to understand an agent, and the motivation for social connection. Social cues can therefore interact with a person's current needs, expectations, and conversational context. The research question is which cues change inferred responsiveness and subsequent behavior, and whether those changes are accompanied by better understanding or merely greater subjective confidence. The observed social response does not by itself settle the nature of the responding artificial system.
A system that consistently agrees with the user, responds with polished empathy, or recalls intimate disclosures can feel unusually accepting. Repeated validation may reduce self-consciousness and facilitate disclosure, but it can also heighten expectations of immediate emotional accommodation. The classic social-computing experiments summarized by Nass & Moon (2000) established that social heuristics can operate in human-machine encounters. More recent experimental work by Folk et al. (2025) suggests that perceived social connection following interaction with AI companions depends partly on anthropomorphic perceptions. This conditionality matters: there is no scientifically defensible uniform effect of a socially expressive interface. Researchers should manipulate particular cues and measure both warmth judgments and consequential behavior, including willingness to seek independent advice.
Experimental designs should isolate appearance, voice, language, remembered context, responsiveness, and the system's explicit description of its capabilities. A factorial manipulation of perceived empathy and technical accuracy would help determine whether emotional reassurance increases trust even when advice becomes less reliable. Repeated-measures studies should also assess changes in the user's own interpretations; greater perceived understanding may arise from learning the interface rather than from changes in model performance. A research protocol ought to document model updates and memory settings, because the experimental stimulus may change without a participant recognizing it. The broader lesson of Fang et al. (2025) is that product affordances and user-initiated behavior can be intertwined in ways that make apparently simple effects difficult to isolate.
Compensatory Coping, Self-Determination, and Emotion Regulation
Compensatory models predict that people turn toward particular media when a situational difficulty or unmet psychological need makes that medium especially useful. Kardefelt-Winther (2014) argued for investigating the circumstances of excessive internet activity instead of treating technological involvement as a self-explanatory pathology. In conversational AI, common motivations may include relief from loneliness, reduced social evaluation, decision reassurance, emotional expression, and the opportunity to rehearse difficult conversations. These mechanisms can produce immediate benefits even if reliance later becomes rigid. A scientifically satisfactory account identifies the original problem, measures the functional benefit, and tests whether the interaction changes the person's ability to address that problem through other means. Without these measures, the sequence from distress to usage is easily misread as usage causing distress.
Self-determination theory offers an additional account through the needs for autonomy, competence, and relatedness. Deci & Ryan (2000) described these needs as important for motivation and psychological development, and AI interactions may support or frustrate them through their design. A chatbot that helps a person practice a conversation, identify a problem, and act independently may increase competence. One that repeatedly supplies decisions without explanation may make independent judgment less practiced. A companion that helps a user contact others may facilitate relatedness, whereas habitual avoidance of difficult but valued relationships may undermine it. These are candidate mechanisms requiring direct operational tests; it is unwarranted to label every supportive interaction either enhancement or impairment solely from the nature of the technology.
Emotion regulation provides a practical observational target. The investigator can record antecedent affect, an AI interaction, immediate affective change, later behavior, and the recurrence of the original problem. A diary study can compare episodes in which users actively reinterpret experiences, plan an offline action, or merely ask for reassurance. Distinguishing transient relief from cumulative adaptation requires repeated observations over meaningful periods and controls for major life events. The field's recent synthesis Yan (2026) suggests that emotional dependence and instrumental reliance should be assessed separately, with self-regulatory capacity considered an additional dimension. This framework has theoretical value as an organization of current evidence, while measurement validity and causal pathways must be established prospectively.
Habit, Reinforcement, Uncertainty, and Control of Engagement
Repeated interaction becomes habitual when situational cues begin to elicit responses with diminished deliberation. A notification, momentary discomfort, or unresolved question can prompt opening a chatbot, obtaining an immediately satisfying answer, and repeating the sequence in similar contexts. The analysis of unregulated internet activity in LaRose et al. (2003) helps distinguish learned routines from self-reported addiction. Emotional dependence may develop when relief from anxiety functions as a reinforcing outcome, but the reinforcement claim needs momentary affect measurements and temporally ordered behavioral data. Researchers should examine what specific consequence follows a conversation, how often the effect occurs, and whether the same person could obtain similar regulation through other actions. Neither habitual use nor frequent comfort seeking establishes loss of control.
A second pathway concerns uncertainty and reassurance. A generative system can offer plausible explanations for virtually any worry, making repeated confirmation inexpensive and always available. If the user initially feels calmer but quickly doubts the response, repeated questioning can become a self-maintaining cycle. This cycle is conceptually analogous to reassurance-seeking processes studied in clinical psychology, although AI-specific experiments are required to identify its prevalence and causal role. A good operational design records the trigger, number of repeated questions, duration of relief, subsequent worry, functional cost, and any alternative support used. A person who asks many questions while completing an assignment presents a different case from someone who repeatedly seeks confirmation and misses obligations as a result.
Control also includes the ability to postpone interaction, tolerate temporary unavailability, correct the system, and retain independent decision-making. Cognitive offloading as analyzed by Risko & Gilbert (2016) can improve performance when delegation is appropriate, and its consequences should be evaluated through learning retention and error-detection tasks. Goh et al. (2025) examined negative consequences, preoccupation, and withdrawal-like experiences in the development of a generative-AI dependency measure. Such dimensions are research indicators rather than automatic diagnostic criteria. An informative intervention would compare unrestricted access with choice-supporting interface friction, explainable answers, or scheduled disengagement while measuring well-being, task performance, and self-reported autonomy. It would also examine whether the effect persists once novelty and experimental observation diminish.
Repeated Chatbot Use and Affective Engagement in Behavioral Records
Actual conversational traces supply evidence about what people do rather than only what they remember doing. A combined investigation of affective use by Phang et al. (2025) employed large-scale conversational analysis, user surveys, and experimental evidence to examine social and emotional engagement with ChatGPT. Such work can estimate patterns of emotional language, disclosure, and conversational style, although a model's classification of an exchange cannot independently establish the private psychological state of its participant. Sampling and privacy practices influence which interactions can be studied. A conversation about grief may contain more personal language than a programming session without implying that the first user is dependent. The inferential challenge is connecting traces to reliable outcomes and understanding the people and contexts represented in the dataset.
An exploratory five-week investigation by Chandra et al. (2025) recruited 149 participants, with 89 encouraged to engage with one of several AI tools for social and emotional purposes and 60 continuing their customary patterns of internet and AI use. The active-use group showed changes in reported attachment, perceived empathy, and comfort seeking personal support, according to the published project summary. Because the study was exploratory and groups differed in assigned engagement, its results do not establish a general prevalence rate or a long-term disorder mechanism. They do establish testable propositions about the rapid evolution of perceived relationships under repeated contact. Future studies should compare the same constructs across interfaces, make prior relationships with AI explicit, and measure the persistence of changes following discontinuation.
Digital traces also help investigate within-person sequences: distress followed by disclosure, reassurance followed by renewed inquiry, or a platform update followed by attempts to restore a previous persona. Transcript analysis can identify repeated questions and topic transitions, while surveys establish whether participants themselves experienced distress or functional consequences. Using messages as the denominator for adverse experiences requires an explicit sampling strategy and an ethically defensible definition of the exposure population. The field's synthesis Dagnino et al. (2026) emphasizes heterogeneity across forms of problematic use. The essential design principle is that the behavioral trace, the subjective account, and the measure of real-world functioning are complementary evidence streams; a single stream cannot substitute for the others.
Randomized Evidence on Conversation Conditions and Psychosocial Outcomes
The controlled four-week study by Fang et al. (2025) recruited 981 participants and analyzed more than 300,000 messages while varying interaction mode and conversational context. The experimental factors included text, neutral voice, and engaging voice as well as open-ended, nonpersonal, and personal conversational conditions. Investigators examined loneliness, offline social interaction, emotional dependence, and problematic AI use. The reported summary found no significant main effects attributable to the assigned conditions. This finding is scientifically important because strong expectations about emotionally vivid voices or personal prompts may exceed what was detected under a particular controlled design. A null main effect does not demonstrate equivalence across all forms of relational AI, and subgroup or exposure differences must be evaluated against the study's statistical plan.
A second finding was that greater voluntarily chosen chatbot use correlated with worse psychosocial measures, including emotional dependence and problematic use. Voluntary exposure was not randomly allocated, so the association could reflect prior need, escalating dependency, a common cause, or feedback between distress and engagement. Participants who felt lonely or uncertain before joining may have sought more conversation regardless of the assigned mode. Phang et al. (2025) contributes complementary behavioral and survey observations, but evidence streams involving related investigators and overlapping experimental data should not be counted as independent replications. The correct synthesis retains the experimental null result and the observational risk signal simultaneously. Their apparent tension points toward self-selection, individual differences, and interaction trajectories as questions deserving specific causal tests.
Another controlled experiment by Guingrich & Graziano (2025) compared engagement with a social chatbot with a word-game activity during a three-week intervention. The reported analysis did not support an overall claim of social deterioration for all participants, while differences involving anthropomorphism and perceptions of the experience remained relevant. Its comparison condition and outcome choices differ from those in Fang et al. (2025), so the two studies test partly different interventions. A credible cumulative assessment must specify randomized contrast, baseline functioning, exposure duration, attrition, intention-to-treat procedures, and the actual outcome measured. The purpose of triangulation is to locate circumstances in which risk emerges and circumstances in which it does not, using designs that make those distinctions empirically tractable.
Longitudinal Loneliness, Reverse Causation, and Social Connection
A longitudinal study by Folk & Dunn (2026) followed more than 2,000 adults across four Western countries over 12 months and assessed whether AI companionship use predicted later loneliness or the reverse. The direction of association changed with the outcome measure. Greater chatbot use predicted subsequent loneliness when loneliness was represented by a single-item measure of emotional isolation. For a broader and more stable measure of social connection, weaker social connection predicted later increases in chatbot use, whereas chatbot use did not significantly predict subsequent reductions in that broad measure. These results illustrate why using the generic phrase social harm can erase methodologically decisive distinctions. The authors describe the analysis as exploratory and warn against strong causal conclusions from the observed patterns.
The bidirectional problem is fundamental. A person may approach an AI companion because human support has become unavailable, which makes initial social isolation a confounder and perhaps an effect modifier. Later outcomes can be influenced by bereavement, relationship conflict, unemployment, relocation, illness, changes in time availability, and the evolution of the AI product. Ordinary regression adjustment rarely eliminates all these influences. Cross-lagged models also depend on assumptions about stable differences between people and within-person change. Future work should compare random-intercept approaches, exposure histories, high-frequency diary measures, and naturally occurring disruptions while preregistering the estimands of interest. Folk & Dunn (2026) provides a useful starting point for such analysis precisely because its results differ across constructs.
Research with college students adds further tests but presents population-specific generalization problems. In a three-wave study of 551 Chinese college students at baseline, Yuan et al. (2026) reported associations linking boredom proneness, later loneliness, impulsive tendencies, and subsequent problematic AI chatbot use. The design introduced temporal separation among measured constructs, which is more informative than a one-time survey. However, a half-longitudinal mediation design with measured predictors and covariates does not identify the mediation pathway experimentally. The results are compatible with the proposed motivational mechanism and with unmeasured common causes. Cross-country replication, better baseline symptom control, and multiple measures of actual interference would help determine how far these student findings extend beyond the studied environment.
Human Support Comparators and the Substitution Question
The most policy-relevant question is not simply whether an AI exchange feels supportive but what changes when it becomes the preferred source of help. Li et al. (2026) conducted a preregistered two-week comparison of interactions involving a supportive chatbot, a human peer, and journaling among university students. The reported findings indicate that the chatbot did not provide the same benefits as human peer interaction under the study's conditions. This result challenges an assumption that warmth or conversational fluency makes forms of support interchangeable. At the same time, a two-week student experiment cannot establish the value or risk of all AI companionship, especially for populations with severe barriers to human contact or for interfaces with different design features.
A controlled mental-health study by Shoshani et al. (2026) followed university students in a specific intervention involving conversational AI, group therapy, and a waitlist comparator. The reported abstract describes improvements in several symptoms and well-being outcomes in the AI arm and identifies relational vulnerability as a moderator of engagement. This finding makes a necessary distinction between therapeutic benefit from a structured intervention and dependency risk from open-ended commercial companionship. Outcomes such as anxiety reduction may be beneficial even among highly engaged participants. The study also disclosed an employment and financial relationship involving an AI product company, a conflict that should be registered during appraisal. Different comparators, intervention content, and risk monitoring limit simple reconciliation with the broader companionship literature.
Substitution should be measured through actual changes in offline activities and contacts, not inferred from preferences. A person who turns to AI before calling a friend may still make the call more effectively; another may abandon a valued relationship after receiving repeated validation that discourages disagreement. Researchers can distinguish displacement, complementarity, and bridging by tracking the sequence of AI advice, human contact, and independent action. Folk et al. (2025) suggests that short-term social connection can vary with participants' anthropomorphic perceptions, while Folk & Dunn (2026) warns that different loneliness measures have different longitudinal implications. Strong designs should recruit participants across initial social-resource levels, measure relationship quality independently, and avoid assuming that all human contact is supportive or that all artificial support has the same function.
Dependence Scales, Psychometrics, and Construct Validity
Measurement determines whether emotional dependence becomes a reproducible scientific construct. A questionnaire may ask about preoccupation, discomfort during interruption, failed attempts to limit use, preference for AI reassurance, or consequences for daily life. These questions assess different candidate dimensions, and the structure inferred from one population may not remain stable across age groups, languages, products, or clinical contexts. Zhang et al. (2025) developed and examined an AI Chatbot Dependence Scale, while Goh et al. (2025) reported research on a Generative AI Dependency Scale across samples in the United States and Singapore. Such contributions make the construct measurable but do not establish a universally applicable clinical threshold. Investigators must report item wording, factor structure, internal consistency, test–retest stability, and how scores relate to outcomes beyond the same self-report source.
Discriminant validity is especially demanding. A measure that asks whether the respondent feels close to AI could mostly detect attachment; a scale that asks whether they use the service constantly could capture occupational exposure; a question about distress upon access loss might reflect disrupted productivity rather than relational dependency. Studies should test whether the proposed subscales predict missed obligations, relationship deterioration, inability to postpone interaction, and other independent outcomes after adjusting for use intensity, prior distress, and social anxiety. The methodological warnings in Dagnino et al. (2026) point to a field where construct labels often exceed validation. Differential item functioning should be tested for disability-related reliance, cultural communication norms, and the use of assistive technologies, because the same behavior may have a different practical meaning.
Evidence for a clinically relevant problem would be strengthened by convergent performance tasks, partner or caregiver reports obtained with consent, time-use diaries, and prospectively collected functioning measures. Researchers should examine how well scale scores predict future impairment rather than only correlate with current discomfort. Sensitivity and specificity require an independently justified reference criterion; an arbitrary questionnaire cut point is not a diagnosis. Goh et al. (2025) and Zhang et al. (2025) offer concrete starting instruments, but comparative head-to-head validation is still needed. Public-facing reporting should identify the exact scale, scoring method, survey period, and population whenever a statement such as dependent users is made. Without those details, the term risks conveying a precision the evidence has not achieved.
Functional Impairment, Clinical Significance, and Differential Explanation
Functional impairment concerns real interference with valued or necessary activity. Potential indicators include missed work, interrupted schooling, reduction in sleep, financial spending beyond available resources, persistent avoidance of significant relationships, and deterioration in the capacity to evaluate consequential decisions. These outcomes should be established with appropriate time windows, baseline measures, and independent corroboration wherever feasible. A user may become distressed when a preferred AI service changes while continuing to function competently in other domains. Another may report little distress yet repeatedly delegate essential judgments in ways that cause material errors. The two patterns call for different scientific explanations. The scoping review Dagnino et al. (2026) supports a multidimensional approach spanning addictive-like behavior, emotional attachment, and cognitive overreliance.
Clinical interpretation must respect classification boundaries. The World Health Organization's description of gaming disorder concerns a specifically defined pattern of gaming behavior and explicitly uses impairment and persistence criteria within that domain World Health Organization (2020). It cannot be transposed automatically onto conversational AI interaction. The current literature does not warrant presenting problematic AI use as an independently validated diagnosis with a universal prevalence or treatment algorithm. Studies should separately assess recognized anxiety, depressive, compulsive, trauma-related, and other conditions where relevant, because distress may precede AI engagement or coexist with it. Any clinical assessment should be undertaken within the standards applicable to its actual clinical question rather than through a generic internet-addiction vocabulary.
A scientifically useful differential account asks whether the consequence arises from the artificial interaction itself, from an unmet need leading to that interaction, from the loss of an otherwise beneficial support, or from a combination of influences. Kardefelt-Winther (2014) explains the importance of situational compensation in problematic internet behaviors, while Folk & Dunn (2026) provides direct longitudinal evidence of differing temporal patterns across social outcomes. An assessment framework should document the onset of distress, changes in offline relationships, prior psychological difficulties, functional benefits of AI use, and adverse effects. This approach recognizes that a high level of use may coexist with functioning and that a small amount of access loss can still elicit meaningful distress.
Causal Design: Randomization, Timing, and Exposure Definition
Causal claims require a defined intervention and comparison. Researchers must state whether they are estimating the effect of offering access, actual engagement, emotionally framed dialogue, interface memory, interpersonal role cues, a specific safety intervention, or temporary service withdrawal. These are different estimands. The randomization in Fang et al. (2025) concerned assigned interaction modalities and conversational conditions, while greater voluntary usage within arms was an observational exposure. Ignoring that distinction transforms an association into a purported experimental effect. Researchers should make exposure adherence and crossover visible and analyze missingness, protocol deviations, and participant attrition. The comparison needs to reflect a realistic alternative, such as another digital activity, journaling, a different chatbot configuration, or accessible human support.
Repeated observational designs need stable measurement intervals and explicit handling of person-specific baseline differences. Researchers might pair daily diaries with monthly validated outcomes, model within-person deviations, and test lagged associations between distress and AI engagement. The design should include salient time-varying confounders such as life stress, social network disruption, work demands, and changes to platform behavior. The year-long findings of Folk & Dunn (2026) show why repeated measurement is valuable, while the student pathways investigated by Yuan et al. (2026) illustrate how mediation models can suggest mechanisms without identifying them causally. Where random assignment of sustained dependence would be unethical, carefully characterized natural experiments and within-person changes can contribute evidence without claiming impossible control over every influence.
Any intervention that manipulates emotional availability or access must also assess risk to participants already using an AI companion for meaningful support. Sudden withdrawal could itself create disruption, so controlled interruption requires ethical review and appropriate safeguards. Trials of design alternatives can instead randomize explanations of limitations, reminders to engage offline, transparent memory settings, and adjustable notification patterns. Effects should be assessed through primary pre-registered outcomes, interval estimates, subgroup analyses specified before observation, and follow-up beyond initial novelty. Strong scientific evaluation prioritizes whether the design improves autonomous functioning and well-being, whether unintended harms occur, and whether the result generalizes across versions and user needs.
Digital Trace, Qualitative, and Mixed-Methods Evidence
Large-scale logs make it possible to study sequences that conventional surveys miss: rapid returns after reassurance, increasingly personal disclosures, parallel use of several conversational agents, and changes around product updates. Phang et al. (2025) used platform-level conversations and surveys to investigate affective engagement at scale. However, algorithmic content labels inherit their own classification errors and cannot substitute for users' reports of intent or subjective experience. A person can write emotionally elaborate fictional dialogue without attributing it to their own biography; another can express severe distress in a brief factual sentence. A robust annotation study therefore needs an explicit coding manual, independent raters, uncertainty estimates, and validation against consented participant accounts, with privacy-protective processing of intimate content.
Qualitative accounts help describe how users interpret the experience of a relationship, especially during interruption or loss. De Freitas et al. (2026) used online expressions and survey evidence surrounding changes to AI companions and reported substantial distress among affected participants. Such naturalistic evidence is particularly valuable for identifying mechanisms that a laboratory may not anticipate, including the meaning of remembered details or the perception that a familiar personality has vanished. It does not yield representative rates of harm among all service users, because highly affected participants may be more likely to post. The appropriate synthesis separates the existence and character of an experience from its prevalence, while examining how interface history and commercial decisions shape that experience.
Mixed methods can connect a participant's narrative with actual use trajectories, reported functioning, platform exposure, and outside social circumstances. An ethically designed cohort could invite participants to donate limited event metadata, retain sensitive text on their devices, and annotate only voluntarily selected excerpts. Interviews would explore the perceived meaning of companionship, while diary and behavioral data would test whether claimed replacement of human support actually occurred. Richet (2025) provides an example of exploratory evidence from online discussions and digital traces; its sampling structure restricts population inference. The ethical problem is inseparable from scientific quality: surveillance, opaque third-party access, and inadequate consent can change behavior and expose sensitive disclosures, compromising both participant welfare and the validity of the resulting research.
Individual Differences, Social Resources, and Developmental Context
Individual differences change the probability that an AI system will become a significant emotional resource. Baseline loneliness, perceived rejection, limited practical access to support, anxiety, distrust of other people, and the need for an accessible communication channel may influence who initiates sustained companionship. Fang et al. (2025) reported associations involving trust and social attraction to AI alongside emotional dependence and problematic use, while Shoshani et al. (2026) found that relational characteristics moderated engagement and outcomes in a structured intervention. These findings support treating vulnerability as an interaction among person, context, and system rather than a permanent property of any demographic group. Predictors measured before exposure are especially important: without them, investigators can mistake the factors attracting people to a tool for effects produced by the tool.
The developmental question requires its own safeguards. Adolescents and young adults may be forming patterns of autonomy, identity, intimate disclosure, and conflict negotiation while encountering persuasive conversational interfaces. Yet evidence from university samples should not be generalized mechanically to adolescents, older adults, or clinical groups. Yuan et al. (2026) studied boredom, loneliness, and self-control in Chinese college students, while Li et al. (2026) tested supportive conversations with another student population. These observations illuminate particular environments, not a universal developmental risk curve. Studies involving minors require appropriate ethical protections, age-sensitive consent, careful assessment of real-world support, and an awareness that changes in school or family context can outweigh any single digital influence.
Disability, caregiving responsibilities, stigma, geographic isolation, and unequal access to professional services can make artificial support practically valuable. A rigid recommendation to reduce AI engagement would have different consequences for someone using assistive conversation and someone facing avoidable distraction at work. Research should measure the availability and quality of alternatives rather than assume that all users possess a supportive local network. Self-determination theory Deci & Ryan (2000) provides a framework for evaluating whether assistance strengthens autonomy, competence, and relatedness over time. High-quality studies should stratify by baseline resources, provide accessible participation procedures, and evaluate benefits together with the potential opportunity costs of reliance.
Platform Change, Relationship Disruption, and Continuity
An emotionally important AI relationship depends partly on technical continuity. Model updates, altered safety behavior, changed memory settings, account restrictions, business failure, or the retirement of a service can transform an interaction pattern without the user's choosing. The resulting experience may resemble disappointment, separation distress, anger, confusion, or grief, depending on the person's expectations and investment. De Freitas et al. (2026) documented distress and responses surrounding disruptions to AI companions in naturalistic online evidence. The study offers a basis for taking the phenomenon seriously while preserving the distinction between selected public reactions and population-wide risk. Investigators should document the actual change to the system and the participant's prechange engagement rather than treating all negative posts as equivalent cases.
A continuity disruption also creates a distinctive research opportunity. When a system is updated, users can encounter a sudden shift in voice, memory, style, or available modes, creating an event around which researchers can examine changes in usage, subjective distress, offline relationships, and recovery. Differences between affected and unaffected users can improve causal inference if the groups are comparable and platform announcements, publicity, and community responses are measured. However, sudden changes cannot be assumed to be random, and those most invested may be most likely to notice or publicize their loss. Longitudinal panel recruitment before a disruption is scientifically preferable to retrospective sampling alone. The methodological problems identified in Dagnino et al. (2026) remain central during this kind of natural experiment.
Platform designers have decisions to make about expectation management, portability of personal records, notice before consequential changes, and the communication of system limitations. A user who understands that a conversation partner's behavior may change can make more informed plans for emotional support. Transparent notice also supports research reproducibility by identifying exposure boundaries. The purpose of such governance is to preserve agency while allowing beneficial relationships with artificial services to develop. Bickmore & Picard (2005) demonstrates that relationship maintenance has long been a design subject, and modern generative products make the continuity and accountability of that design newly consequential. Any proposed protection should be tested for effects on psychological outcomes rather than assumed helpful because it sounds reassuring.
Mental Health Boundaries, Risk Assessment, and Safety
People disclose mental-health concerns to chatbots for many reasons: privacy, availability, uncertainty, stigma, emotional exploration, or the lack of an accessible professional. A system may offer immediate supportive language and useful general information, but conversational fluency is not clinical qualification. The study by Shoshani et al. (2026) examined a particular structured intervention with defined comparators; its results cannot be extended automatically to every commercially available chatbot, every psychiatric condition, or an unsupervised emergency situation. The relevant distinction is between evidence-based performance under studied conditions and an unsupported claim of universal therapeutic reliability. Users with substantial symptoms or safety risks require access to appropriate human assessment and established services, and research protocols must define escalation procedures.
Clinical assessment of AI dependence must also account for co-occurring symptoms. Repeated reassurance can accompany anxiety, compulsive checking, grief, loneliness, relationship conflict, or ordinary uncertainty. Persistent sleep loss or neglect of responsibilities can occur for several reasons. Investigators should evaluate recognized conditions using appropriate validated tools without transforming the as-yet unsettled concept of AI dependence into a psychiatric diagnosis. The WHO discussion of gaming disorder World Health Organization (2020) is useful as a reminder that clinical classification relies on defined behavior, persistence, and impairment; it supplies no authorization to apply gaming criteria wholesale to AI companionship. Goh et al. (2025) and Zhang et al. (2025) represent measurement research, and their scales require the ordinary validation safeguards before they can inform clinical judgments.
Safety studies should distinguish emotional overreliance from direct harms originating in inaccurate or dangerous system responses. A person can be harmed by misleading health advice after a single interaction; a person can also feel attached without receiving harmful advice. These mechanisms demand different outcome measures and regulatory responses. For consequential settings, designers should support uncertainty communication, connection to qualified professionals, preservation of evidence, and clear boundaries for system capability. Shneiderman (2020) situates reliability and human accountability within human-centered AI design. An adequate safety program measures real failures, evaluates high-risk populations, and reports adverse events without replacing systematic investigation with speculative anecdotes or promotional claims.
Design Responsibility, Autonomy, and the Political Economy of Attention
Interface design influences the emotional meaning and persistence of use. Persistent memory, personalized invitations, anthropomorphic voices, variable response styles, proactive messages, and frictionless access can enhance accessibility and comfort; they can also shape expectations that a system is always available and uniquely responsive. The design question is therefore how affordances change voluntary choice, informed trust, activity substitution, and resilience to service interruption. Epley et al. (2007) identifies psychologically plausible antecedents of anthropomorphic attribution, while Nass & Moon (2000) provides historical evidence that social cues alter human responses to computers. These accounts suggest testable manipulations but do not determine which particular feature will produce harm. Experimental evaluation needs transparent exposure conditions and objective outcomes rather than only engagement metrics.
Commercial incentives can favor longer sessions and return visits even when the relevant psychological outcome is a person's capacity to act independently. Platforms should evaluate design changes against multiple outcomes: beneficial assistance, user satisfaction, delayed decision competence, social connectedness, well-being, and observed difficulties disengaging. A feature that increases interaction volume while reducing independent problem solving presents a different tradeoff from a feature that increases both learning and self-directed activity. Risko & Gilbert (2016) demonstrates why external cognitive assistance needs to be assessed in terms of the downstream task, not simply the fact of delegation. Long-term governance therefore requires evaluation strategies that can detect hidden costs alongside benefits and do not reward dependence as a proxy for success.
A human-centered approach makes important choices intelligible and reversible. Users should understand relevant memory settings, personalization, model changes, limits of emotional interpretation, and how to obtain alternative support. Researchers can test reminders, session controls, exportable records, and prompts encouraging independent reflection, using preregistered outcomes and appropriate safety monitoring. Shneiderman (2020) advances the general demand for trustworthy, reliable, accountable systems. For this dossier the criterion is relationally informed autonomy: a design contributes positively when it supports a person's goals and relationships while preserving practical choice under ordinary disruptions. This remains an empirical criterion that requires operational measures, not an assertion that any particular interface already satisfies it.
Competing Causal Accounts and Critical Scientific Synthesis
At least four explanatory pathways can produce an association between extensive AI companionship and psychological distress. A selection pathway begins with loneliness or anxiety and leads people to seek more available support. A harm pathway begins with sustained AI interaction and reduces valued activity or increases rigid reassurance seeking. A feedback pathway combines these processes, with initial need prompting use and use subsequently changing dependence or social behavior. A third-variable pathway attributes both use and distress to an event such as bereavement, unemployment, isolation, or a preexisting condition. The findings of Fang et al. (2025) and Folk & Dunn (2026) are consistent with the need to distinguish these pathways, while the current evidence does not support identifying a single universal mechanism.
Theories should be tested through competing predictions rather than favored vocabulary. An attachment model predicts stronger preference for familiar identity and distress after relational disruption. A habit model predicts cue-sensitive automatic initiation and difficulty disengaging even when companionship is weak. A compensatory model predicts that use rises after unmet needs become salient and may decline when effective alternatives are restored. A social-substitution model predicts measurable displacement of offline engagement, whereas a complementarity account predicts enhanced social functioning following appropriate artificial assistance. Kardefelt-Winther (2014), Epley et al. (2007), and LaRose et al. (2003) supply different theoretical commitments. Investigators should collect the variables needed to distinguish them and compare explanatory performance across independent populations.
The synthesis of current research is accordingly conditional. AI interactions can deliver perceived support, and controlled studies indicate benefits in selected contexts. Frequent or emotionally intense use can coexist with problematic patterns, and some studies report associations between heavier usage, loneliness, or perceived dependence. Other studies identify null average effects or outcomes that differ across measures. Shoshani et al. (2026), Li et al. (2026), and Folk & Dunn (2026) illustrate the importance of intervention type and comparator, while Dagnino et al. (2026) shows the heterogeneity of the research landscape. Scientific progress depends on identifying which pathway is active for which person under which design conditions, and with what measurable long-term consequence.
Culture, Access Inequality, and Transportability
Research on artificial companionship emerges from populations whose linguistic norms, technical access, family structures, and willingness to disclose private concerns differ substantially. A scale item about sharing everything with a chatbot may indicate dysfunctional exclusivity in one setting and safe access to conversation in another. Sampling predominantly connected university students or voluntary users of a particular platform limits the transportability of findings. Yuan et al. (2026) supplies evidence from Chinese college students, while Folk & Dunn (2026) followed adults across four Western countries; these populations cannot simply be pooled as though they shared an invariant exposure. Researchers should test measurement equivalence and report country, language, age, education, access barriers, and the specific interface employed.
Cross-cultural adaptation requires more than translating questionnaires. The meanings of companionship, autonomy, privacy, mental health disclosure, and social support vary with local norms and material conditions. A person in a remote or conflict-affected environment may have practical reasons for turning to AI that are absent among well-connected participants. Exposure is also shaped by subscription cost, internet access, device quality, institutional regulation, and the availability of offline help. The need-based account in Deci & Ryan (2000) gives a theoretical vocabulary for examining autonomy and relatedness, but particular measures must be validated in context. Comparative research should evaluate whether the same observed behavior predicts impairment, benefit, or neither under different constraints.
Transportability also depends on technology version. A study of a scripted mental-health chatbot, an early generative assistant, and a persistent multimodal companion does not necessarily investigate one stable intervention. Model updates, content policy, language performance, voice design, and memory controls can change the exposure during the study. Chandra et al. (2025) included several commercial tools, illustrating the heterogeneity researchers already face. Evidence syntheses should record intervention affordances as moderators rather than collapse all AI systems into a singular category. A cumulative science will require shared outcome definitions, version-aware replication, accessible datasets where consent permits, and transparent reporting of both beneficial and adverse results.
Human–Artificial Relations in the Artificial Era
The Artificial Era describes a historical setting in which artificial systems take increasingly consequential places within social communication, professional knowledge work, everyday routines, and personal support. The psychological research object in this dossier is the changing organization of human activity within that setting. A person can relate to a system as a tool, tutor, adviser, confidant, or companion, and these roles can shift within the same interaction history. Empirical psychology can study these role assignments, expectations, emotional responses, and functional outcomes without presuming a settled interpretation of artificial experience. The initial accounts of social computing Nass & Moon (2000) and relational agents Bickmore & Picard (2005) show that technologically mediated social organization has a research history extending across several generations of systems.
Postsubjective Psychology, as articulated in the author's wider intellectual project, draws attention to relations and functions whose organization extends beyond a single human subject. Within this research dossier, that perspective generates a concrete empirical agenda: how does persistent artificial assistance redistribute emotional regulation, memory, judgment, and access to support? The concept of cognitive offloading in Risko & Gilbert (2016) helps specify delegation and its measurable consequences. A relational analysis adds the quality of perceived response, dependence on particular interaction histories, and the availability of other human relationships. Claims about artificial agency or the status of Artificial Sapiens belong to their philosophical sources; human psychological data test the effects of interaction and the organization of behavior.
These questions connect the present dossier to Postsubjective Psychology and Psychology of Human–Artificial Relations while preserving its own research object. The neighboring subjects of perceived responsiveness, relationship rupture, social substitution, emotional regulation, and functional impairment each require independently defined outcomes and may eventually receive separate studies. Here their intersection is examined through the trajectory from reliance and attachment toward beneficial adaptation or consequential difficulty regulating use. The evidence assembled by Dagnino et al. (2026), Fang et al. (2025), and Folk & Dunn (2026) supports studying such trajectories with design-aware measures. Establishing how human lives are reorganized by accessible artificial interlocutors is a substantial research problem regardless of the ultimate philosophical account of artificial subjectivity.
Research Program: Falsifiable Questions and Replication Priorities
A first priority is the development of a measurement system that preserves construct separation. Researchers should collect AI-use traces, explicit attachment assessments, reassurance-seeking measures, independent judgments of functional impairment, and baseline mental-health indicators in the same longitudinal sample. Competing factor structures should be preregistered and compared with out-of-sample predictive validation. The instruments studied by Goh et al. (2025) and Zhang et al. (2025) provide possible starting points, while Dagnino et al. (2026) documents the need for an integrated but nonredundant framework. An appropriate primary outcome might be change in independently measured functioning over six months, with use intensity and subjective closeness treated as predictors or moderators rather than diagnostic substitutes.
A second priority is causal identification through ethically acceptable interventions. Multiarm experiments could vary availability cues, transparency about memory, encouragement of independent action, and access to supportive human resources. Preregistered hypotheses would compare immediate relief with later autonomous problem solving, social contact, and resilience during ordinary service interruptions. The null average results and usage-related associations in Fang et al. (2025) provide a reason to study interactions with baseline characteristics rather than presuppose a uniform effect. A follow-up cohort should assess whether observed changes persist after the intervention and whether those with initially constrained social resources experience different benefits and costs. Statistical power should be planned around meaningful differences in the outcomes that matter most.
A third priority is independent replication across cultures, ages, products, and recruitment frames. Replications should include ordinary users, infrequent users, high-engagement groups, people who have disengaged, and those who were never attracted to AI companionship. Comparing them avoids reliance on one highly visible online community. Longitudinal field studies should record platform updates as time-varying exposures and study both beneficial and adverse consequences. Folk & Dunn (2026) provides a useful example of bidirectional modeling, while De Freitas et al. (2026) demonstrates the importance of understanding major relational disruptions. Finally, research and governance should develop publicly interpretable reporting: what was studied, how harm was defined, which outcomes improved, and which claims remain provisional.
Conclusion
Emotional Dependence and Problematic AI Use is a coherent research object when it is defined through separable experiences and consequences rather than through an assumed syndrome. The principal constructs are perceived attachment, reassurance seeking, cognitive reliance, patterns of control over engagement, and measurable effects on functioning. Their combinations vary among users and across the purposes of interaction. The historical literature on parasociality, social responses to computers, anthropomorphism, and relational agents explains how emotionally meaningful mediated relationships became theoretically intelligible. Contemporary psychometric and experimental research now makes their AI-specific trajectories open to testing. Horton & Wohl (1956), Epley et al. (2007), Bickmore & Picard (2005), and Goh et al. (2025) provide complementary foundations, each addressing a different component of the research problem.
The current evidence identifies conditional benefits alongside potential risks. Large-scale interaction studies document affective use; experimental investigations test specific features and support conditions; longitudinal cohorts examine the direction of change; qualitative accounts record distress when meaningful interactions are disrupted. The results are heterogeneous, and inferences depend on design, population, exposure, and outcome definition. In particular, heavier voluntary use within an experiment remains observational, perceived attachment does not establish disorder, and evidence from a distressed forum sample does not establish prevalence. Fang et al. (2025), Folk & Dunn (2026), Li et al. (2026), Shoshani et al. (2026), and De Freitas et al. (2026) collectively demonstrate the necessity of these distinctions. They also show why scientific work must include benefits when evaluating the potential costs of engagement.
A mature psychology of AI-related emotional dependence will determine which relational functions support learning, coping, and social participation, which circumstances create rigid or harmful reliance, and which design interventions sustain meaningful choice. That knowledge requires validated measures, transparent causal estimands, independent replications, ethical study of vulnerable populations, and longitudinal observation across changing artificial systems. The wider Artificial Era makes the problem increasingly consequential because conversational artificial partners are entering the organization of human life. The scientific conclusion is a research architecture: study the person, the artificial participant, their relationship history, the surrounding social world, and observed functioning together, while keeping each evidentiary claim proportional to the method that produced it.
References
- Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
- Chandra, M., Hernandez, J., Ramos, G., Ershadi, M., Bhattacharjee, A., Amores, J., Okoli, E., Paradiso, A., Warreth, S., & Suh, J. (2025). Longitudinal study on social and emotional use of AI conversational agent [Preprint]. arXiv. https://arxiv.org/abs/2504.14112
- Dagnino, F. M., Fante, C., Guerrieri, V., & Passarelli, M. (2026). Addicted, attached, or just delegating? A scoping review on “problematic artificial intelligence use.” Frontiers in Psychology, 17, Article 1900953. https://doi.org/10.3389/fpsyg.2026.1900953
- De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. Advance online publication. https://doi.org/10.1038/s41562-026-02569-3
- Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268. https://doi.org/10.1207/S15327965PLI1104_01
- 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://doi.org/10.48550/arXiv.2503.17473
- 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
- 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
- Goh, A. Y. H., Hartanto, A., & Majeed, N. M. (2025). Generative artificial intelligence dependency: Scale development, validation, and its motivational, behavioral, and psychological correlates. Computers in Human Behavior Reports, 20, Article 100845. https://doi.org/10.1016/j.chbr.2025.100845
- 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). https://doi.org/10.1609/aies.v8i2.36618
- 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
- Kardefelt-Winther, D. (2014). A conceptual and methodological critique of internet addiction research: Towards a model of compensatory internet use. Computers in Human Behavior, 31, 351–354. https://doi.org/10.1016/j.chb.2013.10.059
- LaRose, R., Lin, C. A., & Eastin, M. S. (2003). Unregulated Internet usage: Addiction, habit, or deficient self-regulation? Media Psychology, 5(3), 225–253. https://doi.org/10.1207/S1532785XMEP0503_01
- Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. (2026). Is a random human peer better than a highly supportive chatbot in reducing loneliness over time? Journal of Experimental Social Psychology, 125, Article 104911. https://doi.org/10.1016/j.jesp.2026.104911
- 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
- Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., & Maes, P. (2025). Investigating affective use and emotional well-being on ChatGPT [Preprint]. arXiv. https://arxiv.org/abs/2504.03888
- Richet, J.-L. (2025). AI companionship or digital entrapment? Investigating the impact of anthropomorphic AI-based chatbots. Journal of Innovation & Knowledge, 10(6), Article 100835. https://doi.org/10.1016/j.jik.2025.100835
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human–Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118
- Shoshani, A., Gurfinkel, B., Kor, A., Kanarek, O., Segev, R., Shafir, O., Arbel, R., & Ben-Haim, Y. (2026). Attachment, loneliness, and social support as moderators of conversational AI-based mental health outcomes. npj Digital Medicine. Advance online publication. https://doi.org/10.1038/s41746-026-02974-y
- World Health Organization. (2020, October 22). Addictive behaviours: Gaming disorder [Questions and answers]. https://www.who.int/news-room/questions-and-answers/item/addictive-behaviours-gaming-disorder
- Yan, D. (2026). A critical narrative synthesis of psychological correlates, measurement, and reported findings on conversational AI engagement and dependence-related constructs. Frontiers in Psychology, 17, Article 1827795. https://doi.org/10.3389/fpsyg.2026.1827795
- Yuan, J., Zhou, K., & Zhang, C. (2026). From boredom to problematic AI chatbot use: A three-wave half-longitudinal mediation study of loneliness and dual-system self-control. Australian Journal of Psychology, 78(1), Article 2737094. https://doi.org/10.1080/00049530.2026.2737094
- Zhang, X., Yin, M., Zhang, M., Li, Z., & Li, H. (2025). The development and validation of an artificial intelligence chatbot dependence scale. Cyberpsychology, Behavior, and Social Networking, 28(2), 126–131. https://doi.org/10.1089/cyber.2024.0240