Trust-driven healthy engagement with conversational AI for mental health support in young adults: a mixed methods study Authors: Yanling Lan, Sihang Liu, Chao Liu, Hao Chen Published: 2026 ABSTRACT AbstractYoung adults aged 18–35 years increasingly engage with conversational artificial intelligence (C-AI) in everyday contexts in which they may perceive emotional relief, cognitive clarification, and psychological support. However, the mechanisms through which C-AI interactions are associated with users’ perceived psychological health support remain insufficiently understood. Drawing on an extended stimulus-organism-response (S-O-R) framework, this mixed-methods study examined the associations among Conversational AI Perceived Trust (C-AIPT), Conversational AI Perceived Warmth (C-AIPW), Technology Healthy Use (THU), and Perceived Psychological Health Support (PHS). In this study, PHS was conceptualized as users’ subjective appraisal of psychological support arising from everyday C-AI interactions, rather than as mental health-specific use or clinically verified improvement in mental health symptoms. THU was conceptualized not as objectively observed use behavior or usage frequency, but as a self-regulatory orientation toward C-AI engagement encompassing moderate use, reflective judgment, goal orientation, and boundary awareness. Survey data from 1,006 participants with prior C-AI use experience, collected between July and August 2025, were analyzed using partial least squares structural equation modeling (PLS-SEM) with 5,000 bootstrap resamples. These quantitative analyses were complemented by asynchronous in-depth interviews with 50 experienced C-AI users. The results showed that C-AIPT significantly predicted both C-AIPW (β = 0.774, P < 0.001) and THU (β = 0.409, P < 0.001), whereas only THU was directly associated with PHS (β = 0.382, P < 0.001). Mediation analyses identified THU as the primary self-regulatory pathway linking C-AIPT to PHS. In addition, AI autonomy (AIAu) weakened the association between C-AIPT and THU, whereas AI attachment (AIAt) weakened the association between THU and PHS. Qualitative findings further suggest that perceived warmth primarily functions as an interactional cue rather than a direct therapeutic mechanism. Overall, the findings suggest that perceived psychological health support in everyday C-AI use is more closely related to goal-oriented, self-regulated, and boundary-aware healthy engagement than to emotional warmth alone.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57498-y.Keywords: Conversational Artificial Intelligence, Mental Health Support, Technology Healthy Use, Mixed Methods, Human–AI InteractionSubject terms: Health care, Psychology, Psychology Introduction In recent years, mental health concerns among young adults have increasingly been recognized as a global social and public health challenge with substantial structural implications. According to recent statistics from the World Health Organization, approximately one in seven young people experience mental health disorders during development, with increasing symptoms of anxiety and depression reported amid rising educational pressure, evolving social interactions, and rapidly changing digital environments (Zhang et al.,2025)1. Educational difficulties, stigma, and risk behaviors further exacerbate these vulnerabilities, contributing to what has been described as a “youth mental health crisis.” Between 2015 and 2025, the prevalence and diagnosis rates of depression among young populations in countries such as the United States, the United Kingdom, Australia, and China have shown sustained increases, and global economic losses attributable to mental disorders are projected to reach US $16 trillion by 2030 due to reduced productivity and labor output (Bosmans et al.,2025)2. The determinants of young adult mental health are complex and multifaceted. Academic stress has been identified as a key driver of persistent anxiety and learning burnout, depleting cognitive, emotional, and behavioral resources and thereby intensifying depressive and anxiety symptoms (Zhuang et al.,)3. Excessive reliance on digital media further exposes young adults to social comparison and online harassment, both of which have been associated with poorer psychological well-being (Nesi et al.,2021)4. In particular, cyberbullying undermines social trust and amplifies negative emotional experiences, contributing to increased loneliness and diminished self-worth. Moreover, social isolation and chronic stress may erode perceived social support, heighten threat-oriented interpretations of social interactions, and ultimately increase the risk of psychological imbalance and maladjustment among young adults (Yazıcı-Kabadayı et al.,2025)5. Conversational AI (C-AI), encompassing chatbots, conversational agents, and embodied dialogue systems, has emerged as an increasingly prominent digital support modality for mental health support (Wu and Chen, 2025)6. With advances in generative artificial intelligence, human-AI interaction has shifted from rule-based static responses to real-time exchanges driven by large language models, enabling psychological education, emotional support, and behavioral guidance to be delivered in low-barrier and high-frequency formats (Torous and Roberts, 2017)7. Previous studies suggest that certain text-based conversational systems, such as Woebot, Wysa, and Tess, are associated with short-term improvements in depressive and anxiety symptoms and enhanced positive affect; however, existing evidence remains largely based on preliminary trials and short-term evaluations, and no consensus has yet been reached regarding their long-term effectiveness and safety (Magni et al., 2024)8. More importantly, the short-term psychological benefits of C-AI do not necessarily translate into perceived social support or broader psychological support appraisal. Hu et al. (2025)9 found that AI-assisted venting can reduce negative affect, including anger, frustration, and fear, but does not necessarily increase users’ perceived social support. Related research on chatbot use among adolescents has also shown associations between socially supportive chatbot use, higher loneliness, and lower perceived social support (Herbener and Damholdt, 2025)10. These findings suggest that the immediate emotional buffering or conversational comfort provided by C-AI interaction should not be equated with stable perceptions of social support or a broader appraisal of psychological support. Compared with traditional static digital tools, C-AI may enhance user engagement through instant feedback, personalized language generation, and sustained interaction, while also shaping users’ experiences through relational mechanisms such as perceived warmth, trust, and social presence. However, such relational interaction is not a uniformly supportive process. Existing research has highlighted emerging risks associated with C-AI interaction, including overreliance, heightened anthropomorphic perceptions, and blurred human-AI boundaries. In contexts lacking professional oversight or involving intensive interaction, algorithmically generated content may reinforce pre-existing cognitive patterns, while iterative interaction cycles may amplify underlying anxiety or depressive tendencies (Hua et al., 2024)11. Ho et al. (2025)12, in a systematic review of romantic AI companions, identified mixed effects of AI companion relationships: such relationships may provide emotional connection and perceived support, while also generating overreliance, vulnerability to manipulation, erosion of offline interpersonal relationships, privacy risks, and discomfort associated with excessive anthropomorphism. Laestadius et al. (2024)13, in a study of Replika, similarly indicated that emotional dependence on social chatbots may produce mental health harms, particularly when users’ relational expectations exceed the system’s capacity for reciprocal response. Accordingly, C-AI should not be understood as a uniformly beneficial psychological support tool, but as a digital mental health technology with mixed and conditional effects. It can provide immediate emotional buffering and conversational comfort; yet, its longer-term supportive value may be weakened by overreliance, blurred boundaries, displacement of offline support, and reinforcement of maladaptive coping patterns. Anthropomorphism is commonly defined as a cognitive process through which individuals attribute human-like characteristics, intentions, or emotional capacities to non-human technologies during interaction. With advances in generative language models, AI increasingly displays anthropomorphic features through natural language expression, affective responsiveness, and sustained dialogue, thereby exerting a dual influence on Technology Healthy Use (THU). On the one hand, anthropomorphic interaction may enhance users’ perceived emotional support. On the other hand, excessive anthropomorphism may lead to trust miscalibration and emotional attachment, gradually reshaping users’ judgment and self-regulation during repeated engagement. Accordingly, THU in the present study does not refer to the frequency, duration, or intensity of technology use. Rather, it captures users’ ongoing regulation of their engagement with anthropomorphic C-AI, specifically their ability to maintain moderate involvement, reflective judgment, information verification, and boundary control when trust, warmth, and relational cues become salient. Research on digital self-control and digital well-being likewise suggests that healthy technology use is defined less by the intensity of use than by individuals’ capacity to continuously monitor and regulate their interaction with technology while maintaining a balance between technological competence and dependency risk (Monge-Roffarello and De Russis, 202314; Chen et al., 2025)15. When users begin to perceive AI as a socially responsive agent with autonomous capabilities, technology use may shift from instrumental assistance to relational dependence (Klein, 2025)16. In mental health contexts, such a relational shift may redefine the boundaries of healthy engagement and, in turn, shape users’ evaluation of the support derived from the interaction. In this sense, perceived psychological health support (PHS) is treated not as a clinically verified mental health improvement but as users’ perceived emotional relief, cognitive assistance, and psychological resources derived from interacting with C-AI. Existing research indicates that psychological support in digital media environments is often reflected in users’ perceived emotional support and the perceived availability of support resources, whereas the mental health value of AI-based conversational agents is more appropriately understood as perceived support rather than as a stable clinical intervention outcome (Shensa et al., 201617; Li et al., 2023)18. Against this background, AI attachment (AIAt), jointly shaped by anthropomorphic cues and AI autonomy (AIAu), may weaken users’ sense of agency and self-efficacy, thereby undermining the balanced engagement, information verification, and boundary regulation that characterize healthy technology use (THU) (Rheu et al., 2021)19. Furthermore, excessive attachment may shift users’ interaction with technology from task-oriented, functional engagement to emotionally driven, repetitive involvement, reduce users’ tendency to seek real-world help, intensify anxiety-related rumination through confirmatory algorithmic feedback, and ultimately diminish users’ perceived psychological support (Nagata et al., 2026)20. Within digital mental health contexts, anthropomorphism should not be treated merely as a design attribute of conversational artificial intelligence (C-AI); it may also operate as a relational condition within interaction, shaping users’ trust, regulatory engagement, and perceived support appraisal. This study incorporates AI autonomy (AIAu) and AI attachment (AIAt) into an extended stimulus–organism–response (S-O-R) framework, not as general anthropomorphic factors, but as relational boundary conditions operating at different stages of the proposed pathway. Specifically, AIAu, as a system-side autonomy cue, is theorized as a boundary condition in the relationship between C-AIPT and THU. Existing research on AI-assisted decision-making suggests that users may over-rely on AI recommendations under uncertain conditions; thus, when system autonomy increases, users’ trust in C-AI may be expressed more through delegated reliance or default following than through reflective adoption, information verification, and boundary management (Klingbeil et al., 2024)21. By contrast, AIAt, as a user-side relational attachment, is theorized as a boundary condition in the relationship between THU and PHS. Research on social chatbots suggests that sustained interaction may lead users to reinterpret chatbots not only as functional tools but also as relational agents occupying roles such as friend, mentor, or lover, which may further contribute to psychological dependence. Thus, when users develop stronger emotional attachment to C-AI, the way they interpret and derive support from C-AI interaction may shift from functional regulation toward relationship maintenance and emotional reliance, thereby altering the association between THU and perceived support appraisal (Xie et al., 2023)22. Based on this positioning, this study conceptualizes C-AI-based psychological support as a perceived process shaped by relational conditions, rather than as a linear, stable, or inevitable mental health outcome. Theoretically, this study contributes to refining the application of the S-O-R framework in C-AI-based psychological support by clarifying how system-side autonomy and user-side attachment may operate at the “trust–regulatory engagement” and “regulatory engagement–perceived support appraisal” stages, respectively. Practically, the findings provide only tentative considerations for future C-AI research and design evaluation. They suggest that user control, reflective engagement, and relational boundary awareness may be relevant factors to examine in future studies, but the present cross-sectional evidence does not warrant platform-level intervention guidance or public health recommendations. Theoretical framework and research hypothesis Theoretical framework The stimulus-organism-response (S-O-R) model, originally proposed by Mehrabian and Russell (1974)23, explains how external stimuli influence subsequent responses through individuals’ internal cognitive and affective processes. The framework generally consists of three components: stimulus, referring to external cues or information; organism, referring to internal psychological, cognitive, and affective states; and response, referring to subsequent behavioral or psychological outcomes. It has been widely applied to explain digital media use, online learning, and human-AI interaction. Compared with the Technology Acceptance Model (TAM), which mainly emphasizes rational evaluations of technology (Klein, 2025)16, and with frameworks such as the Theory of Planned Behavior (TPB) (Ajzen, 1991)24, the Expectation-Confirmation Model of Information Systems Continuance (ECM-IS) (Bhattacherjee, 2001)25, and Social Cognitive Theory (SCT) (Bandura, 1999)26, which emphasize attitudes, performance expectations, and self-efficacy, the S-O-R framework is better suited to capturing the dynamic pathway linking external cues, internal processing, and subsequent responses (Vieira, 2013)27. Existing research further indicates that the organism layer is not limited to affective states but also includes perceptions, cognitive appraisals, and regulatory processing formed after individuals encounter external stimuli. For example, Bigne et al. revisited the S-O-R model by incorporating automatic processing patterns and conscious perceptions into the analysis of the organism layer, suggesting that this layer captures the internal transformation of external stimuli (Bigne et al., 2020)28. Similarly, Vafaei-Zadeh et al. noted that the strength of the S-O-R framework in digital contexts lies in its ability to integrate affective and cognitive responses within the organism layer, thereby explaining multidimensional mechanisms in complex technology-mediated interactions (Vafaei-Zadeh et al., 2025)29. Based on this logic, this study conceptualizes Conversational AI Perceived Trust (C-AIPT) as an external technological stimulus and positions Conversational AI Perceived Warmth (C-AIPW) and Technology Healthy Use (THU) within the organism layer. In this study, THU does not refer to the frequency, duration, or intensity of C-AI use. Rather, it captures users’ internal regulatory feedback after trust in C-AI is formed, including goal-oriented engagement, moderate involvement, reflective judgment, information verification, self-control, boundary management, and dependency risk control. Existing research on Internet use suggests that the core issue in problematic technology use is not technology use itself, but internal regulatory deficits such as deficient self-regulation, cognitive preoccupation, and impaired control (LaRose et al., 200330; Caplan, 201031; Spada, 201432). In AI-mediated interaction, this regulatory issue is further manifested as judgment calibration and dependency control after trust is established. Klingbeil et al. showed that users may over-rely on AI recommendations in uncertain contexts, even when such recommendations conflict with contextual information or users’ own interests (Klingbeil et al., 2024)33. Bucinca et al. further argued that reducing overreliance in AI-assisted decision-making requires not only system explainability but also users’ reflective judgment and deeper cognitive processing (Bucinca et al., 2021)34. Ma et al. similarly showed that appropriate human-AI collaboration is not grounded in uncritical trust in AI, but in users’ ability to determine when to trust AI and when to rely on their own judgment (Ma et al., 2023)35. Naiseh et al. also indicated that explanations in AI recommendation systems may produce trust calibration errors, suggesting that users still need to verify information and calibrate their adoption of AI recommendations (Naiseh et al., 2021)36. Accordingly, THU in this study is more accurately understood as an internal regulatory feedback and engagement calibration process following the formation of trust in C-AI, rather than as an overt behavioral outcome at the end of the pathway. Correspondingly, this study positions Perceived Psychological Health Support (PHS) within the response layer, not because it is equivalent to an objective behavioral consequence or a clinical treatment effect, but because it represents users’ evaluative appraisal of psychological support after C-AI interaction and internal regulation. In digital mental health contexts, the response layer need not be restricted to observable actions; it may also be reflected in users’ subjective evaluation of support, understanding, emotional buffering, and resource availability. Shensa et al. showed that psychological support in digital media use among young adults can be assessed through users’ perceived emotional support (Shensa et al., 2016)37. Yan et al. further demonstrated that social support in online health communities, especially emotional support, can influence users’ mental health status (Yan et al., 2014)38. Lin et al. also found that problem-focused and emotion-focused support in online support communities can enhance individuals’ coping resources (Lin et al., 2015)39. Together, these studies suggest that, in digital health interactions, psychological support is not limited to objective intervention outcomes but may also be reflected in users’ perceived emotional relief, cognitive assistance, and coping resources. In the context of AI-based conversational agents, Sanjeewa et al.‘s systematic review of empathic conversational agents for mental health showed that user experience, satisfaction, engagement, and perceived support are important indicators for evaluating their supportive role (Sanjeewa et al., 2024)40. Li et al., in a systematic review and meta-analysis of AI-based conversational agents, also emphasized that the mental health value of such systems should be interpreted cautiously in relation to user experience, human-AI relationships, and interaction quality, rather than being equated with stable clinical intervention outcomes (Li et al., 2023)41. On this basis, PHS in this study represents users’ appraisal of psychological support resources rather than a pre-interaction psychological state or an organism-level processing variable. It is a perceived support response formed after C-AI interaction, trust appraisal, perceived warmth, and THU-based regulation, and is reflected in perceived emotional relief, cognitive assistance, and support resources. Thus, within the S-O-R pathway of this study, PHS more accurately corresponds to a response-level support appraisal formed after interaction and internal processing. Furthermore, anthropomorphism-related factors, including AI autonomy (AIAu) and AI attachment (AIAt), are incorporated to extend the S-O-R model and enhance its theoretical precision and contextual relevance in digital mental health settings. By integrating these factors, this study further specifies how technological cues are linked to perceived support responses through affective-cognitive processing and internal regulatory feedback in C-AI interactions. Anthropomorphism refers to the cognitive process through which individuals attribute human-like characteristics, motivations, and psychological states to non-human entities (Heider and Simmel, 1944)42. In intelligent human-AI interaction, anthropomorphism in C-AI emerges when users perceive conversational systems as socially responsive agents through human-like linguistic cues, relational experiences, and perceived social presence (Ma et al., 2025)43. With advances in large language models, C-AI has moved beyond scripted responses toward real-time, context-sensitive generative interaction. In frequent-use contexts, AI autonomy (AIAu) has emerged as a critical dimension of anthropomorphic C-AI interaction: moderate autonomous responsiveness may enhance interaction continuity, whereas excessive autonomy or excessive anthropomorphism may reduce perceived controllability and evoke “uncanny valley” discomfort, thereby attenuating the pathway from trust to calibrated and self-regulated engagement (Lee and Yong, 2024)44. Moreover, AI attachment (AIAt), shaped by anthropomorphic cues and perceived autonomy, has been identified as a key relational factor shaping the quality and regulation of C-AI engagement. Elevated levels of attachment may shift engagement from task-oriented problem-solving toward emotion-driven reliance, potentially weakening the association between technology healthy use (THU) and perceived psychological health support (PHS) (Schillings et al., 2024)45. Building on this perspective, this study integrates AIAu and AIAt into the S-O-R framework as relational boundary conditions for examining C-AI use in digital mental health contexts. This framework explains how technological cues are linked to perceived support responses through intertwined affective, cognitive, and regulatory processes. The proposed model not only clarifies the roles of THU and conversational AI perceived warmth (C-AIPW) within the relationship between C-AI trust and perceived support but also highlights how AIAu and AIAt function as critical moderators, shaping the pathway from C-AI trust to internal regulatory engagement and from THU to perceived support appraisal. By integrating perspectives from artificial intelligence, mental health, public health, and behavioral science, this study refines the application of the S–O–R framework to everyday C-AI use by clarifying how trust, perceived warmth, self-regulated use, autonomy, and attachment are associated with perceived psychological support. Research question and hypothesis Within digital mental health support contexts, conversational AI (C-AI) is increasingly moving beyond a purely instrumental technology to function as an interactive agent characterized by autonomous responsiveness and relational attributes. This shift raises new theoretical questions about how young adults form perceptions of psychological health support through C-AI interaction. On the one hand, prior research generally suggests that when technology is perceived as reliable and functionally valuable, user trust may support more reflective engagement, boundary management, and regulated use. On the other hand, anthropomorphic interaction, reinforced by autonomy cues and relational attachment, may reshape users’ judgments about technological control, interaction boundaries, and sources of support, potentially shifting C-AI engagement from functional assistance toward relational involvement. Accordingly, young adults’ engagement with C-AI may be shaped by two sets of mechanisms: a regulatory pathway centered on reflective adoption, information verification, and boundary management, and relational boundary conditions represented by AI autonomy (AIAu) and AI attachment (AIAt). Against this theoretical background, this study proposes the following research questions: RQ1: Under conditions of higher AI autonomy (AIAu), does conversational AI perceived trust (C-AIPT) remain positively associated with Technology Healthy Use (THU), or is this association attenuated, suggesting that users may shift from self-regulated and boundary-managed engagement toward more delegated or relationally driven engagement? RQ2: During ongoing interaction with conversational AI among young adults, does Technology Healthy Use (THU) remain positively associated with Perceived Psychological Health Support (PHS), or does stronger AI attachment (AIAt) attenuate this association by redirecting engagement toward relational dependence and weakening the perceived support value associated with THU? Conversational AI perceived trust (C-AIPT) represents an important antecedent of users’ interactional appraisals and regulatory engagement (Seo, 2022)46. In the present study, C-AIPT refers to users’ cognitive trust judgments formed after evaluating system competence, information reliability, and potential technological risks, rather than affective responses or relational investment. Within the S-O-R framework, C-AIPT is conceptualized as an external technological stimulus associated with subsequent affective interpretation and regulatory engagement, rather than as an organism-level response itself. From the perspectives of uncertainty reduction and expectation confirmation, higher levels of C-AIPT may be associated with lower perceived ambiguity about system responses and with a greater tendency to interpret C-AI outputs through a benevolent and supportive lens. This pattern may correspond to higher levels of conversational AI perceived warmth (Zhang et al., 2024)47. In addition, C-AIPT is not merely associated with increased usage frequency. By reducing interactional uncertainty and the burden of monitoring, C-AIPT may provide cognitive conditions for goal-oriented engagement, pacing regulation, and reflective adoption, which are central to technology healthy use (THU) (Khan and Shehawy, 2025)48. Prior research also suggests that cognitive trust in conversational systems is generally associated with higher perceived warmth and more reflective, self-regulated forms of engagement, potentially through reduced perceived risk and cognitive load (Kong et al., 2024)49. Accordingly, the following hypotheses are proposed: H1a: Conversational AI perceived trust is positively associated with Conversational AI perceived warmth. H1b: Conversational AI perceived trust is positively associated with technology healthy use. Within intelligent human–AI interaction contexts, conversational AI perceived warmth does not imply that AI possesses genuine emotions, nor is it equivalent to anthropomorphism. Whereas anthropomorphism emphasizes users’ attribution of human-like qualities to technological agents, conversational AI perceived warmth focuses on users’ affective evaluations formed through interactional tone, conversational pacing, and supportive language cues (Fitzpatrick, 2017)50. Distinct from conversational AI perceived trust, which captures cognitive judgments regarding technological competence, information reliability, and system-related risks, conversational AI perceived warmth can be understood as a proximal organism-level affective response. It reflects how users interpret C-AI interaction at the emotional level and may be associated with subsequent regulatory engagement and perceived support appraisal (Croes and Antheunis, 2021)51. In this study, conversational AI perceived warmth is conceptualized as users’ socio-emotional evaluation of a supportive communication climate, rather than as evidence of genuine empathy or autonomous agency in AI. When users perceive C-AI responses as nonjudgmental and supportive, they may experience lower interactional anxiety and evaluative threat, a pattern that may be associated with greater self-disclosure and relational engagement. Related research suggests that supportive interaction climates may be linked to perceived psychological health support through emotional regulation and perceived social support processes (Fulmer et al., 2018)52. At the same time, warm interaction should be distinguished from emotional dependence. Compared with purely instrumental feedback, warm interaction, as a supportive communication cue, may be associated with more reflective and self-regulated technology healthy use, partly because it may reduce perceived threat and stress burden (Bickmore and Picard, 2005)53. Furthermore, although conversational AI perceived trust primarily reflects users’ cognitive evaluations of system reliability and technical capability, trust alone may not fully account for perceived psychological health support. Whether users also perceive interactional warmth may shape the extent to which they interpret system responses as benevolent and supportive resources during interaction (Croes and Antheunis, 2021)54. Accordingly, the following hypotheses are proposed: H2a: conversational AI perceived warmth is positively associated with technology healthy use. H2b: conversational AI perceived warmth is positively associated with perceived psychological health support. H2c: conversational AI perceived warmth mediates the association between conversational AI perceived trust and perceived psychological health support. In C-AI use contexts, technology healthy use (THU) refers to a self-regulated, goal-aligned orientation toward engagement formed through users’ processing of cognitive and affective cues, with an emphasis on the quality of interaction rather than usage frequency or intensity (Rapp et al., 2021)55. In this study, THU is conceptualized not as a normative description of an “ideal” pattern of technology use, but as a context-specific psychological regulatory construct in C-AI-based mental health support contexts. Specifically, THU reflects users’ goal-directed, boundary-aware, and risk-sensitive regulation of C-AI engagement, including moderate involvement, reflective judgment, information verification, boundary management, and dependency risk control. Unlike general self-regulation, THU does not refer to a broad, cross-contextual capacity for self-control. Research on digital self-control primarily focuses on users’ regulation of digital distractions, impulsive use, and attentional interference, whereas THU emphasizes how these regulatory mechanisms are enacted in trust-based C-AI interaction, particularly when users respond to system suggestions, emotional feedback, and relational cues (Biedermann et al., 2021)56. Unlike digital literacy or AI literacy, THU does not primarily concern whether users possess the knowledge or skills required to understand, evaluate, and use digital technologies or AI systems. Research on digital competence and AI literacy focuses more on competence structures, technological understanding, practical application, evaluative judgment, and ethical awareness, whereas THU concerns whether these capacities are enacted in actual C-AI interaction through reflective adoption, information verification, and boundary management (Zhao et al., 202157; Ng et al., 2021)58. Unlike adaptive technology use, THU is not limited to adjustments in technology use according to task demands, system functions, or contextual feedback. User adaptation in the post-adoption stage emphasizes task–technology fit, affective appraisal, and continued use, whereas THU further highlights trust calibration, boundaries of support, and dependency risk control in digital mental health contexts (Zhou et al., 2024)59. Therefore, THU in this study does not represent a general self-regulatory capacity, a level of digital literacy, or an adaptive technology use behavior. Rather, it is an internal regulatory construct situated between technological stimuli and perceived support responses. The core of THU lies not in greater use or normatively preferred use, but in users’ capacity to sustain reflective adoption, verify information, maintain relational boundaries, and control dependency risks when trust, warmth, and relational cues become salient. Furthermore, THU differs from perceived usefulness and perceived ease of use in the technology acceptance model. The latter two constructs primarily reflect users’ evaluations of instrumental utility and operational effort, whereas THU focuses on how users engage with C-AI in a reflective and boundary-aware manner. Based on this understanding, THU can be viewed as a regulatory orientation toward engagement that may help clarify the association between cognitive appraisal and perceived psychological health support (Skjuve, 2022)60. Prior research also suggests that structured and supportive human–AI interaction may be associated with paced and goal-directed engagement, and that self-regulated interaction patterns may be linked to emotional regulation, perceived support, and perceived psychological health support (Qin et al., 2025)61. In addition, conversational AI perceived trust, as users’ cognitive judgment of system capability and information reliability, may be associated with lower interactional uncertainty and a reduced burden of monitoring, as well as with higher levels of reflective adoption and self-regulated engagement. In this sense, technology healthy use (THU) can be understood as an internal regulatory process relevant to the association between users’ trust in C-AI and perceived psychological health support (Lv et al., 2022)62. Accordingly, the following hypotheses are proposed: H3a: technology healthy use is positively associated with perceived psychological health support. H3b: technology healthy use mediates the association between conversational AI perceived trust and perceived psychological health support. In C-AI interactions, conversational AI perceived trust alone may not be sufficient to account for perceived psychological health support (Li et al., 2023)63. Users often construct relational judgments through affective interpretations of interactional intent. When users develop higher levels of trust in C-AI, they may be more likely to interpret system responses as supportive socioemotional cues, and this pattern may be associated with higher conversational AI perceived warmth (Chin et al., 2023)64. As an organism-level affective response, conversational AI perceived warmth may be associated with lower evaluative threat and interaction-related stress, as well as with self-regulation-oriented patterns of technology healthy use. Research on stress buffering and emotion regulation similarly indicates that supportive interaction climates and health-oriented technology engagement may be linked to the integration of psychological resources and perceived support, both of which are relevant to perceived psychological health support (Provoost et al., 2017)65. Accordingly, the following hypothesis is proposed: H3c: conversational AI perceived warmth and technology healthy use jointly exert a significant serial mediating effect on the association between conversational AI perceived trust and perceived psychological health support. In C-AI interaction contexts, AI autonomy refers to users’ perceptions of the system’s behavioral agency, autonomous responsiveness, and capacity to guide the progression of interaction. Unlike AI anthropomorphism or affective cues, AI autonomy more directly concerns the extent to which the system undertakes action initiation, suggestion generation, and decision guidance during interaction, and may therefore shape users’ judgments about the allocation of control, monitoring responsibility, and boundaries of reliance (Laitinen and Sahlgren, 2021)66. Related research suggests that the autonomy level of AI agents can alter human–AI interaction patterns, information-sharing requirements, and perceptions of the AI teammate (Hauptman et al., 2023)67. In research on AI service adoption, Frank et al. (2022)68 also found that AI autonomy moderates the relationship between trust and AI service adoption, and that higher autonomy weakens the positive association between trust and adoption intention. Similarly, Choo and Nam (2022)69 noted that although competence-based trust is generally associated with lower demands for continuous monitoring and greater interaction efficiency, greater system autonomy may shift the allocation of human–AI control from user-led to system-driven, which may in turn alter how trust is expressed in later interactional engagement. Existing research thus suggests that the key issue associated with AI autonomy is not whether trust is formed, but whether trust is expressed through reflective adoption, information verification, and boundary management, or through delegated reliance and default following. Research on autonomous AI delegation indicates that delegating decisions to autonomous AI involves a transfer of power and a loss of perceived control, making trust more likely to be expressed as delegated reliance rather than active monitoring and selective adoption (Candrian and Scherer, 2022)70. Moreover, research on human–AI delegation shows that system-initiated delegation increases users’ self-threat and reduces their willingness to accept delegation, and that this effect is stronger when users perceive lower levels of control (Adam et al., 2024)71. Diebel et al. (2026)72 likewise showed that stronger system-driven decision processes may induce a default-following tendency, which may reduce users’ active regulation of AI suggestions. At the same time, as system autonomy increases, effective interaction and oversight place greater demands on situation awareness, and calibrated trust is more likely to emerge when users retain opportunities for supervision, intervention, and correction (Endsley, 202373; Schwartz et al., 202274. In this sense, conversational AI perceived trust reflects users’ cognitive judgments of system capability and information reliability, whereas technology healthy use captures reflective adoption, information verification, boundary management, and dependency risk control in C-AI interaction. Higher levels of AI autonomy may alter how trust is expressed in interaction and, in turn, weaken the positive association between conversational AI perceived trust and technology healthy use. Based on this reasoning, the following hypothesis is proposed: H4: AI autonomy negatively moderates the association between conversational AI perceived trust and technology healthy use. Within conversational AI interaction contexts, AI attachment is theorized as a relatively stable emotional bond and psychological reliance associated with sustained human-AI interaction, reflected in users’ affective expectations of system companionship and motivation for continued engagement. Unlike conversational AI perceived trust, anthropomorphism, or conversational AI perceived warmth, AI attachment represents relational investment and psychological connectedness rather than users’ cognitive evaluation of technical capability or informational reliability (Laestadius et al., 2024)75. This conceptual distinction is also supported by longitudinal and human-chatbot relationship research, which shows that relationship-oriented responses to chatbots tend to emerge through repeated interactions and that users may experience chatbots as relational partners offering companionship and socio-emotional value (Croes and Antheunis, 202176; Skjuve et al., 2021)77. Therefore, AI attachment is specified as a moderator of the association between technology healthy use and psychological health support because it is more closely related to how users perceive and emotionally organize the psychological value of AI-mediated engagement than to their initial evaluations of trust, warmth, or anthropomorphic cues. Technology healthy use is expected to be positively associated with psychological health support because structured engagement, self-regulation, and goal-aligned interaction may help users experience conversational AI as a bounded and purposeful psychological support resource (Abd-Alrazaq et al., 2020)78. Prior research indicates that companion chatbots can provide everyday forms of social support, including companionship, emotional support, informational support, and appraisal support, while systematic review evidence also emphasizes that the psychological benefits of mental health chatbots should be interpreted in relation to safety, clinical evidence, and appropriate use boundaries (Ta et al., 2020)79. However, higher AI attachment may be associated with stronger relationship-maintenance orientation, greater emotional investment, and dependence-oriented interaction patterns, which may weaken the positive association between technology healthy use and psychological health support (Pentina et al., 2023)80. When users perceive conversational AI as a primary emotional support source without clear boundaries or self-regulatory frameworks, attachment-oriented engagement may become less aligned with goal-directed self-regulation, making reflective processing and active regulation less salient (Nißen et al., 202281; Yirmiya and Fonagy, 2025)82. Thus, AI attachment is not treated as a direct determinant of psychological outcomes but as a contextual moderator that may condition the strength of the association between technology healthy use and psychological health support (Balan and Gumpel, 2025)83. Based on this reasoning, the following hypothesis is proposed: H5: AI attachment negatively moderates the association between technology healthy use and psychological health support, such that this positive association is weaker when AI attachment is higher Fig. 1. Conceptual Model and Research Hypotheses. Methods Study design This study adopted a mixed-methods design grounded in the stimulus-organism-response (S-O-R) framework to examine how young adults develop perceived psychological health support through everyday interactions with conversational AI (C-AI). A large-scale online survey was first conducted to test structural relationships and moderation effects within the S-O-R model, followed by qualitative interviews that contextualized the psychological processes and behavioral motivations underlying the statistical findings. This sequential explanatory design enabled the integration of theoretical modeling with experiential accounts. In the quantitative phase, structured measurement scales were used to operationalize variables across the stimulus, organism, and response components and to test the moderating roles of AI autonomy and AI attachment within the proposed pathways. The primary goal of this phase was to identify structural relationships among variables and evaluate the overall fit of the theoretical model. In the qualitative phase, purposive sampling was used to recruit 50 young adult participants with extensive C-AI use experience from the pool of valid survey respondents. Semi-structured, text-based asynchronous interviews explored how participants perceived AI autonomy, how they described the development of AI attachment, and how these experiences shaped their interaction behaviors and subjective appraisals of psychological support. Qualitative data were used as an explanatory complement to deepen understanding of the transition from organism-level processes to response-level appraisals within the S-O-R mechanism, rather than for statistical inference. In this study, quantitative and qualitative components were not treated as separate stages but were integrated within a unified theoretical framework. The quantitative phase focused on structural modeling to examine the pathways linking conversational AI perceived trust, technology healthy use, and perceived psychological health support, as well as the moderating roles of AI autonomy and AI attachment in everyday C-AI interactions. The qualitative interviews extended the quantitative findings by clarifying how participants interpreted certain C-AI interactions as psychologically supportive and by enriching the contextual understanding of the statistical results. Specifically, quantitative analyses identified overall structural relationships, whereas qualitative narratives provided insight into the psychological processes and behavioral motivations shaping these associations. Integration was conducted during the interpretation stage by comparing statistical patterns with emergent interview themes, thereby enabling a multidimensional understanding of the human-AI relational mechanisms involved in perceived psychological support. Participants and procedure This study defined perceived psychological health support as an interaction-derived appraisal of support in everyday C-AI use, rather than as an indicator of therapeutic use or clinically verified improvement. Specifically, PHS captures the extent to which users experience C-AI exchanges as providing emotional relief, cognitive clarification, perceived understanding, and coping-related resources. On this basis, the study examined whether Chinese young adults, drawing on their general prior experience with C-AI, interpreted system responses, feedback, and sustained dialogue as psychologically supportive. Given variations in institutional definitions of young adulthood across global and regional statistical standards, and considering the active engagement of individuals aged 18–35 years with digital media, intelligent interactive applications, and online support resources, this study operationally defined young adults as individuals aged 18–35 years. Participants were recruited through an online sample panel and completed the survey via Questionnaire Star, a large-scale Chinese survey platform with approximately 2.6 million registered users across diverse regions and social backgrounds. Screening items were embedded at the beginning of the survey to ensure eligibility; only participants who met the inclusion criteria could proceed to the full questionnaire. Eligibility criteria included: (1) being aged between 18 and 35 years; (2) having prior experience using C-AI; and (3) being able to understand and complete a Chinese-language online survey. Participants who did not meet these criteria were automatically excluded. Use of C-AI specifically for mental health purposes was not required as an inclusion criterion because the study did not sample help-seeking behavior in dedicated digital mental health services. Instead, the sampling criterion was intended to identify young adults with sufficient exposure to C-AI interaction, while PHS was used to capture their subsequent appraisal of whether such interactions provided psychological support value. Thus, the study did not evaluate the effectiveness of dedicated mental health chatbots, clinical intervention systems, or AI-based therapeutic tools. Rather, it examined whether general C-AI users perceived emotional relief, cognitive assistance, a sense of being understood, and coping-related support during everyday interaction. Upon completion, respondents received standardized compensation of 10 RMB, approximately US $1.3–1.5, for their time, regardless of their responses or the study hypotheses. Formal data collection was conducted between July 8 and August 8, 2025. A total of 1,097 responses were obtained. Based on predefined data quality standards, responses completed in less than 60 s were considered invalid and removed, resulting in the exclusion of 91 responses. The final analytic sample comprised 1,006 valid responses, with a valid response rate of 91.7%, exceeding the minimum sample size requirement. All retained responses passed duplicate-submission checks, logical consistency checks, and eligibility screening prior to statistical analysis. For the qualitative component, purposive sampling was used to recruit 50 participants with substantial C-AI use experience from the valid survey pool for follow-up asynchronous text-based in-depth interviews. These interviews were used to unpack the interpretive process through which support appraisals emerged in ordinary C-AI use. In particular, they clarified how general C-AI users made sense of certain system responses as sources of emotional relief, cognitive assistance, relational understanding, or coping resources, thereby complementing the quantitative analysis with contextual evidence on how everyday interaction experiences may acquire perceived psychological support value. For measurement development, this study adopted a contextual adaptation process based on established scales. Three refinement stages were conducted prior to formal data collection. First, three scholars with expertise in human-computer interaction and digital mental health evaluated content validity and revised items with semantic ambiguity or contextual mismatch. Second, a cognitive pretest with 30 young C-AI users was conducted to gather comprehension feedback and refine wording that was overly abstract or involved double-negative phrasing. Third, preliminary reliability analysis of the pilot sample indicated acceptable internal consistency, with Cronbach’s α values exceeding 0.70 across constructs. In the measurement design, PHS items were framed to ask respondents to evaluate support-related qualities of their own C-AI interactions, such as feeling emotionally relieved, understood, cognitively assisted, or better equipped to cope. These items were not intended to assess clinical benefits, professional counseling outcomes, or the use of C-AI as a dedicated mental health service. Accordingly, the rationale for using PHS as an outcome rests on the conceptual and measurement distinction between mental health-specific service use and perceived psychological support arising from everyday C-AI interaction. Detailed questionnaire items and raw measurement materials are provided in Supplementary Information 1 and 2. To ensure adequate statistical power for model testing, an a priori minimum sample size estimation was conducted using G*Power 3.1.9.7, following the effect size guidelines proposed by Faul et al., (2009)84. Given that the proposed model included both mediation and moderation paths, a conservative estimation strategy was adopted. The analysis type was specified as F tests—Linear multiple regression: Fixed model, R² deviation from zero, and the calculation was based on the structural equation containing the largest number of incoming paths to a single endogenous construct. In the present model, both technology health use and psychological health support included a maximum of three predictors (including interaction terms); therefore, the number of predictors was set to u = 3. The parameters were defined as follows: significance level α = 0.05 and statistical power = 0.90. Based on Cohen’s effect size classification, small (f² = 0.02), medium (f² = 0.15), and large (f² = 0.35) effects were examined. The results indicated that the minimum required sample size was N = 713 for detecting small effects, N = 99 for medium effects, and N = 45 for large effects. Considering that moderation effects in digital mental health research are typically small, the small-effect criterion was adopted as the reference benchmark to ensure sufficient statistical power and robustness of the model estimation. Measurement The measurement instruments used in this study were contextually adapted from established theoretical frameworks and established measurement scales rather than directly adopting the original items without modification. Given that the study focused on Chinese young adults’ everyday interactions with C-AI, particularly their trust, perceived warmth, regulatory use orientation, and perceived psychological support, the research team first clarified the conceptual boundaries of each construct based on the theoretical model. Relevant items were then selected and adjusted to ensure consistency between the measurement content and the construct definitions used in this study. For scale items derived from English-language sources, two researchers holding doctoral degrees and bilingual research experience independently translated the items into Chinese. The translated versions were compared and reconciled through discussion. Another researcher holding a doctoral degree and bilingual research experience then conducted back-translation and semantic review. The research team compared the back-translated items with the original English items to identify potential semantic shifts, loss of construct meaning, or cultural-contextual mismatches. Items that were not fully aligned with the C-AI interaction context were contextually revised according to the intended construct meaning, so that they could more accurately reflect Chinese young adults’ perceptions and interaction experiences in everyday C-AI use. For items that were originally available in Chinese or developed based on the Chinese context, the research team refined the wording and adjusted the contextual framing to avoid overly abstract wording, semantic overlap, or deviation from the C-AI interaction context. Before formal data collection, expert review and cognitive pretesting were conducted to further enhance the content appropriateness of the measurement instruments. Three scholars with expertise in human–computer interaction, digital mental health, and quantitative research independently evaluated the items, focusing on whether each item was consistent with its corresponding construct and whether semantic ambiguity, conceptual overlap, or cultural-contextual mismatch was present. Subsequently, 30 young adults with C-AI use experience participated in a cognitive pretest. Their feedback on item comprehensibility, clarity, and contextual relevance was used to further revise the questionnaire. After formal data collection, the adapted measurement instruments were assessed for reliability and construct validity. Internal consistency reliability was examined using Cronbach’s alpha and composite reliability (CR). Convergent validity was assessed using indicator outer loadings and average variance extracted (AVE), whereas discriminant validity was evaluated using the Fornell–Larcker criterion. The results indicated that the adapted scales showed acceptable internal consistency, convergent validity, and discriminant validity based on the Fornell–Larcker criterion in the present sample. It should be noted that this procedure represents contextual adaptation and psychometric assessment within the current sample based on existing theories and established scales, rather than the development and full validation of new scales across independent samples. Accordingly, the Limitations section further acknowledges that emerging human–AI relational constructs, particularly AIAt, require further examination using validated AI-specific attachment measures in future research. Detailed reliability, convergent validity, and Fornell–Larcker discriminant validity results are reported in Tables 1 and 2. Table 1Demographic and C-AI use characteristics of interview participants.VariablesOptionsn (%)(N = 50)GenderMale23 (46.0%)Female27 (54.0%)Age18–225 (10.0%)23–2620 (40.0%)27–3019 (38.0%)31–356 (12.0%)EducationUndergraduate11 (22.0%)Bachelor28 (56.0%)Postgraduate11 (22.0%)Duration of C-AI use1 year6 (12.0%)2 years30 (60.0%)≥ 3 years14 (28.0%) Demographic and C-AI use characteristics of interview participants. Table 2Convergent validity assessment.VariableItemsFLCACRAVEVIFConversational AI Perceived TrustC-AIPT10.7920.8250.8840.6561.666C-AIPT20.7871.656C-AIPT30.8191.770C-AIPT40.841.944Conversational AI Perceived WarmthC-AIPW10.8170.8470.8970.6851.786C-AIPW20.8411.999C-AIPW30.8331.954C-AIPW40.8191.856Technology Healthy UseTHU10.6530.7880.8630.6151.272THU20.8281.859THU30.8271.788THU40.8151.711AI AutonomyAIAu10.8720.8180.8920.7331.000AIAu20.8301.666AIAu30.8671.921AI AttachmentAIAt10.7050.8560.8930.6261.000AIAt20.7931.880AIAt30.8271.988AIAt40.7971.768AIAt50.8281.847Psychological Health SupportPHS10.8630.9580.9630.7243.290PHS20.8503.162PHS30.8623.294PHS40.8372.991PHS50.8623.219PHS60.8633.189PHS70.8512.991PHS80.8312.742PHS90.8463.181PHS100.8443.105 Convergent validity assessment. Conversational AI perceived trust (C-AIPT) The conversational AI perceived trust scale consisted of 4 measurement items adapted and integrated from established technology trust frameworks, automation trust models (Hoff and Bashir, 2015)85, and prior research on trust in artificial intelligence service contexts. The original scales were primarily designed to assess individuals’ perceptions of system competence, reliability, and intentionality (Flavian and Casalo, 2021)86. In this study, the item wording was contextually refined to reflect C-AI interaction scenarios while maintaining conceptual equivalence with the original construct, allowing assessment of users’ overall perceived trust toward C-AI. Representative items included statements such as “I believe the information provided by C-AI is reliable” and “I trust that C-AI can effectively respond to my needs.” All items were measured using a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). The scale demonstrated acceptable internal consistency (Cronbach α = 0.825; mean = 3.755, SD = 0.622). Conversational AI perceived warmth (C-AIPW) The conversational AI perceived warmth scale was adapted from the stereotype content model, which conceptualizes competence and warmth as core dimensions shaping individuals’ impressions of social agents. Carpinella et al., (2017)87 further extended this framework by developing the Robotic Social Attributes Scale (RoSAS), incorporating psychological attributes such as sincerity and benevolence to better capture emotional communication in human–robot interaction. Building on these foundational studies, Li et al., (2025)88 introduced the concept of conversational AI perceived warmth in the context of social chatbots and examined the role of perceived warmth in users’ behavioral intentions. Following this approach, the present study measured perceived warmth during C-AI interactions using a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). The conversational AI perceived warmth scale consisted of 4 items assessing users’ subjective emotional warmth experiences toward C-AI. The scale demonstrated acceptable reliability (Cronbach α = 0.847; mean = 3.811, SD = 0.672). Technology healthy use (THU) The technology healthy use scale was developed through contextual adaptation of item construction approaches from behavioral intention measurement research (Trafimow and Finlay, 1996)89 and relevant scale designs commonly used in artificial intelligence adoption research (Wang and Chen, 2024)90. Following established principles of construct measurement, scale adaptation should be guided by a clearly specified construct domain, with item content adjusted according to how the construct is manifested in a specific research context (MacKenzie et al., 2011)91. Accordingly, this study did not measure THU as usage intention, usage frequency, or an overt behavioral outcome. Instead, consistent with its positioning in the organism layer of the S-O-R framework, the THU scale was used to assess users’ perceived regulatory capacity during C-AI engagement. Specifically, the adaptation did not merely substitute C-AI for the original technology context; rather, it shifted the measurement focus from general technology-use tendencies to reflective regulation, moderate involvement, independent judgment, information verification, and boundary management following the formation of trust in C-AI. Research on digital self-control also suggests that healthy digital engagement depends less on usage time or exposure intensity than on individuals’ ability to maintain regulatory control over technology interaction through goal setting, usage monitoring, and distraction management (Biedermann et al., 2021)92. Therefore, the THU scale in this study was designed to measure users’ reflective regulation and boundary management during C-AI engagement, rather than their endorsement of an “ideal” pattern of technology use. The scale consisted of four items, all measured on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). The scale demonstrated acceptable internal consistency (Cronbach’s α = 0.788; mean = 3.634, SD = 0.681). Psychological health support (PHS) The measurement of perceived psychological health support was adapted from established frameworks in digital mental health and online emotional support research (Provoost et al., 2017)93, and was grounded in established theories of social and emotional support. Previous studies have identified perceived emotional and psychological support as key dimensions reflecting the availability of emotional regulation and buffering resources in digital interaction environments (Chin et al., 2022)94. Research on young adults’ digital media use also suggests that perceived emotional support can be assessed as a subjective appraisal of support in digital interaction environments, rather than being limited to formal mental health service contexts (Shensa et al., 2016)95. Within digital mental health and conversational system research, psychological support is commonly conceptualized in terms of users’ perceived understanding, comfort, and emotion-regulation resources during interaction (Tremain et al., 2020)96. Building on these foundations, the present study contextually adapted the scale items to everyday C-AI interaction scenarios to assess participants’ perceived psychological support arising from C-AI engagement. Consistent with the sampling strategy, the PHS items were not intended to classify participants according to whether they had used C-AI specifically for mental health purposes. Instead, they were designed to assess whether participants with general prior experience using C-AI perceived emotional relief, cognitive assistance, a sense of being understood, and coping-related resources in their prior C-AI interactions. Related research further suggests that evaluations of conversational agents for mental health and well-being should consider not only clinical effectiveness but also user experience, interaction quality, content engagement, and human-AI relational factors (Li et al., 2023)97. Accordingly, PHS was measured in this study as a subjective appraisal of psychological support arising from everyday C-AI interactions, rather than as an indicator of mental health-specific use, clinical treatment effects, or professional intervention outcomes. The scale consisted of 10 items rated on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). The instrument demonstrated strong internal consistency (Cronbach’s α = 0.958; M = 2.76, SD = 1.094). AI Autonomy (AIAu) The AI autonomy scale was adapted from Stepp’s anthropomorphic interaction framework (Steep, 2022)98, which conceptualizes autonomy as users’ perceptions of an intelligent system’s self-initiated responsiveness, interaction quality, and behavioral agency during human-AI interaction. The original scale was designed to assess social and anthropomorphic characteristics of intelligent agents, reflecting users’ experiences of social presence and affective engagement when interacting with AI systems. Subsequent studies have applied and refined this instrument within conversational voice assistant contexts to quantify perceived AI autonomy in interactive environments (Pal et al., 2022)99. In the present study, AI autonomy was measured using a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree) to capture users’ perceptions of autonomy during interactions with C-AI. The adapted scale consisted of three items. The instrument demonstrated strong internal consistency (Cronbach α = 0.818; mean = 3.978, SD = 0.65). AI Attachment (AIAt) The AI attachment scale was developed based on the intimate relationship attachment framework (Brennan et al., 1998)100 and extends attachment theory to the human–artificial intelligence interaction context, capturing the degree of emotional bonding that individuals develop through sustained interactions with C-AI (Yang and Oshio, 2025)101. The scale emphasizes users’ perceived psychological closeness, emotional reliance, and relational investment during ongoing interaction, serving as an important indicator of relational experience in human–AI engagement. In this study, AI attachment was measured using five items, all rated on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). The items focused on users’ emotional dependence on C-AI, perceived relational intimacy, and psychological involvement during continuous interaction. The scale demonstrated strong internal consistency (Cronbach α = 0.856; mean = 3.88, SD = 0.594). Data analysis For quantitative data analysis, this study employed partial least squares structural equation modeling (PLS-SEM) as the primary analytical approach, with model estimation conducted using Smart PLS V4.1. PLS-SEM is particularly suitable for research contexts involving multiple structural paths and moderating effects, especially when the objective is explanatory and predictive model estimation and strict distributional assumptions are not required. Given that the structural model in this study was developed based on the S-O-R framework and incorporated interaction terms, PLS-SEM was adopted to support robust estimation of the proposed path relationships. Model evaluation followed a two-stage procedure. First, the measurement model was assessed by examining indicator outer loadings, Cronbach’s alpha, composite reliability (CR), and average variance extracted (AVE) to evaluate internal consistency reliability and convergent validity. Discriminant validity was assessed using the Fornell-Larcker criterion, which compares the square root of each construct’s AVE with its correlations with other constructs. Second, the structural model was evaluated by reporting path coefficients and conducting bootstrapping with 5,000 resamples to obtain standard errors and confidence intervals. In addition, R² and Q² values were reported to assess explanatory power and predictive relevance. The moderating effects of AI autonomy and AI attachment were tested by constructing interaction terms in Smart PLS and examining their significance using the same bootstrapping procedure. To reduce the risk of common method bias, procedural remedies such as anonymous responses and randomized item presentation were implemented. At the statistical level, full collinearity variance inflation factors (VIFs) were examined as a supplementary assessment of common method bias. IBM SPSS Statistics version 29 was used for descriptive statistics, correlation analysis, and multicollinearity diagnostics to support the interpretation of the structural model results. All tests were two-tailed, with the significance level set at p < 0.05. Qualitative data in this study were collected through asynchronous online text-based question-and-answer exchanges, allowing participants to engage in reflective and considered responses outside of real-time interaction contexts. Qualitative interviews were conducted in a semi-structured, asynchronous text-based question-and-answer format to clarify the subjective experiences underlying the relationships observed in the quantitative model, including C-AI perceived trust, perceived warmth, AI autonomy, AI attachment, technology healthy use, and perceived psychological support. The interview questions were developed based on the stimulus-organism-response framework and the quantitative model, and were refined through research team discussion. They primarily covered C-AI use experience, C-AI perceived trust, perceived warmth, AI autonomy, AI attachment, technology healthy use, and subjective appraisals of psychological support. The interview guide focused on three main areas: first, how participants developed trust, perceived warmth, and perceived understanding during everyday C-AI interactions; second, how AI-initiated suggestions, topic elaboration, and sustained dialogue influenced participants’ judgment, reliance, and boundary management; and third, how participants interpreted certain C-AI responses as emotional relief, cognitive assistance, perceived understanding, or coping resources. To reduce the risk of leading responses, the interview questions were open-ended. The interview design avoided presupposing that C-AI was an effective support tool and encouraged participants to describe supportive, ineffective, and risk-related experiences. Representative questions included: Please describe the situations in which you usually use C-AI and what role these interactions typically play for you. In your everyday use of C-AI, what types of responses strengthen or weaken your trust in it? Please describe an interaction in which a C-AI response made you feel understood, comforted, or helped you clarify your thoughts. When C-AI actively provides suggestions, elaborates on a topic, or guides the conversation, how do you usually respond? Does this type of proactive response influence your judgment or how you use C-AI? When C-AI provides advice or comfort, how do you decide whether to verify the information, limit your reliance, or seek real-world support? Because this study used asynchronous text-based question-and-answer interviews rather than real-time audio or face-to-face interviews, interview duration was not applicable as a continuous measure of data collection time. Following COREQ and SRQR recommendations for transparent reporting of qualitative data collection and analysis, this study used the volume of participants’ written responses as a supplementary indicator of data adequacy in the asynchronous interview context (Tong et al., 2007; O’Brien et al., 2014). The corpus size was calculated based only on participants’ open-ended responses, excluding interview prompts, section headings, participant identifiers, and researcher instructions. The final open-ended response corpus included in the analysis comprised approximately 67,000 Chinese characters. Before formal coding, the research team assessed the suitability of the interview texts for qualitative analysis. The assessment criteria considered whether each text addressed the core open-ended questions, contained specific descriptions of C-AI interaction experiences, and provided substantive statements suitable for meaning-unit extraction. Relevance was assessed in relation to the key themes of this study, including C-AI trust, perceived warmth, AI autonomy, AI attachment, technology healthy use, and subjective appraisals of psychological support. Following qualitative content analysis guidance on data preparation and the selection of units of analysis, open-ended responses containing identifiable meaning units were retained for formal coding (Elo and Kyngäs, 2008). Two researchers independently conducted the initial assessment and resolved discrepancies through discussion. Texts meeting these criteria were included in formal coding. In addition, because qualitative sample adequacy depends not only on sample size but also on the specificity of the research aim, the information power of the sample, and the quality of textual data, the analysis considered whether the responses provided sufficient information to contextualize the core pathways in the quantitative model (Malterud et al., 2016). Participants for the qualitative phase were recruited sequentially from the pool of valid survey respondents. In line with prior qualitative interview research in which recruitment was stopped when subsequent interviews no longer generated new themes (Kornder et al., 2026), and drawing on methodological research on thematic saturation and sample size determination (Hennink and Kaiser, 2022), the research team continuously evaluated thematic repetition and the stability of meaning units during sequential recruitment and preliminary analysis. In addition, research on online semi-structured interviews suggests that 30 to 70 interviews are a common planned sample range, and that code saturation often occurs toward the later stages of planned interviews (Squire et al., 2024). Accordingly, during recruitment, the research team continuously examined whether newly collected interview texts produced new meaning units, subthemes, or thematic categories. Recruitment was stopped after 50 interviews. Comparison of the later-stage interviews, including the 50th interview, showed that these texts mainly repeated, supplemented, or refined the core viewpoints, experiential patterns, and meaning constructions identified in earlier interviews, without generating substantively new meaning units, subthemes, or thematic categories. This indicated that the qualitative data had reached thematic sufficiency for the explanatory aims of the qualitative phase. Because qualitative sample adequacy depends not only on sample size but also on the specificity of the research aim, the information power of the sample, and the quality of the data (Malterud et al., 2016), the research team determined that the existing data were sufficient to support the interpretation of the quantitative structural relationships. After the interview texts were screened for suitability and thematic sufficiency was confirmed, the research team analyzed the qualitative data using Colaizzi’s phenomenological analytic procedure. Following Colaizzi’s descriptive phenomenological approach, the research team sequentially conducted data familiarization, extraction of significant statements, formulation of meanings, theme clustering, and distillation of the essential structure of the phenomenon. This procedure helped ensure that the analysis remained grounded in the original data and analytically traceable (Morrow et al., 2015). Specifically, participants’ written responses were retained in their original wording and were anonymized, numbered, and organized for analysis. The researchers then repeatedly read the full textual corpus to become familiar with participants’ accounts of C-AI interaction experiences, technology healthy use, and subjective appraisals of psychological support. Significant statements directly relevant to the research questions were extracted from the original texts and organized into meaning units for coding. The research team then derived more abstract, formulated meanings from these meaning units and grouped them into theme clusters and higher-level thematic categories according to conceptual similarity. Two researchers independently conducted initial coding, meaning formulation, and theme clustering. Coding outcomes were refined through iterative comparison and discussion until consensus was reached. Throughout the analytical process, the research team maintained an audit trail linking original text segments, significant statements, meaning units, formulated meanings, theme clusters, and thematic categories to enhance transparency and traceability. To ensure interpretive consistency with the original context, the final thematic framework was repeatedly checked against the raw textual responses. Peer debriefing within the research team was also conducted to strengthen analytical credibility. Statement This study confirms that all methods were conducted in compliance with the relevant ethical guidelines and regulations.All procedures were reviewed and approved by the Academic Ethics Committee of Xiamen University of Technology prior to data collection (Approval No.: XUT-AEC-2025-05; Date: May 1, 2025). Adult participants provided electronic informed consent after being informed of the study’s purpose, procedures, potential risks, and benefits. Participation was voluntary, and participants were free to withdraw at any time without penalty. All data were anonymized using participant identification codes. Results Demographic description analysis The demographic characteristics of the participants are summarized in Table 3 (N = 1006). The gender distribution was relatively balanced: 56.3% (566/1006) were female and 43.7% (440/1006) were male. Participants were primarily aged 23–30 years, with 43.2% (435/1006) aged 27–30 and 38.1% (383/1006) aged 23–26; the 18–22 and 31–35 age groups accounted for 8.2% (82/1006) and 10.5% (106/1006), respectively. Most participants reported at least one year of C-AI use (86.9%, 874/1006), including 2 years (34.1%, 343/1006) and 1 year (33.4%, 336/1006); 19.4% (195/1006) reported ≥ 3 years of use, whereas 13.1% (132/1006) reported < 1 year. In terms of use frequency, 79.5% (800/1006) used C-AI at least a few times per week, including a few times per week (46.8%, 471/1006) and a few times per day (32.7%, 329/1006). Less frequent use was reported by 13.3% (134/1006) who used C-AI once per week and 7.2% (72/1006) who used it less than once per week. Overall, the sample primarily comprised young adults with sustained and frequent C-AI use. Table 3Sociodemographic characteristics of our research sample.VariablesOptionsN (%)(N = 1006)GenderMale440(43.7%)Female566(56.3%)Age18–2282(8.2%)23–26383(38.1%)27–30435(43.2%)31–35106(10.5%)Duration of UseLess than one year132(13.1%)One year336(33.4%)Two years343(34.1%)Three years or more195(19.4%)Frequency of UseA few times a day329(32.7%)A few times a week471(46.8%)Once a week134(13.3%)Less than once a week72(7.2%) Sociodemographic characteristics of our research sample. Following the demographic description of the survey sample, 50 young adults were purposively selected from the pool of valid survey respondents to participate in semi-structured, asynchronous, text-based in-depth interviews. The demographic and C-AI use characteristics of the interview participants are summarized in Table 1. The interview subsample included 23 male participants (46.0%) and 27 female participants (54.0%), and most participants were aged between 23 and 30 years (39/50, 78.0%). In terms of education, 28 participants held a bachelor’s degree (56.0%), while undergraduates and postgraduates each accounted for 22.0%. All interviewees reported at least one year of C-AI use, with most reporting two years or more of use (44/50, 88.0%). To protect participant privacy, all interview data were anonymized and de-identified prior to analysis; no names, contact details, institutional identifiers, or other potentially identifiable information were retained in the analytic dataset or reported materials. Detailed participant-level demographic information, presented only in anonymized form, is provided in Supplementary Information 3. Quantitative results of structural equation models Validity & common method bias test The measurement model showed acceptable internal consistency reliability and convergent validity. Following established PLS-SEM reporting recommendations, internal consistency reliability was assessed using Cronbach’s alpha and composite reliability (CR), whereas convergent validity was evaluated using indicator outer loadings and average variance extracted (AVE) (Hair et al., 2012)102. The indicator outer loadings ranged from 0.653 to 0.872. Cronbach’s alpha values ranged from 0.788 to 0.958, and CR values ranged from 0.863 to 0.963, all exceeding the recommended threshold of 0.70. The AVE values ranged from 0.615 to 0.733, all above the recommended threshold of 0.50, indicating acceptable convergent validity. Although THU1 had an outer loading of 0.653, it was retained because PLS-SEM guidelines recommend that reflective indicators with loadings between 0.40 and 0.70 should be evaluated based on their effects on CR, AVE, and content validity rather than removed automatically (Hair et al., 2017)103. Because the THU construct met the CR and AVE criteria and THU1 contributed to the content coverage of self-regulation and boundary management, the item was retained. Additionally, the square root of the AVE for each construct was higher than its correlations with other constructs, satisfying the Fornell-Larcker criterion for discriminant validity. Furthermore, indicator-level variance inflation factor (VIF) values were calculated to assess potential collinearity among measurement indicators. The VIF values ranged from 1.000 to 3.294, all below the threshold of 3.3, indicating no substantial collinearity among the indicators. Specific details are provided in Tables 2 and 4. Table 4Fornell–Larcker criterion for discriminant validity.AIAtAIAtAIAuC-AIPTC-AIPWPHSTHU0.791AIAu0.7380.856C-AIPT0.7410.6960.81C-AIPW0.7560.6830.7740.828PHS0.0660.0550.1690.1170.851THU0.6150.5960.6770.620.2770.784 Fornell–Larcker criterion for discriminant validity. Model fit The overall performance of the PLS-SEM model was further evaluated using the standardized root mean square residual (SRMR) and the normed fit index (NFI). The SRMR value was 0.061, which is below the recommended threshold of 0.08, indicating acceptable model fit. According to (Demler et al., (2015)104; Hair et al., 2021)105, the NFI typically ranges from 0 to 1, with higher values indicating better model fit; the NFI value in this study was 0.887, falling within the acceptable range. To further assess the model’s out-of-sample predictive capability, the PLS-Predict procedure was conducted to evaluate predictive validity, while comparing prediction errors across the PLS-SEM model, a linear benchmark model (linear model, LM). Following established recommendations, the analysis first examined the Q²_predict values for each indicator and subsequently compared RMSE and MAE metrics to determine the relative predictive advantage of the model. The results showed that all psychological health support items yielded positive Q²_predict values, ranging from 0.018 to 0.037, suggesting stable yet moderate out-of-sample predictive relevance for this construct. The Q²_predict values for conversational AI perceived warmth ranged from 0.396 to 0.429, while those for technology healthy use ranged from 0.153 to 0.364, indicating a gradient pattern of predictive strength from affective experience to behavioral tendency. In terms of prediction error comparison, among the 10 psychological health support indicators, nine items demonstrated lower RMSE values in the PLS-SEM model compared with the linear benchmark model, with only PHS7 slightly exceeding the LM. Overall, the positive Q²_predict values indicate predictive relevance, while the RMSE comparison suggests modest and construct-specific predictive performance rather than uniformly superior prediction over the linear benchmark model. ( specific details are provided in Tables 5 and 6). Table 5Structural equation modeling (SEM) statistics.SRMRSaturated ModelEstimated Model0.050.061d_ULS1.1511.726d_G0.380.423Chi-square value2386.2222576.881NFI0.8950.887 Structural equation modeling (SEM) statistics. Table 6Assessing the predictive power of models.PHS1Q²_predictPLS-SEM_RMSEPLS-SEM_MAELM_RMSELM_MAE0.0191.2161.0121.221.012PHS40.0271.2121.0131.2161.013PHS30.0181.1890.9911.1950.996PHS100.0371.1860.991.1940.997PHS90.021.2121.0151.221.021PHS20.0261.2251.0241.2331.038PHS50.0291.2351.041.2361.038PHS60.0321.2261.0281.231.034PHS70.0311.2371.0511.2341.044PHS80.0231.1940.9941.2061.004C-AIPW20.4290.7660.5410.7330.514C-AIPW30.3960.7650.5490.7230.512C-AIPW40.410.7660.5490.7380.526C-AIPW10.40.750.5430.7120.5THU20.3490.8080.6050.8050.593THU30.3640.7850.5970.7870.588THU40.3250.8290.6380.8290.633THU10.1531.1530.9351.1430.924 Assessing the predictive power of models. Hypothesis testing The hypothesis testing was conducted using PLS-SEM V4.1, constructing an extended S-O-R model based on anthropomorphism theory, which includes both mediation and chain mediation effects, to validate the hypothesis’s effectiveness. PLS-SEM Bootstrapping with 5000 resamples was used to estimate the path coefficients and significance levels of the relationships between the variables. Table 6 presents the path coefficients, t values, p values, and confidence intervals for the direct effects, while Fig. 2 illustrates the structural relationships among the variables. C-AI perceived trust had a significant positive effect on C-AI perceived warmth (β = 0.774, p < 0.001, t = 46.803, 95% CI [0.741–0.806]), supporting H1a. C-AI perceived trust also showed a significant positive effect on technology healthy use (β = 0.409, p < 0.001, t = 9.54, 95% CI [0.325–0.492]), supporting H1b. C-AI perceived warmth significantly influenced technology healthy use (β = 0.168, p < 0.001, t = 4.118, 95% CI [0.090–0.251]), supporting H2a. However, C-AI perceived warmth did not have a significant effect on psychological health support (β = 0.011, p = 0.837, t = 0.205, 95% CI [− 0.132–0.087]), and therefore H2b was not supported. Technology healthy use demonstrated a significant positive effect on psychological health support (β = 0.382, p < 0.001, t = 8.872, 95% CI [0.286–0.456]), supporting H3a. Result of structural model. The mediating and serial mediating effects were examined using the PLS-SEM bootstrapping procedure (Table 7). Before interpreting the specific indirect effects, the total effect of C-AIPT on PHS was inspected and found to be significant (beta = 0.215, p < 0.001, 95% CI [0.071, 0.286]) (see Table 8), indicating an overall positive association explained through the modeled indirect pathways. Table 7Specific indirect effect results.PathβSEtSignificance95% CIFindingC-AIPT -> C-AIPW -> PHS0.0090.0430.206p = 0.837[−0.102, 0.068]H2c Not SupportedC-AIPT -> THU -> PHS0.1560.0256.173p < 0.001[0.105, 0.205]H3b SupportedC-AIPT -> C-AIPW -> THU -> PHS0.0500.0133.714p < 0.001[0.024, 0.077]H3c SupportedC-AIPW -> THU -> PHS0.0640.0173.774p < 0.001[0.032, 0.099]C-AIPT -> C-AIPW -> THU0.1300.0324.059p < 0.001[0.069, 0.195] Specific indirect effect results. Table 8Direct and total effect results.PathβSEtSignificance95% CIFindingC-AIPT -> C-AIPW0.7740.01746.803p < 0.001[0.741, 0.806]H1a SupportedC-AIPT -> THU0.4090.0439.540p < 0.001[0.325, 0.492]H1b SupportedC-AIPW -> THU0.1680.0414.118p < 0.001[0.090, 0.251]H2a SupportedC-AIPW -> PHS0.0110.0550.205p = 0.837[-0.132, 0.087]H2b Not SupportedTHU -> PHS0.3820.0438.872p < 0.001[0.286, 0.456]H3a SupportedAIAu -> THU0.1810.0394.64p < 0.001[0.108, 0.26]AIAt -> PHS-0.2080.063.470p < 0.01[-0.266, -0.015]Total effectC-AIPT -> PHS0.2150.0543.999p < 0.001[0.071, 0.286] Direct and total effect results. Total effect C-AIPT -> PHS Regarding specific indirect effects, the indirect path through C-AIPW alone (C-AIPT -> C-AIPW -> PHS) was not significant (beta = 0.009, SE = 0.043, t = 0.206, p = 0.837, 95% CI [-0.102, 0.068]), indicating that perceived warmth did not function as a significant single mediator; thus, H2c was not supported. In contrast, the indirect effect through THU (C-AIPT -> THU -> PHS) was significant (beta = 0.156, SE = 0.025, t = 6.173, p < 0.001, 95% CI [0.105, 0.205]), supporting H3b. Furthermore, the serial mediation pathway (C-AIPT -> C-AIPW -> THU -> PHS) showed a significant chain indirect effect (beta = 0.050, SE = 0.013, t = 3.714, p < 0.001, 95% CI [0.024, 0.077]), supporting H3c. In addition, the indirect effect of C-AIPW on PHS through THU was significant (beta = 0.064, SE = 0.017, t = 3.774, p < 0.001, 95% CI [0.032, 0.099]), suggesting that perceived warmth was associated with psychological health support primarily through technology healthy use rather than through a significant direct pathway. Table 9 shows that the interaction term between conversational AI perceived trust and AI autonomy was significantly and negatively associated with technology healthy use (β = -0.028, p = 0.035, 95% CI [-0.055, -0.002]), indicating that AI autonomy negatively moderated the relationship between conversational AI perceived trust and technology healthy use. Specifically, the positive association between conversational AI perceived trust and technology healthy use became weaker at higher levels of AI autonomy, thereby supporting H4. In addition, the interaction term between AI attachment and technology healthy use was significantly and negatively associated with perceived psychological health support (β = -0.065, p = 0.009, 95% CI [-0.105, -0.006]), suggesting that AI attachment negatively moderated the relationship between technology healthy use and perceived psychological health support. In other words, higher AI attachment weakened the positive association between technology healthy use and perceived psychological health support, supporting H5. Further simple slope analyses based on the mean ± 1 standard deviation of AI autonomy and AI attachment are shown in Figs. 3 and 4. Table 9Moderating effect results.Moderated pathRelationshipβSEtSignificance95% CIFindingAIAu moderates C-AIPT -> THUAIAu x C-AIPT -> THU−0.0280.0132.104p = 0.035[−0.055,−0.002]H4 SupportedAIAt moderates THU -> PHSAIAt x THU -> PHS−0.0650.0252.602p = 0.009[−0.105,−0.006]H5 Supported Moderating effect results. C-AIPTxAIAu-THU. THUxAIAt-PHS. Qualitative results During the interviews, most participants demonstrated a clear understanding of the functional role and limitations of conversational AI (C-AI) as an assistive tool. They rationally framed it as a temporary resource for emotional support and cognitive regulation rather than a substitute for interpersonal relationships, reflecting their capacity for self-regulated, technology healthy use. Based on Colaizzi’s phenomenological analysis procedure, Fig. 5 presents the thematic structure and illustrative subthemes; representative quotations are provided below, and the full coding framework is reported in Supplementary Information 4. Thematic Mapping Based on Respondents’ Narratives. Theme 1. Emotional Support Function. At the level of emotional expression and release, participants generally perceived C-AI as a low-risk and nonjudgmental space for emotional disclosure. Anonymity and privacy reduced concerns about social evaluation, making it easier for them to express negative emotions and prevent emotional suppression. Some interviewees emphasized that this mode of interaction provided a sense of being heard and acknowledged, thereby offering immediate emotional buffering. However, a small number of participants noted that such support tended to remain at a basic level, with relatively limited emotional depth. “When I feel anxious, I turn to C-AI for support… it makes me feel like there is a place to vent my emotions, which helps me quickly calm down.” [Male, 29 years old]. “Sometimes, my thoughts become a mess… although it cannot completely solve my problems, its answers provide emotional relief.” [Female, 30 years old]. At the level of emotional comfort and anxiety reduction, participants consistently mentioned that concise and stable comforting feedback helped to reduce emotional intensity. Empathetic yet non-intrusive responses were seen as an immediate psychological buffer, keeping the interaction controllable and low-pressure. Several participants believed that these standardized responses helped to organize their thoughts and restore emotional stability. However, a few perspectives pointed out that this comfort primarily served as short-term regulation rather than providing sustained, deep support. “It always provides me with concise comforting responses without offering intrusive advice… I no longer feel so burdened.” [Male, 27 years old]. “Although I know that C-AI cannot fully understand my emotional dilemma, its simple feedback gives me enough comfort.” [Female, 28 years old]. At the level of privacy protection and emotional safety, participants closely associated emotional support with the anonymity and low-interference nature of the interaction environment. The independence afforded by the technological setting enhanced feelings of self-acceptance and reduced anxiety related to being judged or misunderstood. Most participants viewed C-AI as a stable yet clearly bounded listening channel that helped maintain a sense of security during emotional expression. However, a minority of participants noted that this sense of safety stemmed more from the technological structure itself rather than from relational emotional attachment. “Although C-AI is not a real human, it doesn’t judge, just listens to me, and I feel a hint of warmth in its feedback.” [Male, 29 years old]. “Sometimes I feel extremely helpless… but C-AI’s responses help me calm down through its calm tone.” [Female, 27 years old]. Theme 2. Cognitive Support Function. At the level of cognitive restructuring and emotional reframing, participants generally perceived C-AI as a supportive tool that facilitates emotional organization and cognitive reconstruction. Through guided interactions, they were able to reinterpret sources of stress, attenuate negative thinking patterns, and gradually enhance their ability to recognize and articulate their emotional states. Several participants reported that the problem-oriented responses of C-AI helped shift their focus from uncontrollable future-related anxiety toward actionable present tasks, thereby clarifying both emotional and cognitive structures. In addition, some participants described the interaction process as a form of ongoing cognitive practice that encouraged more rational reflection on personal challenges. However, a minority noted that such support was primarily reflected in structured guidance and reframing rather than in deeper or more complex judgment. “It helps me sort out emotions through questions and see pressure from different perspectives.” [Female, 24 years old]. “Through these questions, I began to re-examine the source of my emotions.” [Male, 29 years old]. At the level of problem structuring and goal-oriented thinking, participants further described C-AI as a cognitive aid that helped transform vague stress into clearly articulated problems. Through contextual decomposition, priority differentiation, and goal setting, they gradually identified the sources of occupational pressure and developed more coherent decision pathways. Several participants noted that structured questioning encouraged a shift from emotional reactions toward task-focused thinking, enabling previously disorganized stress to be translated into actionable steps and concrete objectives. At the same time, a few participants emphasized that this form of support functioned more as a cognitive framework or thinking tool rather than as a source of direct solutions. “It reorganized my fragmented thoughts and turned vague anxiety into manageable steps.” [Male, 24 years old]. “Structured questioning helped me define problems first and then find executable goals.” [Female, 24 years old]. Theme 3. Behavioral Regulation and Healthy Use Function. At the level of shaping healthy interaction behaviors, participants generally viewed C-AI as an auxiliary tool that supports daily behavioral regulation and emotional management training. Through breathing relaxation guidance, adjustments to work rhythm, and periodic self-monitoring, interactions gradually shifted from immediate emotional relief toward sustained self-management practices. Several participants noted that C-AI’s reminders and structured suggestions enabled rapid state adjustment under high-pressure conditions and helped integrate emotional regulation into everyday work routines, thereby fostering more stable behavioral rhythms. In addition, some participants emphasized that this form of support extended beyond emotional comfort to include actionable regulation strategies and rhythm planning. However, a small number of participants suggested that such guidance primarily functioned as basic behavioral assistance rather than a mechanism for long-term behavioral change. “It guided me to track emotions daily and use self-monitoring to manage stress instead of reacting impulsively.” [Male, 20 years old]. “It reminds me to regulate my pace during overtime and maintain a healthier rhythm at work.” [Male, 33 years old]. At the level of goal-oriented task support, participants further described C-AI as a cognitive-behavioral assistive tool that facilitates problem-solving and communication strategy optimization. Through guided questioning and situational analysis, they were able to overcome creative bottlenecks, clarify task objectives, and develop clearer action paths in complex work scenarios. Several participants mentioned that C-AI helped them reassess problems from multiple angles and reduce uncertainty in high-pressure tasks through structured communication suggestions, making behavioral decisions more rational and executable. However, a few participants felt that this support was more akin to an auxiliary thinking framework rather than a direct substitute for professional judgment. “It guided me to analyze the core goal of the design and rethink communication strategies.” [Female, 23 years old]. “It helps me translate complex technical ideas into clearer expressions for clients.” [Male, 29 years old]. Theme 4. Relationship Safety Function. At the level of digital support alliance and relational safety, participants generally regarded C-AI as an interactive channel that provides stable companionship and continuous, accessible support. Because the interaction is unrestricted by time and free from interpersonal evaluation pressure, this non-interpersonal environment gradually forms functional trust and relational safety, allowing participants to maintain emotional stability in a low-burden state. Several participants mentioned that the consistently predictable response patterns enhanced the predictability of the interaction process and strengthened the experience of “stable companionship,” rather than traditional emotional attachment. Additionally, some participants described this relationship as a low-pressure support alliance, with the core being stability and retrievability, rather than deep emotional understanding. However, this relationship was more understood as functional safety support rather than a real interpersonal connection. “It feels like a stable companion… a low-burden alliance that builds functional trust over time.” [Male, 28 years old]. “I treat it as a functional support object… a continuous sense of relational security rather than an emotional substitute.” [Female, 26 years old]. At the level of functional boundaries and rational dependence, participants generally exhibited a clear understanding of C-AI’s limitations. While receiving emotional support and cognitive organization, they actively maintained the priority of real-life interpersonal relationships and emphasized avoiding excessive attachment. Several participants pointed out that C-AI’s understanding ability remains limited in complex emotional and career pressure situations, and thus it is more often used as an auxiliary tool rather than a primary source of support. This rational awareness of functional boundaries leads to a relational support system characterized by controllability, limitations, and a focus on real social networks. “I clearly understand it cannot replace human relationships… I still rely on friends and family for deeper support.“[Female, 34 years old]. “My reliance on C-AI is rational… I use it first to organize thoughts, but still seek real people for complex issues.” [Male, 29 years old]. Discussion Grounded in an extended stimulus–organism–response (S–O–R) framework and incorporating anthropomorphism-related boundary conditions, this mixed-methods study examined cross-sectional associations among conversational AI perceived trust (C-AIPT), conversational AI perceived warmth (C-AIPW), technology healthy use (THU), AI autonomy (AIAu), AI attachment (AIAt), and perceived psychological health support (PHS) in everyday C-AI use. Given the cross-sectional design, the findings do not support causal or mechanistic claims about C-AI producing psychological health support; rather, they should be interpreted as statistical associations supplemented by qualitative contextualization. The quantitative results showed that C-AIPT was positively associated with both C-AIPW and THU. In the structural model, PHS was significantly associated with THU, whereas the direct path from C-AIPW to PHS was nonsignificant. In other words, C-AIPW is better understood as an affective interactional cue associated with engagement and self-regulated use, rather than as an independent correlate of PHS. Indirect-effect analyses further indicated that the association between C-AIPT and PHS was primarily represented by the indirect association through THU and by the serial indirect association through C-AIPW and THU. In addition, AIAu attenuated the positive association between C-AIPT and THU, and AIAt attenuated the positive association between THU and PHS, suggesting that autonomy- and attachment-related perceptions conditioned the strength of these associations. The qualitative interviews further contextualized these quantitative findings: participants tended to describe warm C-AI responses as lowering expressive pressure and increasing willingness to interact, whereas they more often linked perceived psychological support to THU-related experiences, including reflective judgment, goal-oriented use, information verification, and boundary management. Taken together, the findings indicate that PHS in everyday C-AI interaction was more consistently associated with goal-oriented, self-regulated, and boundary-aware THU than with perceived warmth alone. C-AIPT and Its associations with C-AIPW and THU Existing research suggests that, in intelligent human-AI interaction contexts, users’ trust in artificial intelligence is an important antecedent of affective evaluations and functional perceptions (Afroogh et al., 2024)106. The present results showed that C-AIPT was positively associated with C-AIPW, supporting H1a. This pattern can be interpreted in light of Bullock et al. (2025)107, who suggest that trust may be associated with lower perceived risk and ethical burden, stronger social attribution, and reduced cognitive monitoring demands, which may help explain why users with higher trust tend to interpret C-AI responses as supportive and warm interactional feedback. In immersive and high-frequency conversational environments, C-AIPT may also reflect a sense of perceived relational safety. This interpretation is consistent with Xu et al. (2025)108 and McKee et al. (2024)109, who suggest that higher levels of trust are generally associated with lower defensive processing, greater willingness to self-disclose, and higher perceived interactional warmth. C-AIPT was also positively associated with THU, supporting H1b. This finding is consistent with Visser et al. (2025)110, who suggest that stable trust in human-AI interaction may be related to calibrated engagement, the management of task goals, and lower attentional fragmentation. It also aligns with Choudhury and Shamszare’s (2023)111 discussion of trust in relation to decision efficiency and task-management quality. In the present study, users with higher C-AIPT were more likely to report goal-oriented, self-regulated, and boundary-aware THU, including selective adoption of suggestions, active adjustment of interaction pacing, and use of AI as a supportive knowledge resource rather than as a source of dependency (Abdelwahed et al., 2025)112. Wu et al. (2024)113 and Suzuki et al. (2026)114 similarly suggest that stable trust does not necessarily correspond to greater dependence or higher use frequency, but may instead be associated with information verification, cautious judgment, and moderate use. Taken together, these findings suggest that C-AIPT is relevant not only to users’ acceptance of C-AI but also to goal-oriented, self-regulated, and boundary-aware THU. This pattern helps explain why the association between C-AIPT and PHS was primarily captured by the indirect association through THU, rather than by the separate indirect association through C-AIPW. C-AIPW as an interactional cue and THU as the primary association pathway This study conceptualized users’ responses to C-AI use at two levels: affective interactional appraisal, represented by C-AIPW, and self-regulated use orientation, represented by THU. The model results showed that C-AIPW was positively associated with THU, supporting H2a. This finding is consistent with prior research suggesting that when users perceive benevolent, supportive, and low-threat responses during C-AI interaction, they are more likely to report stable engagement and more structured patterns of use (Jiang et al., 2022)115. In this sense, C-AIPW should not be equated directly with psychological support itself but is better understood as an affective interactional cue. Sustained and benevolent human-AI interaction may be associated with clearer conversational pacing, lower expressive pressure, and more stable willingness to engage, thereby linking C-AIPW to goal-oriented and self-regulated THU (Lee and Yong, 2024)116. Furthermore, the present study found that the direct path from C-AIPW to PHS was nonsignificant; therefore, H2b was not supported. Similarly, the separate indirect association between C-AIPT and PHS through C-AIPW was nonsignificant, and H2c was not supported. This differentiated pattern highlights the need to distinguish zero-order associations between variables from independent path estimates in the structural model after THU and other related constructs are included. C-AIPW did not show a significant direct structural path to PHS, nor was it observed as a separate mediator between C-AIPT and PHS; rather, its role was mainly evident in its association with THU. Related research likewise suggests that affective interactional experiences, perceived closeness, or a sense of safety do not necessarily correspond to stable appraisals of psychological support. Herbener and Damholdt (2025)117 showed that affective perceptions in human-AI interaction may enhance users’ perceived closeness and safety with C-AI, but do not necessarily correspond to more stable psychological support appraisals. Prior discussions in digital mental health research on the distinction between experiential evaluations and outcome-related evaluations also suggest that feeling understood, accepted, or comforted does not mean that such experiences are independently associated with PHS (McKee et al., 2023)118. In the absence of self-regulation, information verification, and behavioral engagement, stronger anthropomorphic expectations or warmth experiences may also be related to relational expectation bias rather than to stable support appraisal (Hong et al., 2023)119. Sawyer et al. (2023)120 further noted that the supportive value of affective experiences may be limited when they are not accompanied by behavioral engagement or skill learning. In line with Tremain et al. (2020)121 and Sanjeewa et al. (2024)122, who emphasized interaction quality and user experience in digital mental health contexts, the role of C-AIPW is better understood as lowering the threshold for expression, strengthening interactional safety, and supporting subsequent self-regulated use, rather than as an independent direct correlate of PHS. By contrast, THU was positively associated with PHS, supporting H3a. The indirect-effect results further showed that the indirect association between C-AIPT and PHS through THU was significant, supporting H3b. The serial indirect association between C-AIPT and PHS through C-AIPW and THU was also significant, supporting H3c. Taken together, these findings indicate that THU is the key construct linking C-AIPT, C-AIPW, and PHS in the present model. Unlike C-AIPW, which represents an immediate affective interactional appraisal, THU more directly reflects users’ goal orientation, reflective judgment, information verification, pacing regulation, and boundary management during C-AI use. These features are more closely aligned with the conceptual meaning of PHS, which in this study does not refer to clinical mental health improvement, but to users’ perceived emotional buffering, cognitive clarification, sense of being understood, and coping-related resources in everyday C-AI interaction. Therefore, the association between C-AIPT and PHS was not primarily represented by perceived warmth alone, but was more clearly observed in the indirect association through THU. The role of C-AIPW also appears mainly in its connection with THU. Several strands of related research support this interpretation. Schillings et al. (2024)123 suggested that sustained and structured interaction quality is more likely to be related to users’ appraisal of support value than occasional affective feedback alone. Allen et al. (2026)124 further indicated that when users engage in problem focusing, meaning organization, and strategic responding during interaction, technology-mediated interaction is more likely to be interpreted as a supportive resource. Gratch and Essig’s (2025)125 analysis of the relationship between trust and interaction regulation also helps explain the indirect pattern observed in the present study: as a trust-related cognitive appraisal, C-AIPT may be associated with more regular and boundary-aware use patterns, which are further linked to users’ subjective appraisal of PHS. Research on emotion regulation (Kettle and Lee, 2024)126, human-AI relational experience (Skjuve et al., 2021)127, and perceived support resources (Volpato et al., 2025)128 further suggests, from different perspectives, that the supportive value of C-AI is better understood in relation to users’ sustained engagement, cognitive processing, and boundary awareness, rather than being attributed solely to warmth perception or relational experience itself. Accordingly, THU should not be understood as general use frequency or continued interaction, but as a self-regulated use orientation more closely aligned with PHS. This interpretation helps explain why PHS was more consistently associated with goal-oriented, reflective, and boundary-aware use than with C-AIPW as a direct correlate. Moderating roles of AIAu and AIAt in key association paths Building on the finding that THU was the key construct linking C-AIPT, C-AIPW, and PHS, the moderation results further clarified how AIAu and AIAt conditioned two key associations in the model. Specifically, AIAu negatively moderated the positive association between C-AIPT and THU, supporting H4. This result indicates that when users perceived higher AI autonomy, the positive association between C-AIPT and THU became weaker. In other words, AIAu should not be understood as a technical feature that inherently supports healthy use, but rather as a system-side moderator that conditions how trust is associated with self-regulated use. Prior research suggests that effective trust in human-AI interaction should be balanced with user control, preserved human judgment, and trust calibration (Park et al., 2024)129. When C-AI presents stronger autonomous suggestions, proactive guidance, or anthropomorphic decision cues, users’ trust may be more likely to be expressed as delegated reliance or default acceptance of system outputs, rather than as reflective judgment, information verification, and boundary management. Wrightson-Hester et al. (2025)130 also suggest that higher autonomy and anthropomorphic interaction features may complicate users’ perceptions of control boundaries, responsibility allocation, and interactional control. From this perspective, the negative moderating role of AIAu in the C-AIPT-THU association does not indicate that AI autonomy directly weakens healthy use; rather, it suggests that under high-autonomy interaction conditions, the link between trust and goal-oriented, self-regulated, and boundary-aware THU may become weaker. The present study also found that AIAt negatively moderated the positive association between THU and PHS, supporting H5. This result indicates that when users reported higher attachment to C-AI, the positive association between THU and PHS became weaker. Unlike AIAu, which operates as a system-side autonomy cue, AIAt is closer to a user-side relational investment, reflecting users’ emotional connection, companionship expectations, and psychological reliance formed through sustained C-AI interaction. Békés and Aafjes-van Doorn (2026)131 noted that when users develop stronger emotional connections during interaction with AI systems, their interaction motivation may no longer be organized solely around task completion or problem solving, but may increasingly involve companionship seeking, emotional reassurance, and relationship maintenance. Zhou and Hu (2026)132 further showed that higher levels of AI attachment may change how users interpret the value of human-AI interaction, making them more likely to evaluate C-AI interaction from the perspective of emotional companionship and relational continuity rather than solely from functional support or goal attainment. In light of the present findings, higher AIAt may be associated with users interpreting psychological support less in terms of THU-related goal orientation, self-regulation, and boundary management, and more in relation to companionship expectations and relationship-maintenance orientation. Accordingly, AIAt should not be understood as a factor that directly reduces PHS, but as a user-side moderator that conditions the strength of the association between THU and PHS. Taken together, AIAu and AIAt operate at two key stages of the model: AIAu moderates the association between C-AIPT and THU, whereas AIAt moderates the association between THU and PHS. Together, these findings suggest that anthropomorphic and relational factors do not necessarily strengthen supportive associations in C-AI use. Instead, when system autonomy is higher or users’ relational attachment is stronger, the THU-centered association path may be attenuated. This finding further reinforces the central argument of this study: PHS in everyday C-AI use is more consistently associated with goal-oriented, self-regulated, and boundary-aware THU, rather than being more strongly associated with higher AI autonomy or stronger AI attachment. In other words, the perceived support value of C-AI is better understood in relation to clear human-AI boundaries, appropriate user control, and reflective use orientation. Qualitative contextualization of the quantitative association patterns The qualitative interviews further contextualized the quantitative findings. Participants’ accounts suggested that users did not simply equate warm C-AI responses with psychological support. Instead, they more often understood supportive experiences in terms of low-pressure expression, cognitive clarification, reflective use, and relational boundary awareness. The analysis identified four experiential themes related to perceived support: affective conditions for expression, cognitive clarification, behavioral self-regulation, and relational boundary safety. These themes help explain the core pattern in the quantitative model: C-AIPW is better understood as an affective interactional cue associated with THU, whereas THU is more consistently associated with PHS. First, affective conditions for expression were mainly reflected in a low-pressure, relatively private, and less judgmental interaction environment. Participants often described warm C-AI responses as lowering expressive pressure, increasing willingness to disclose, and making interaction feel more approachable, rather than as psychological support itself. Harris-Watson et al. (2023)133 similarly suggested, in their work on digital therapeutic alliance, that emotional resonance in conversational agents often functions as an entry point for user engagement and self-expression rather than as an independent support outcome. Relatedly, cognitive clarification was reflected in how users drew on C-AI prompts, follow-up questions, and structured responses to reframe unclear stressors, emotional concerns, or complex situations. Participants did not usually perceive support simply because C-AI responses were “warm”; rather, they were more likely to interpret the interaction as supportive when it helped them identify sources of distress, organize their concerns, and develop a clearer interpretive frame. In this sense, affective conditions for expression and cognitive clarification together suggest that the role of C-AIPW lies less in a direct association with PHS than in lowering the threshold for expression, strengthening interactional safety, and supporting later self-regulated use. Second, behavioral self-regulation was the theme most closely aligned with THU in the qualitative data. Participants more often linked supportive experiences in C-AI interaction to planned use routines, targeted questioning, selective adoption of suggestions, moderate reliance, and active verification, rather than to emotional comfort alone. Wang et al. (2021)134 similarly indicated that behavioral engagement and self-regulation are important dimensions for understanding technology-mediated support experiences in digital mental health contexts. In line with the quantitative findings of this study, this theme suggests that THU should not be understood as general use frequency or continued interaction, but as an orientation toward reflective judgment, control over use, and boundary awareness during C-AI interaction. This also helps explain why THU was more consistently associated with PHS than C-AIPW: in this study, PHS primarily refers to users’ subjective appraisal of emotional buffering, cognitive clarification, perceived understanding, and coping-related resources, and these appraisals are more closely linked to self-regulated patterns of use. Finally, relational boundary safety was reflected in users’ tendency to understand C-AI as an assistive tool rather than as a primary emotional substitute. Participants’ accounts suggested that when users maintained clearer functional boundaries in human-AI interaction, C-AI use was more often organized around rational help-seeking, situational analysis, and assisted decision-making. By contrast, when users developed stronger companionship expectations toward C-AI, interaction was more likely to shift toward relationship maintenance or emotional reassurance. Bryant et al. (2026)135 similarly suggested that blurred emotional boundaries in human-AI relationships may increase short-term engagement and emotional involvement, but do not necessarily correspond to more stable support appraisals. This theme is consistent with the quantitative finding for H5, namely that higher AIAt attenuated the positive association between THU and PHS. In other words, the qualitative data further suggest that PHS is more consistently associated with boundary-aware, self-regulated, and reflective THU than with warmth perception or relational attachment itself. Conclusions This mixed-methods study suggests that psychological health support reported by young adults in C-AI interactions is more likely to be realized through a trust-driven behavioral pathway—namely, perceived trust fostering healthier, goal-oriented and self-regulated use (THU), which is in turn associated with psychological health support. Perceived warmth appears to function primarily as an interactional cue that facilitates engagement and supports the formation of healthier usage patterns, rather than directly predicting psychological support outcomes. Extending the S–O–R framework, the study also identifies boundary conditions linked to anthropomorphism: higher perceived AI autonomy attenuates the trust–THU association, and stronger AI attachment weakens the positive association between THU and psychological health support. Practically, these findings imply that C-AI systems designed for mental health–related support should not rely solely on human-like emotional expression. Instead, design and governance should prioritize trust formation grounded in transparency and controllability, promote goal-oriented and self-regulated use patterns, and mitigate attachment-related risks through boundary prompts, usage feedback, and risk identification mechanisms. Future research may further examine whether user guidance on when to use C-AI, how to use it, when to stop, and when to seek offline support is associated with more boundary-aware C-AI engagement. Overall, the study provides evidence-based insight into how interaction experiences translate into perceived psychological support in everyday C-AI use among young adults, while highlighting the need to balance engagement benefits with boundary-related risks. Limitations and future research The findings should be interpreted in light of several limitations. First, although the model was grounded in the extended S–O–R framework, the cross-sectional design precludes causal and temporal inference. Accordingly, the mediation and moderation results should be understood as statistical indirect and conditional associations, not as evidence of causal mechanisms or temporally ordered psychological processes. The findings cannot determine whether C-AIPT precedes THU, whether THU predicts subsequent PHS, or whether users with higher perceived support retrospectively evaluate their C-AI use more positively. Future studies should use longitudinal, experimental, or cross-lagged designs to examine the temporal ordering and stability of these associations. Second, several important confounding variables were not fully controlled. Baseline mental health status, loneliness, personality traits, attachment orientation, prior help-seeking experience, prior use of digital mental health tools, and familiarity with C-AI may influence both C-AI use patterns and appraisals of PHS. Thus, the observed associations may partly reflect pre-existing individual differences rather than C-AI-specific interaction patterns. Future research should include these variables as controls, moderators, or grouping factors to provide a more robust test of the proposed association model. Third, this study relied primarily on online questionnaires and self-reported scales. Although procedural controls, collinearity checks, and qualitative interviews helped reduce methodological concerns and strengthen interpretation, single-source self-reported data may still be affected by cognitive consistency, recall, and social desirability biases. In particular, THU was measured as a self-reported regulatory use orientation rather than through directly observed behavioral records. Future research should integrate interaction logs, behavioral trace data, or platform usage data to examine whether self-reported THU corresponds to actual human–AI interaction patterns. Fourth, an important measurement limitation concerns AIAt. Although the AIAt items were adapted from existing attachment-related frameworks, item adaptation does not constitute formal construct validation, particularly because AI attachment may be conceptually distinct from traditional human attachment. Without a fully validated AI-specific attachment measure, the AIAt scale may not have fully distinguished attachment to AI from adjacent constructs such as trust, liking, perceived warmth, companionship motivation, or frequency of use. Future research should employ or compare validated AI-specific attachment measures, such as the AI Attachment Scale developed and validated by Kasturiratna and Hartanto (2025)136, and further examine the discriminant validity of AIAt in relation to related human–AI interaction constructs. Finally, the sample and context also limit generalizability. Rural participants, users with lower digital literacy, and groups with limited access to C-AI tools were underrepresented. Platform-level differences in algorithmic feedback, language style, anthropomorphic cues, and interaction mechanisms may also shape trust, warmth, attachment, and healthy use. Future research should compare different user groups and C-AI platforms to test whether the proposed association patterns remain stable across contexts. Although the mixed-methods design enhanced interpretive depth, the quantitative–qualitative integration remained primarily explanatory and complementary; future studies could adopt more integrated designs informed by quantitative path results. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Supplementary Material 1 Author contributions L-YL and L-SH designed the research study. L-YL performed the research. L-SH analyzed the data. L-SH and H-C reviewed and edited the manuscript. L-YL and C-L supervised the study. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work. Funding This research received no specific grant from any funding agency in the public, commercial, or non-profit sectors. Data availability All data generated or analyzed during this study are included in this published article and its supplementary information files. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Sihang Liu and Chao Liu contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Supplementary Material 1 Data Availability Statement All data generated or analyzed during this study are included in this published article and its supplementary information files.