Problematic use of AI chatbots for mental health concerns among adolescent psychiatric outpatients: severity, impairment, and coping

Abstract Background Generative artificial intelligence (AI) chatbots are being increasingly adopted by adolescents to address mental health–related concerns, such as seeking emotional reassurance, clarifying symptoms, and obtaining coping guidance. In clinical settings, however, some youths may develop maladaptive patterns of use that become maladaptive and interfere with therapeutic processes. This study aimed to characterize problematic AI chatbot use among adolescents receiving child and adolescent psychiatry outpatient care and to examine associated psychosocial correlates. Methods We conducted a cross-sectional study in a Tunisian child and adolescent psychiatry outpatient clinic (July–December 2025). The participants were adolescents who reported using generative AI chatbots for mental health concerns. Clinically meaningful problematic use was identified via a clinician-based, impairment-centered classification aligned with the World Health Organization’s International Classification of Diseases, 11th Revision (ICD-11) principles for disorders due to addictive behaviours. Severity was assessed dimensionally using an adapted 20-item Internet Addiction Test for AI chatbots (AIAT-20). A concise 10-item short form (Concise internet Addiction Test for AI chatbots; CIAT-10) was derived from the AIAT-20 item pool and tested in the study sample for internal consistency and criterion-related discrimination against the clinician-based classification. Coping strategies were measured with the Brief Coping Orientation to Problems Experienced Inventory (Brief COPE), loneliness with the University of California, Los Angeles Loneliness Scale (UCLA Loneliness Scale; Version 3), and family functioning with the Family Satisfaction Scale (FSS) from the Family Adaptability and Cohesion Evaluation Scales, Fourth Edition (FACES IV) package. Discrimination of severity scores against the clinician classification was examined via receiver operating characteristic (ROC) analyses, and correlates of CIAT-10 severity were examined via correlational and multivariable regression analyses with robust standard errors. Results Ninety-two adolescents were included (mean age 13.2 ± 1.1 years; 60.9% female). The most frequently endorsed psychiatric-related purposes were emotional support during distress (71.7%) and seeking information about mental health conditions (70.7%). On the basis of the clinician-based impairment classification, 21 participants met the criteria for clinically meaningful problematic use (22.8%; 95% confidence interval [CI] 15.4–32.4). The internal consistency was high for the AIAT-20 (Cronbach’s alpha [α] = 0.888) and acceptable to good for CIAT-10 (α = 0.794). The CIAT-10 showed good discrimination against the clinician classification (area under the ROC curve  = 0.789) and high sensitivity at the empirical screening threshold (≥ 24). Greater CIAT-10 severity was independently associated with lower trust but higher perceived usefulness of chatbot responses, along with lower family satisfaction and weaker problem-focused coping, suggesting that problematic reliance may reflect perceived regulatory value rather than simple confidence in AI-generated advice. Conclusions Among adolescent psychiatric outpatients, problematic AI chatbot use may reflect a digitally mediated strategy for managing distress, reassurance seeking, and unmet coping or relational needs rather than excessive exposure alone. Larger-scale studies are needed to clarify unresolved patterns, particularly the apparent dissociation between trust and perceived usefulness, while clinical assessment should focus on function, impairment, and therapeutic impact.