Effect of a Cognitive Behavioral Therapy–Based AI Chatbot on Depression and Loneliness in Chinese University Students: Randomized Controlled Trial With Financial Stress Moderation

The study investigates whether a culturally adapted, CBT-based AI chatbot can improve mental health outcomes (depression, anxiety, and loneliness) among Chinese university students and whether financial stress moderates its effectiveness. Context: University students have high rates of depression, anxiety, and loneliness, exacerbated postpandemic by financial stress. Despite need, service use is low in non-Western settings like China due to stigma, access, cost, and cultural barriers. Digital interventions and AI chatbots can overcome barriers via accessible, anonymous, and personalized support, but engagement and cultural relevance remain challenges. Cultural adaptation of CBT is essential in Chinese contexts given collectivistic norms, face concerns, emotional restraint, and somatic expressions, and prior evidence shows adapted CBT yields better outcomes and adherence in Chinese populations. Research gaps include limited RCTs in Chinese students, lack of co-production approaches, and little evidence on moderation by financial stress. Hypotheses: (1) Students using the culturally adapted chatbot will show greater improvements in depression, anxiety, and loneliness versus waitlist control. (2) Effectiveness will be moderated by baseline financial stress, with greater benefits for students with higher financial stress.

The paper reviews evidence that digital mental health interventions can reduce depression, anxiety, and stress in students and that AI chatbots offer 24/7 access, reduced stigma, and personalization, with early trials showing efficacy (eg, Woebot). It emphasizes cultural adaptation of CBT as necessary in Chinese contexts due to collectivistic values, face-related stigma, emotional restraint, and somatization. Prior systematic reviews and trials demonstrate that culturally adapted CBT improves outcomes and adherence among Chinese populations, though effectiveness varies by adaptation approach. The review also notes low mental health service utilization in non-Western student populations, particularly China, and highlights financial stress as a growing determinant of student mental health. Gaps identified include limited RCTs of culturally adapted AI chatbots for Chinese students and insufficient analysis of financial stress as a moderator of intervention effects.

Design: Parallel-group, randomized controlled trial with a 2 (group: intervention vs waitlist control) × 3 (time: baseline day 1 [T1], day 3 [T2], day 7 [T3]) mixed design. Participants: Recruited from multiple public universities via email and flyers. Inclusion: full-time undergraduate/graduate students aged 18-24, Android smartphone ownership, willingness to engage daily for 7 days. Exclusion: current psychotherapy, history of severe mental illness (eg, schizophrenia, bipolar disorder), current psychiatric medication use. Screening via online questionnaire and follow-up interviews; written informed consent obtained. Ethics: Approved by Nanjing Normal University IRB (NNU202301018); registered at ChiCTR (ChiCTR2500100868). Sample size: A priori power analysis (G*Power 3.1) for mixed ANOVA detecting small effect (f=0.2), 80% power, α=.05 indicated N=42; accounting for attrition, target set to N=100. Randomization: 1:1 allocation (n=50 intervention, n=50 control) using computer-generated block randomization by independent statistician; allocation concealed via digital system post-baseline. Intervention: A 7-day culturally adapted CBT-based AI chatbot (“Psy-Bot”) delivered via Android app. Content grounded in Beck’s CBT model and included cognitive restructuring, behavioral activation, problem-solving, mindfulness, and social skills training. Cultural adaptation followed Bernal’s 8-dimension framework (language, persons, metaphors, content, concepts, goals, methods, context), incorporating indirect communication, collectivistic and face-related themes, and culturally resonant metaphors; expert ratings indicated strong cultural relevance (mean 4.1/5). Module structure (approx 20-30 min/day): Day 1 CBT introduction; Day 2 coping with campus-related stressors; Day 3 adapting to new environments; Day 4 managing anxiety (including pandemic-related anxiety); Day 5 financial stress management; Day 6 emotion regulation and mood improvement; Day 7 resilience and relapse prevention. Technical architecture: Hybrid response generation using RASA-based dialog management; BERT-based NLU fine-tuned on Chinese therapeutic conversations; responses combined pre-scripted (70%), template-based (20%), and limited contextual generation (10%); safety monitoring with automated risk detection; data logged in secure Firebase. Validation: (1) Content validation by three experts (CBT adherence mean 4.2/5; cultural relevance 4.1/5; engagement 3.9/5). (2) Technical validation and stress-testing. (3) Pilot test with 10 students: System Usability Scale mean 82.3 (SD 3.83) indicating excellent usability; Mobile Application Rating Scale indicated good comprehension (mean 4.4/5). Delivery: One module unlocked per day with reminders; conversational interface with multimedia; gamification for engagement. Control: Waitlist control received daily check-in messages and completed assessments at T1, T2, T3; access to app after T3. Measures: - Depression: CES-D-10 (0-30; higher=worse); Cronbach α=.89, .89, .91 at T1-T3. - Anxiety: GAD-7 (0-21); α=.92, .93, .94. - Loneliness: UCLA Loneliness Scale v3 (20-80); α=.93, .94, .95. - Financial stress: Psychological Inventory of Financial Scarcity (PIFS; 12 items, 1-7 Likert; higher=greater stress). Analytic plan: Data screened for outliers (z>±3.29) and assumption violations; missing data handled via last observation carried forward (LOCF) consistent with intention-to-treat; attrition 11%. Baseline differences tested via independent t tests and χ² tests. Primary outcomes analyzed via mixed-design ANOVAs (group × time) per measure, with Bonferroni-corrected pairwise tests; effect sizes reported as partial eta-squared and Cohen d for change scores. Moderation by financial stress examined using mixed linear models with group, time, financial stress, and interactions as fixed effects and subject as random effect; exploratory subgroup analyses used median split on PIFS (high vs low) with repeated-measures ANOVAs within intervention subgroup.

- Sample: N=100 (mean age 20.8 years, SD 2.24; 62% female). Attrition 11% overall (intervention 6%; control 16%). No significant baseline group differences in demographics or outcomes. - Primary effects (mixed ANOVAs): Significant group×time interactions for depression (CES-D: F(2,196)=8.63; P<.001; ηp²=.08) and loneliness (UCLA: F(2,196)=5.57; P=.004; ηp²=.05), indicating greater improvements in the intervention group versus waitlist. Anxiety showed no significant interaction (GAD-7: F(2,196)=1.31; P=.27; ηp²=.01). - Post hoc within-group changes (intervention): Significant reductions from T1 to T3 in depression (t=3.85; P<.001; Cohen d=0.71, 95% CI 0.30-1.12) and loneliness (t=4.28; P<.001; Cohen d=0.60, 95% CI 0.20-1.00); no significant changes in waitlist control. - Change-score effect sizes (between groups T1→T3): CES-D Cohen d=0.71 (95% CI ~0.30-1.17); UCLA Cohen d=0.60 (95% CI 0.20-1.00), reflecting moderate-to-large effects favoring the intervention. - Moderation by financial stress (mixed models): For CES-D, significant main effects of time (F(2,192)=6.96; P=.001) and financial stress (F(1,96)=12.96; P<.001), significant Group×Time (F(2,192)=7.92; P<.001) and Time×Financial stress (F(2,192)=5.90; P=.003) interactions, and a marginal Group×Time×Financial stress interaction (F(2,192)=2.86; P=.06). For loneliness, time (F(2,192)=10.54; P<.001), financial stress (F(1,96)=23.08; P<.001), and Group×Time (F(2,192)=5.19; P=.006) were significant; the 3-way interaction was not (P=.101). - Exploratory subgroup analyses (intervention group): High financial stress (n=27) showed large time effects: CES-D F(2,52)=11.56; P<.001; ηp²=.31; UCLA F(2,52)=11.18; P<.001; ηp²=.30. Low financial stress (n=23) showed no significant effect for CES-D (F(2,44)=1.22; P=.31; ηp²=.05) and a smaller, marginal effect for UCLA (F(2,44)=3.44; P=.04; ηp²=.14). - Safety/engagement: High adherence in intervention (47/50 completed all 7 sessions).

Findings support the primary hypothesis that a culturally adapted, CBT-based AI chatbot can reduce depressive symptoms and loneliness among Chinese university students relative to a waitlist control, with moderate-to-large effects. The null effect on anxiety suggests differential responsiveness across symptom domains. Potential reasons include a brief 7-day duration, emphasis on depression- and loneliness-relevant CBT components (cognitive restructuring, behavioral activation, social skills) versus anxiety-specific techniques, the physiological and avoidance components of anxiety that may require more intensive or longer interventions, and cultural norms around anxiety. Moderation analyses indicate that financial stress likely shapes treatment response: students with higher financial stress benefitted more, consistent with the chatbot’s targeted modules for financial worries, its accessibility and anonymity reducing cost and stigma barriers, and emerging evidence that socioeconomic factors influence intervention effectiveness. The results underscore the value of culturally adapted, AI-delivered CBT in expanding access within university stepped-care systems, particularly for financially stressed students in non-Western contexts, and illustrate how cultural adaptation (eg, collectivistic values, face concerns) can enhance acceptability and outcomes.

This RCT demonstrates that a culturally adapted, CBT-based AI chatbot can effectively reduce depression and loneliness among Chinese university students, with especially strong benefits for those experiencing high financial stress. The work advances scalable, culturally sensitive digital mental health solutions for underserved student populations. Future research should: (1) extend intervention duration and include follow-ups to assess sustainability; (2) incorporate and test anxiety-specific components (eg, exposure, safety behavior reduction, somatic regulation); (3) use active control conditions to isolate specific therapeutic effects; (4) evaluate implementation in real-world university settings without monetary incentives; (5) include multimethod assessments (clinician ratings, behavioral/physiological markers); and (6) further refine cultural adaptations and explore mechanisms by which financial stress influences treatment response.

- Sample size, while adequate for primary effects, limited power for moderation and subgroup analyses; exploratory findings require confirmation in larger, pre-registered moderation designs. - Exclusion of participants in psychotherapy, on psychiatric medications, or with severe mental illness limits generalizability to more clinically complex students. - Waitlist control may inflate between-group differences due to expectancy/attention effects; active controls are needed. - Reliance on self-report measures may introduce reporting and cultural biases (eg, somatization). - Brief 7-day duration may be insufficient to affect anxiety or ensure durable effects; no follow-up assessments were conducted. - High adherence may have been influenced by monetary compensation; real-world adherence could be lower. - Technical constraints: predominantly rule-/template-based responses may limit personalization; further NLP advancements could improve responsiveness while ensuring safety. - Cultural adaptation, though systematic, may not have captured all relevant nuances; broader stakeholder involvement could enhance relevance.