Turn to chatbots for sharing feelings or seeking solutions? Differential associations of preferences for chatbot-mediated self-disclosure and instrumental help-seeking with adolescent interpersonal behaviors in the United States and Hong Kong

Synthesis: Hu, Habibi Asgarabad, Jen and Cheung ask whether what adolescents talk to AI chatbots about matters more than how much they use them. From an open-ended item, they sort 2,023 adolescents (1,276 in the United States, 747 in Hong Kong, mean age 15.52) into an instrumental Help-Seeking orientation — asking for problem-solving guidance or factual information — or a self-disclosure orientation centered on feelings, personal experience and relationships, with a third group preferring to say nothing. Instrumental Help-Seeking behaved the same way in both places: it related to more adaptive coping (β = 0.16, p < 0.001) and, through it, to more prosocial behavior (β = 0.09) and less bullying (β = −0.02). Self-disclosure diverged. In Hong Kong it tracked higher adaptive coping, more prosocial behavior and less bullying; in the United States it tracked more maladaptive coping and, through that, more bullying (self-disclosure → maladaptive coping: US b = 0.16, HK b = −0.38, z = 4.84, p < 0.001). The takeaway for adults is that "chatbot use" is not one behavior: tool-like question asking looks benign to beneficial in both contexts, whereas emotional disclosure is where culture, platform design and expectations can change the sign of the association.

Key Findings

1. A screened cross-national sample of 2,023 adolescents. Qualtrics panels in the United States and Hong Kong were recruited through parents or guardians, with parental approval and informed consent plus attention and speed checks; 2,272 adolescents passed screening, 249 "others" responses (ambiguous or missing, e.g. "anything", "life") were dropped from the focal comparison, leaving 1,276 US (63.1%) and 747 Hong Kong (36.9%) participants, mean age 15.52 (SD = 1.24).
2. The two sites preferred different conversations. Overall, 51.3% preferred instrumental Help-Seeking, 31.2% preferred to say nothing and 17.5% preferred self-disclosure about feelings, stress and relationships. Hong Kong adolescents leaned harder toward engagement (62.0% instrumental, 30.0% self-disclosure, only 8.0% nothing), while US adolescents were far more likely to decline the premise (44.8% nothing, 45.0% instrumental, 10.2% self-disclosure).
3. Both engaged groups looked better than the disengaged group. One-way ANOVA on the whole sample put the self-disclosure group highest on adaptive coping (M = 2.86) then instrumental help-seeking (M = 2.70) then nothing (M = 2.53; F = 35.14, p < 0.001), the same ordering for prosocial behavior (3.74 / 3.52 / 3.42; F = 16.14, p < 0.001), and the self-disclosure group lowest on maladaptive coping (2.06 vs 2.29 and 2.26; F = 13.70, p < 0.001) and bullying (0.40 vs 0.51 and 0.55; F = 4.80, p < 0.01).
4. Instrumental help-seeking was culturally stable. In the pooled structural equation model (CFI = 0.957, TLI = 0.939, RMSEA = 0.048), instrumental help-seeking predicted adaptive coping (β = 0.16, p < 0.001) and nothing else directly; adaptive coping then predicted less bullying (β = −0.12) and more prosocial behavior (β = 0.55), yielding significant indirect effects of β = −0.02 (bullying) and β = 0.09 (prosocial), with the maladaptive-coping routes non-significant (β = 0.01, p > 0.05 for bullying; β = −0.00, p > 0.05 for prosocial behavior).
5. Self-disclosure was positive in the pooled model. It predicted more adaptive coping (β = 0.22, p < 0.001) and less maladaptive coping (β = −0.09, p < 0.01), with significant indirect effects on bullying through both coping styles (β = −0.03, p < 0.001 via adaptive; β = −0.03, p < 0.01 via maladaptive) and on prosocial behavior (β = 0.12, p < 0.001 via adaptive; β = 0.01, p < 0.05 via maladaptive).
6. The pooled result was carried by Hong Kong, and reversed in the United States. Self-disclosure related to adaptive coping in Hong Kong but not in the US (b = 0.42 vs 0.09; z = −3.74, p < 0.001), and to maladaptive coping positively in the US but negatively in Hong Kong (b = 0.16 vs −0.38; z = 4.84, p < 0.001). The adaptive-coping routes to prosocial behavior and to bullying were also significantly stronger in Hong Kong (z = −3.79, p < 0.001; z = 2.16, p < 0.05).
7. Region-specific indirect effects followed the same split. Self-disclosure's indirect effect on bullying through adaptive coping was significant in Hong Kong (β = −0.06, p < 0.001) and null in the US (β = −0.00, p > 0.05), and its effect on prosocial behavior was significant in Hong Kong (β = 0.25, p < 0.001; maladaptive route β = 0.02, p < 0.05) and not in the US (β = 0.02 and −0.01, both p > 0.05). Conversely, instrumental help-seeking's route to reduced bullying ran through adaptive coping in the US (β = −0.01, p < 0.05) but not in Hong Kong (β = −0.02, p > 0.05).
8. Invariance testing licensed the split, and the scales held up. Coping, bullying and prosocial measures all showed at least metric invariance across sites, but the fully constrained model fitted worse than the metric-invariant one (ΔCFI = 0.011 > 0.01), which the authors read as evidence against a single pooled estimate and for analyzing and comparing the two groups separately. The US sample was 66.6% female against 30.1% in Hong Kong, and only age and gender were controlled.

Samples and instruments in two sites

Participants were reached through Qualtrics, a commercial panel that contacts Parents and Families and asks them to invite their adolescent children to complete a self-report battery online; the authors cite the platform's use in prior work for reaching demographically diverse respondents. Ethical approval came from a university committee, consent was electronic and parental approval preceded participation, and data quality was guarded by attention and speed checks that removed respondents and, in this analysis, also removed the 249 "others" responses that did not fit the topic coding scheme.

The instrument set is deliberately conventional, chosen to operationalize a mediation model spanning Social-Emotional Learning, coping and peer behavior rather than to probe chatbot experience. Coping styles came from the Brief COPE, using 10 of its 14 strategies grouped into adaptive coping (instrumental support, emotional support, active coping, acceptance, positive reframing, planning) and maladaptive coping (self-blame, substance use, behavioral disengagement, denial), with Cronbach's α of 0.88 and 0.84 pooled (0.86/0.81 US; 0.91/0.89 Hong Kong). Interpersonal behavior was measured by the 9-item bullying-behaviors subscale of the Illinois Bullying Scale (α = 0.92 pooled; 0.91 US; 0.95 Hong Kong) and the 16-item Prosocial Behavior Scale (α = 0.94 pooled; 0.94 US; 0.95 Hong Kong). Multi-group invariance testing found metric invariance for coping (ΔCFI = 0.004, ΔRMSEA = 0.001) and bullying (ΔCFI = 0.001, ΔRMSEA = 0.003) and scalar invariance for the prosocial scale, so the cross-national comparison rests on scales that behave comparably in both samples.

Measuring self-disclosure preference and instrumental help-seeking preference

The predictor was a single open-ended Generative AI prompt — "If you have access to generative AI (e.g., ChatGPT, Microsoft Bing, and Google Bard), what topics would you like to discuss with generative AI?" — chosen to capture spontaneous preferences in adolescents' own words and to avoid priming from fixed options. Two researchers coded responses independently into four categories, resolving disagreement by discussion and assigning multi-topic answers to their predominant theme.

The self-disclosure scheme follows the Collins and Miller framework of personal experience, feelings, emotions, reflection and beliefs about oneself, with examples such as "how to deal with stress and social anxiety" and interpersonal concerns such as making friends. The instrumental help-seeking scheme follows work on utilitarian student–chatbot conversation and the typology of instrumental social support, covering Problem Solving guidance ("how to improve English") and factual or analytical information seeking. A "nothing to say" category covers "I don't know", refusal and "nothing", and everything else — ambiguity, missing data — became "others" and was excluded. Topic preference was then dummy-coded with "nothing to say" as the common reference group for both predictors, age and gender were controlled, mediation was estimated with 1,000 bootstrap samples and 95% confidence intervals, and multi-group models were compared through nested model testing.

The measurement boundary matters for reading the results: the study captured what adolescents say they would like to discuss, not whether or how often they actually used a Conversational AI chatbot, and no platform data were collected.

Results: coping, peers and interpersonal behavior

The mediator and outcome differences are clean in the pooled sample and messy in the US sample alone. Across all 2,023 respondents the ANOVA ordering was self-disclosure > instrumental help-seeking > nothing on adaptive coping and prosocial behavior, and self-disclosure < the other two on maladaptive coping and bullying, so engaged adolescents looked better adjusted on every measure of coping, peer behavior and Well-Being. Split by site, Hong Kong reproduced that pattern strongly (adaptive coping 3.00 vs 2.71 vs 2.57 under self-disclosure, instrumental, nothing, F = 23.63; maladaptive 1.88 vs 2.27 vs 2.27, F = 23.21; prosocial 3.86 vs 3.35 vs 3.16, F = 35.39), while in the US the only significant contrasts were between the instrumental and nothing groups, with maladaptive coping and bullying showing no group differences at all (F = 1.76 and 0.34).

The Large Language Models (LLMs) chatbot question sharpens in the path model. In the pooled sample, adaptive coping is the dominant mediator: it carries β = 0.55 to prosocial behavior and β = −0.12 to bullying, more than twice the apparatus around maladaptive coping (β = 0.34 to bullying, β = −0.10 to prosocial). Instrumental help-seeking reaches prosocial behavior almost entirely through that adaptive route, which is why its profile travels intact across cultures — an instrumental Help-Seeking orientation toward a tool behaves like help-seeking in any other setting. Self-disclosure, by contrast, is the only predictor with a reliable direct path to reduced maladaptive coping in the pooled model, and that path is precisely the one that changes sign between the two sites: Hong Kong adolescents who preferred to disclose to chatbots reported the lowest maladaptive coping of any group (M = 1.88), while US adolescents who preferred the same reported the highest (M = 2.39, non-significant in that subsample).

Interpretation: substitution or complementarity?

The authors read the whole-sample self-disclosure results as evidence that chatbot conversation can complement rather than substitute for human relationships. Their mechanisms are social cognitive theory, where behavior is acquired through repeated interaction with the environment, and the computers-are-social-actors paradigm, where people apply social heuristics — reciprocity, politeness — to machines that show social presence. On that account a helpful chatbot response can be generalized into real life, raising Self-Efficacy and the willingness to seek support from people, which is what the adaptive-coping route measures. They add a self-focus explanation: a disposition toward disclosure may reflect self-reflection, which clarifies goals and processes emotion, rather than rumination, which dwells on threat — and the pooled data look like reflection dominating.

The cross-national divergence is explained culturally, not clinically. East Asian animistic traditions are described as affording weaker human exceptionalism and greater comfort anthropomorphizing technology, so Hong Kong adolescents may experience a Conversational AI chatbot as genuinely socially present; Confucian emphasis on self-reflection as a route to growth, combined with a non-judgemental partner, may make the chatbot a safe workspace for that habit. Western human exceptionalism, by contrast, differentiates humans from the rest of the world, so US adolescents may find chatbot responses emotionally unresponsive in a way that feeds rumination, maladaptive coping and diminished Trust in the exchange. Practical recommendations follow the split: designers for Chinese users can position Conversational AI chatbots as non-judgemental empathic companions and build in journalling or guided self-inquiry prompts, designers for Western users should frame them as "thinking partners" rather than "empathic listeners", and mental-health and clinical guidance should proactively name the limits of emotional responsiveness for Western clients. Instrumental and cognitive support, being governed by functional criteria such as accuracy and speed that are not culturally variable, can be deployed broadly with little adaptation. The authors also flag platform affordances as an untested alternative explanation, since relationally oriented companion apps and search-integrated assistants are not equally popular in the two sites, and they state plainly that mechanisms invoking actual interaction quality remain tentative because only preferences were measured.

What this means for practice

• Educators. Teach the difference between asking a chatbot for a solution and confiding in it: instrumental Help-Seeking predicted adaptive coping (β = 0.16, p < 0.001) in both the United States and Hong Kong, so steer AI guidance toward problem-solving and factual requests and keep emotional disclosure a conversation with people.
• Educators. Name the limits of a chatbot's emotional responsiveness in US-facing programs instead of assuming it transfers: preference for self-disclosure tracked maladaptive coping in the United States (b = 0.16) and the opposite in Hong Kong (b = −0.38, z = 4.84, p < 0.001).
• Administrators. Keep counseling staff and peer-support routes visible alongside any chatbot deployment, because the lower bullying that followed instrumental help-seeking ran through adaptive coping (β = −0.02) — the same route human support seeking takes.
• Researchers. Measure what adolescents say they want to discuss to a chatbot, not how often they use one: the predictor here was a single open-ended preference item with no platform data, which is why the emotional-responsiveness mechanisms stay tentative.

Limitations

• The design is cross-sectional and correlational, so no temporal ordering or causal claim survives; the authors note that coping styles and behavior could equally shape conversation preference, and call for experimental or longitudinal work.
• Recruitment was an online panel limited to the United States and Hong Kong, exposed to self-selection and to whatever coverage Qualtrics achieves in each market, with school- or community-based sampling recommended for future studies; the two samples also differ in gender composition (US 66.6% female, Hong Kong 69.9% male) with only age and gender controlled.
• All measures were self-reported, inviting recall bias and social desirability, and bullying perpetration in particular is likely under-reported; the predictor was a single open-ended item about willingness, not behavior, so the survey never established whether or how frequently participants used a chatbot, and everything the paper says about chatbot responses, AI Sycophancy or emotional responsiveness is an interpretation of a disposition rather than of an interaction, with the cultural and self-focus mechanisms offered as tentative explanations.
• The coding forced each adolescent into one of three categories by predominant theme, erasing mixed motives, and no data were collected on which platforms adolescents used, so the authors explicitly warn that companion-style applications with stronger relational affordances may produce different patterns than the general-purpose assistants named in the prompt.

Connected Concepts

• Help-Seeking — the orientation whose stability across sites carries the paper's central contrast
• Well-Being — adolescent emotional adjustment framed through coping, disclosure and peer behavior
• Social-Emotional Learning — the developmental domain in which disclosure and prosocial behavior are being shaped
• Student-AI Interaction — conversation topic as a moderator of what chatbot contact does to learners
• Conversational AI — general-purpose assistants (ChatGPT, Bing, Bard) as the interaction partner studied
• Generative AI — the technology class adolescents were asked about in the open-ended item
• Large Language Models (LLMs) — the language-model substrate whose humanlike fluency drives the CASA effect
• Self-Efficacy — the mechanism claimed to carry instrumental help-seeking into adaptive coping
• Trust — anthropomorphism, emotional responsiveness and perceived social presence as trust questions
• Privacy — personal disclosure to a commercial system, unmeasured but central to the practice
• Cognitive Offloading — the counter-reading in which chatbots absorb coping work rather than support it
• Anxiety and Stress — adult concerns about adolescent chatbot use that the findings partly complicate

Connected Articles

• An AI-Powered Culturally Aware Chatbot for Stress Detection and Wellness Support among Pakistani University Students Using NLP and Machine Learning — cultural adaptation of stress-support chatbots for students
• Ask Me Anything: Exploring Children's Attitudes Toward an Age-tailored AI-powered Chatbot — children's attitudes toward and trust in chatbot interaction
• Understanding student perspectives on generative AI chatbots: a human-centred mixed-methods study in higher education — student perspectives on generative AI chatbots in learning
• Efficiency vs. Effectiveness: Self-Regulated Learning with LLM-Mediated Help-Seeking — help-seeking and self-regulation with LLM support
• Warning About AI Fallibility Increases Help-Seeking in an Intelligent Tutoring System — how fallibility signals change learners' help-seeking behavior
• Scaffolding Students-AI Dialogue: A Framework for Safe Educational Interactions — The SCAFFOLD framework for steering students-AI dialogue, with its classroom pilot
• Regulating the AI Tutor: Intentions, Help-Seeking, and Self-Regulated Learning in Adolescent GenAI Use — adolescent self-regulation with an AI tutor
• An Experimental Study Exploring Human–AI Complementarity in Early Social-Emotional Learning — complementarity between human and AI support in social-emotional learning

Citation

Hu, H., Habibi Asgarabad, M., Jen, E., & Cheung, H. N. (2026). Turn to chatbots for sharing feelings or seeking solutions? Differential associations of preferences for chatbot-mediated self-disclosure and instrumental help-seeking with adolescent interpersonal behaviors in the United States and Hong Kong. PsyArXiv preprint.