University students face numerous stressors that can impact their mental health and well-being. Depression, anxiety, and loneliness are prevalent concerns in this population [1], with global estimates indicating that over 30% of tertiary students experience clinically significant symptoms [2]. Compared to primary and secondary school students, university students are at a higher risk of developing mental health issues due to the unique challenges they face, such as increased academic pressure, financial burden, and the transition to adulthood [3]. Existing studies consistently illustrate that anxiety and depression lead to a variety of negative outcomes among university students, such as poor academic performance, reduced social networks, low quality of life, and sleep problems [4,5].

In the postpandemic era, financial stress has significantly worsened the psychological health of university students in various countries, including the United States, United Kingdom, Germany, and China [6-8]. The rising costs of higher education, coupled with limited financial resources, place a significant burden on many students, leading to increased vulnerability to mental health problems [9]. Research consistently suggests that financial hardship is associated with higher levels of stress, anxiety, and depression among university students [10,11]. This relationship appears particularly pronounced in the postpandemic context, where economic uncertainties have intensified financial pressures on students [8]. Addressing the impact of financial stress on university students’ mental health is, therefore, increasingly crucial.

Despite the high prevalence of mental health issues, most university students seldom seek professional mental health services provided by their institutions or external providers [4,12]. A systematic review and meta-analysis study reported that 35% (95% CI 22%‐50%) of university students used mental health services, with individual study estimates ranging from 13.7% to 68.6% [13]. This substantial heterogeneity reflects differences in health care systems, cultural contexts, service definitions, and accessibility of services across countries and institutions. Notably, the same review also examined specific subgroups and service types, finding that overall outpatient use rates were particularly low among students in non-Western contexts, such as China (5.1%) and Bangladesh (7.1%), compared to the average rates in Western settings [13]. This East-West disparity highlights the particular challenges faced by students in Asian contexts.

The underuse of mental health services can be attributed to various barriers, such as stigma, time constraints, unaffordability, and limited accessibility [14-16]. These barriers are often more pronounced in collectivistic cultures like China, where mental health stigma is particularly strong and seeking help may be viewed as bringing shame to one’s family [17]. The combination of high need and low service use—especially pronounced in Chinese contexts—underscores the importance of developing innovative, culturally sensitive interventions that can overcome these substantial barriers to care.

Digital mental health interventions, such as mobile apps and web-based self-help programs, show promise in overcoming treatment and access barriers [18]. These interventions can meet the high demand for campus counseling services and reduce access barriers by providing convenient, anonymous, and cost-effective support to students. A meta-analysis by Harrer et al [19] found that internet-based interventions were effective in reducing symptoms of depression, anxiety, and stress among university students, with effect sizes being comparable to those of face-to-face interventions. However, many digital interventions still face challenges with user engagement, personalization, and cultural relevance [20-22].

Recent advancements in artificial intelligence (AI) and natural language processing have paved the way for the development of AI-powered chatbots that can deliver psychological support through interactive conversations [23]. These AI-powered chatbots offer several advantages over traditional digital interventions: they provide 24/7 access to support, reduce stigma associated with help-seeking, and offer personalized, engaging experiences [24]. Emerging research supports their efficacy; for example, Fitzpatrick et al [25] found that a CBT-based conversational agent effectively reduced symptoms of depression and anxiety among young adults, while D’Alfonso [26] reported promising results using an AI chatbot for university students experiencing financial difficulties.

AI-powered chatbots can be particularly beneficial for university students, who are digital natives and may prefer seeking help through technology-based interventions [18]. The COVID-19 pandemic has further highlighted the importance of digital mental health interventions, as social distancing measures have limited access to in-person mental health services [27]. The pandemic has led to widespread social isolation and disruption of daily routines, which may exacerbate the risk of mental health problems among students [28]. Chatbots can provide a convenient and accessible source of support for students who may have smaller social networks or feel hesitant to share their problems with others due to the pandemic. Furthermore, the anonymous nature of chatbot interactions may encourage students who are hesitant to share their problems with parents, friends, or teachers to seek help [29]. Chatbot interventions can provide personalized support through interactive conversations, which may include psychoeducation, reflection prompts, in-session exercises, and personalized feedback based on users’ responses [25]. These interactive features can enhance user engagement, promote the acquisition and application of coping skills, and provide a sense of social support, which may be particularly valuable for students experiencing social isolation or loneliness [30].

Cognitive-behavioral therapy (CBT) has demonstrated effectiveness for treating common mental health problems among university students [31]. However, cultural context significantly influences how mental health problems are experienced, expressed, and addressed, making cultural adaptation of psychological interventions not merely beneficial but essential for maximizing effectiveness.

The need for cultural adaptation is particularly clear in Chinese contexts, where mental health experiences differ markedly from Western norms [17]. In Chinese culture, mental health problems are often viewed through a collectivistic lens where individual psychological issues can threaten family reputation (“face”), creating a “courtesy stigma” that extends beyond the individual to their entire family unit, which can intensify stigma and discourage help-seeking [32]. This family-oriented stigma creates stronger barriers to help-seeking compared to more individualistic Western societies. In addition, traditional Chinese values emphasize emotional restraint and self-reliance and the importance of maintaining social harmony, which may conflict with Western therapeutic approaches that encourage emotional expression and seeking external help [33]. Mental health issues in Chinese contexts are often conceptualized in somatic terms rather than psychological ones, with cultural idioms of distress frequently manifesting as physical complaints [34].

These cultural differences are not merely theoretical but have demonstrated practical implications for intervention effectiveness. Several studies have found that standard Western CBT interventions may have lower engagement and effectiveness when applied directly to Chinese populations without cultural adaptation [35,36]. For example, Hwang et al [36] demonstrated that culturally adapted CBT for Chinese Americans showed not only significantly better clinical outcomes but also substantially lower dropout rates compared to standard CBT. Similarly, Choi et al [37] found that cultural adaptation of CBT significantly improved treatment adherence and outcomes among Chinese Australians.

A recent meta-analysis by Li et al [38] examining culturally adapted interventions for common mental disorders in people of Chinese descent found these adaptations beneficial, though the efficacy did not differ significantly between culturally modified interventions and culturally specific interventions. Rathod et al [39] also reviewed culturally adapted psychological interventions and found varying levels of effectiveness across different cultural groups and adaptation approaches. These highlight that thoughtful adaptation, whether modifying existing approaches or developing culturally specific ones, can improve outcomes when grounded in an understanding of core cultural concepts and values while maintaining essential therapeutic mechanisms.

Despite the promising potential of AI-powered chatbots and the importance of cultural adaptation, significant research gaps remain. First, most studies evaluating digital mental health interventions have been conducted with Western samples, limiting generalizability to Chinese university students [40,41]. Second, as revealed by a systematic review of digital mental health in China [42], Chinese studies rarely adopt co-production approaches in developing digital health technologies and often demonstrate less rigorous methodological quality. Third, few studies have examined whether digital mental health interventions are equally effective across different levels of financial stress, which is an increasingly important consideration given growing economic pressures on students.

Most critically, despite the theoretical promise of combining AI chatbot technology with culturally adapted CBT content, very few randomized controlled trials have evaluated such interventions specifically for Chinese university students experiencing financial stress. This represents a significant missed opportunity to address an urgent public health need through innovative, scalable approaches that could potentially overcome cultural and structural barriers to care.

This study addresses these gaps by evaluating the efficacy of a custom-developed, culturally adapted, CBT-based AI chatbot in reducing depression, anxiety, and loneliness among Chinese university students, with particular attention to the moderating role of financial stress. Using a randomized controlled trial design, we tested two primary hypotheses:

By examining both the overall efficacy of the chatbot and the moderating effect of financial stress, this study aims to provide valuable insights into how AI-driven, culturally adapted interventions might be optimally deployed to improve the well-being of vulnerable student populations.

This study used a parallel-group, randomized controlled trial design. Participants were recruited from multiple public universities through email advertisements and campus flyers. The inclusion criteria were (1) being a currently enrolled full-time undergraduate or graduate student aged 18‐24 years, (2) owning an Android smartphone, and (3) being willing to engage with the chatbot daily for 7 days. The exclusion criteria included (1) current engagement in any type of psychotherapy, (2) having a diagnosis of any severe mental illness in history (eg, schizophrenia, bipolar disorder, etc), and (3) current use of psychiatric medications. Initial screening was conducted via online questionnaire (Wenjuanxing), and eligible participants were invited for further screening interviews conducted by trained research staff to confirm eligibility and obtain informed consent.

The study received ethical approval from the Institutional Review Board of Nanjing Normal University (NNU202301018). All participants provided written informed consent before enrollment. The consent process explicitly detailed the study purpose, procedures, potential risks and benefits, data handling protocols, and participants’ right to withdraw at any time without penalty. Data were anonymized using participant identification codes, and all information was stored on encrypted, password-protected servers compliant with institutional data security standards. Participants who completed the study received monetary compensation for their time and contribution. The study was registered with the Chinese Clinical Trial Registry (ChiCTR2500100868).

A priori power analysis using G*Power 3.1 (Heinrich-Heine-Universität Düsseldorf) was conducted to determine the necessary sample size for detecting a small effect size (Cohen f=0.2) with 80% power at a significance level of α=.05. The anticipated effect size was based on meta-analytic findings by Firth et al [43], who reported effect sizes for smartphone-based mental health interventions targeting depression ranging from g=0.22 to g=0.56, depending on control condition type. We selected f=0.2 as a conservative estimate appropriate for our brief 7-day intervention with waitlist control, representing a clinically meaningful effect size within the empirically derived range.

The analysis was based on a 2 (group: intervention vs control) × 3 (time: T1, T2, T3) mixed ANOVA design, considering the interaction between within-subjects and between-subjects factors. The power analysis indicated that a total sample size of 42 participants (21 per group) was required. Adjusting for an anticipated 31% attrition rate (the weighted average for internet-based treatment programs according to a systematic review by Melville et al [44]), the adjusted total sample size was 60. However, to ensure sufficient power and account for potential dropouts, we aimed for a larger sample of 100 participants, which would maintain adequate statistical power even with higher-than-expected attrition. Detailed power analysis parameters and post hoc calculations for moderation analyses are provided in Multimedia Appendix 1.

A total of 120 students were recruited, of which 100 met the inclusion criteria and were enrolled in the study. Participants were then randomly allocated in a 1:1 ratio to either the intervention group (n=50) or the waitlist control group (n=50) using a computer-generated block randomization sequence created by an independent statistician not involved in recruitment. Group allocation was concealed through a digital randomization system managed by a researcher not involved in participant recruitment or assessment. The system revealed allocations only after participants completed the baseline assessments, ensuring that neither participants nor researchers could predict group assignments.

Quantitative data were analyzed using R software (version 4.3.0; R Core Team). Before the main analyses, data were screened for outliers and violations of statistical assumptions, with no significant issues detected. Our screening process involved visual inspection of boxplots and calculation of z scores, with values exceeding ±3.29 considered potential outliers. No extreme outliers requiring special handling were identified in the dataset. Missing data (attrition rate 11%, 11/100) were handled using the last observation carried forward (LOCF) method, maintaining the intention-to-treat principle while considering the study’s short duration (7 d).

Baseline group differences in demographic characteristics and outcome measures were examined using independent t tests for continuous variables and χ² tests for categorical variables. To assess intervention effects, 2 (group: intervention vs control) × 3 (time: T1, T2, T3) mixed ANOVAs were conducted for each outcome variable, with group as the between-subjects factor and time as the within-subjects factor. Significant group×time interactions were probed with pairwise comparisons using Bonferroni correction to control for multiple testing. Effect sizes were reported using partial eta-squared (η²_p) to provide a measure of the magnitude of the intervention effects. For outcomes with significant group×time interactions, Bonferroni-corrected pairwise comparisons were conducted to identify specific patterns of change. In addition, Cohen d effect sizes were calculated to quantify the magnitude of within-group changes from baseline to postintervention (T1 to T3) and between-group differences in change scores.

To examine the potential moderating role of financial stress on the intervention’s effectiveness, a moderation analysis was first conducted on the entire sample using mixed linear models. Exploratory subgroup analyses were conducted to further examine potential differential intervention effects based on participants’ baseline financial stress levels. The sample was median-split into high and low financial stress subgroups based on baseline PIFS scores. Repeated measures ANOVAs were then conducted separately for each subgroup and outcome variable, with time as the within-subjects factor. This approach allowed us to examine whether the intervention had differential effects depending on participants’ level of financial stress.

The CONSORT (Consolidated Standards of Reporting Trials) flow diagram (Figure 2) illustrates the participant flow through each stage of the study.

The final sample comprised 100 university students (mean age 20.8, SD 2.24 y; 62/100, 62% female) who were included in the intention-to-treat analysis using LOCF imputation. There were no significant baseline differences between the intervention and waitlist control groups on any demographic or outcome measures (P>.05). Detailed sample characteristics are presented in Table 1.

Baseline group differences were examined using independent 1-tailed t tests for continuous variables (age, CES-D, GAD-7, UCLA Loneliness Scale, and PIFS) and χ² tests for categorical variables (sex). There were no significant differences in age (t_80.91=0.80; P=.43), sex (χ²₁=0.04; P=.84), CES-D scores (t_97.81=0.59; P=.56), GAD-7 scores (t_98.00=0.98; P=.33), or UCLA Loneliness scores (t_97.88=1.09; P=.28) between the intervention and waitlist control groups at baseline. Regarding financial circumstances, there was a marginally significant difference between groups in PIFS scores (t_96.05=1.81; P=.073), with the intervention group reporting slightly higher financial stress (mean 49.60, SD 14.67) than the control group (mean 44.10, SD 16.24), though this difference did not reach statistical significance at the conventional alpha level.

The overall attrition rate was 11% (11/100), with no significant difference between the intervention (6%, 3/50) and control (16%, 8/50) groups. In the intervention group, 47 participants completed all seven sessions, while 3 participants did not fully complete the intervention. In the waitlist control group, 3 participants did not complete the midintervention assessment on Day 3, and 5 participants did not provide valid responses to the postintervention assessment on Day 7. Reasons for dropout included lack of time, loss of interest, and technical difficulties with the app.

Mixed-design ANOVAs were conducted to assess the intervention effects on each outcome variable (CES-D, GAD-7, and UCLA Loneliness Scale), with group (intervention vs control) as the between-subjects factor and time (T1, T2, and T3) as the within-subjects factor. The results are presented in Table 2.

Significant group × time interactions were found for CES-D (F_{2, 196}=8.63; P<.001; η²_p=.08) and UCLA Loneliness Scale scores (F_2,196=5.57; P=.004; η²_p=.05), indicating differential changes in depression and loneliness between the intervention and waitlist control groups over time. These effect sizes represent medium effects according to Cohen benchmarks. No significant interaction effect was found for GAD-7 scores (F_2,196=1.31; P=.28; η²_p=.01).

For CES-D scores, post hoc comparisons with Bonferroni correction revealed that the intervention group showed significant reductions from baseline to post-intervention (T1 to T3: t₄₉=3.85; P<.001; P_adj=.003), as well as from midintervention to post-intervention (T2 to T3: t₄₉=2.94; P=.025; P_adj=.026). In contrast, the waitlist control group showed no significant change from baseline to postintervention (all P_adj>.05).

Similarly, for UCLA Loneliness scores, the intervention group demonstrated a significant reduction from baseline to postintervention (T1 to T3: t₄₉=4.28; P<.001; P_adj<.001) and from midintervention to postintervention (T2 to T3: t₄₉=3.44; P=.001; P_adj=.01). No significant changes were observed in the waitlist control group (all P_adj>.05).

To quantify the magnitude of these effects, we calculated Cohen d for the change scores (T1 to T3) between the intervention and waitlist control groups. For CES-D, this analysis yielded Cohen d=0.71 (95% CI 0.30‐1.17), and for UCLA Loneliness Scale, Cohen d=0.60 (95% CI 0.20‐1.00), representing medium to large effects. These effect sizes based on change scores better reflect the differential patterns of improvement indicated by the significant group×time interactions.

Figure 3 illustrates the mean scores for CES-D, GAD-7, and UCLA Loneliness Scale across time points for the intervention and waitlist control groups.

The plots in Figure 3 illustrate the differential changes in CES-D and UCLA Loneliness Scale scores between the intervention and waitlist control groups over time. For both outcomes, the intervention group (represented by the green line) shows a steeper decline in scores compared to the waitlist control group (represented by the orange line), suggesting a greater reduction in depressive symptoms and loneliness.

To examine the potential moderating role of financial stress on the intervention’s effectiveness, a moderation analysis was first conducted on the entire sample using mixed linear models. The models included financial stress, time, intervention, and their interactions as fixed effects, and subject as a random effect.

For CES-D scores, the analysis revealed significant main effects of time (F_2,192=6.96, P=.001) and financial stress (F_1,96=12.96; P<.001), as well as significant Group×Time (F_2,192=7.92; P<.001) and Time×Financial stress (F_2,192=5.90; P=.003) interactions. Most importantly, a marginally significant 3-way interaction between group, time, and financial stress was observed (F_2,192=2.86; P=.06), suggesting that the effectiveness of the intervention on depressive symptoms might vary based on participants’ financial stress levels. A post hoc power analysis indicated that with our sample size of 100 participants, α set at .05, and the corresponding effect size (f=0.17), the achieved power was 96.7% (see Figure S2 in Multimedia Appendix 1). This suggests adequate power for detecting the interaction effect in our primary analysis.

For UCLA Loneliness Scale scores, significant main effects were found for time (F_2,192=10.54; P<.001) and financial stress (F_1,96=23.08; P<.001). A significant Group×Time interaction was also observed (F_2,192=5.19; P=.006), indicating differential changes in loneliness between the intervention and waitlist control groups over time. However, the 3-way interaction between group, time, and financial stress did not reach statistical significance (F_2,192=2.32; P=.101), suggesting that the intervention’s effect on loneliness was not significantly moderated by financial stress.

To further explore potential differential effects based on financial stress levels, we conducted exploratory subgroup analyses for depression and loneliness. Participants in the intervention group were divided into high (n=27) and low (n=23) financial stress subgroups based on a median split of baseline PIFS scores (median=51). Separate repeated measures ANOVAs were conducted for each subgroup. The results are presented in Table 3.

For the high financial stress, significant time effects were found for CES-D (F_2,52=11.56, P<.001, η²_p=.31) and UCLA Loneliness Scale scores (F_2,52=11.18, P<.001, η²_p=.30), representing large effects. In contrast, for the low financial stress subgroup (n=23), no significant time effects were found for CES-D (F_2,44=1.22, P=.31, η²_p=.05) and only a marginally significant effect was observed for UCLA Loneliness in this subgroup (F_2,44=3.44, P=.04, η²_p=.14) with a substantially smaller effect size.

These results suggest that the intervention was particularly effective for participants with high financial stress, demonstrating large effects on both depression and loneliness in this subgroup, while showing limited or no significant effects for those with lower financial stress.

Figure 4 illustrates these differential patterns, with more pronounced reductions in both depression and loneliness scores among participants with high financial stress compared to those with low financial stress. For both outcomes, participants with high financial stress show a steeper decline in scores over time, indicating greater improvements in depressive symptoms and loneliness. These visualizations provide additional support for the differential effectiveness of the intervention based on participants’ level of financial stress, with those experiencing high financial stress benefiting more from the intervention in terms of reducing depressive symptoms and loneliness.

This study demonstrated the efficacy of a culturally adapted, CBT-based AI chatbot intervention in reducing depression and loneliness among Chinese university students, with more pronounced effects observed in students experiencing higher levels of financial stress. The intervention significantly reduced depressive symptoms and feelings of loneliness compared to a waitlist control condition, with moderate to large effect sizes (T1 to T3: Cohen d=0.71 and 0.60, respectively). Notably, the chatbot did not produce significant improvements in anxiety symptoms, suggesting differential effectiveness across mental health domains. Furthermore, exploratory analyses revealed that students with high financial stress benefited substantially more from the intervention than those with lower financial stress, highlighting the potential usability of this approach for addressing mental health disparities related to economic factors.

The significant reduction in depressive symptoms and loneliness among participants who engaged with the chatbot is consistent with previous studies on the effectiveness of AI-driven interventions for mental health. For example, a randomized trial of a CBT-based conversational agent (Woebot; Woebot Health) reported significant decreases in depression among young adults [25]. A scoping review likewise noted that several chatbot studies report lower loneliness, attributing this to enhanced perceived social support and engagement [24].

However, we found that the intervention did not significantly improve anxiety symptoms, which warrants further consideration, though it aligns with patterns observed in other digital mental health research. Meta-analytic data show systematically larger and more reliable effects for depression (Hedges g=0.56, 95% CI 0.38‐0.74) than anxiety (Hedges g=0.45, 95% CI 0.30‐0.6) across smartphone-based interventions [43].

Several factors may explain this differential effectiveness. First, anxiety disorders are characterized by physiological arousal components (eg, increased heart rate, muscle tension, and hypervigilance), which may be less responsive to purely cognitive-behavioral text-based interventions without complementary somatic management techniques or in-person guidance. While our chatbot included some relaxation exercises, these may have been insufficient to address the physiological aspects of anxiety effectively through a digital-only medium.

Second, the content of our chatbot focused more heavily on cognitive restructuring, targeting negative thought patterns about the self and future (which are central to depression) and behavioral activation to increase social connection (addressing loneliness), while potentially giving less attention to anxiety-specific techniques such as systematic desensitization or detailed safety behavior reduction. This content imbalance is consistent with the finding that many digital interventions emphasize depression-focused content over anxiety-specific components [18].

Third, anxiety disorders often feature avoidance behaviors that may require more intensive clinical intervention and monitoring to address effectively. Chatbot interactions may be less effective at detecting subtle avoidance patterns compared to in-person therapy. This aligns with findings from a meta-analysis, which reported stronger effects of digital interventions for depression compared to anxiety disorders, particularly when human guidance was minimal [59].

Fourth, the temporal dynamics of symptom improvement may differ between depression and anxiety. In symptom-trajectory studies, improvements in anxiety symptoms often emerge later in treatment compared to depressive symptoms. The study found that technology-based interventions for anxiety typically required a longer duration to demonstrate substantial effects [60]. Our relatively brief 7-day intervention period may have been sufficient to initiate changes in depressive symptoms but too short to observe meaningful improvements in anxiety.

Finally, cultural factors may have influenced the differential responsiveness of depression versus anxiety symptoms. In Chinese cultural contexts, expressions of depression may be more acceptable to acknowledge and address than anxiety, which is sometimes normalized as a necessary component of academic and professional achievement [32]. Cultural adaptations that more specifically target culturally shaped manifestations of anxiety might be necessary to improve the effectiveness of the intervention for anxiety symptoms in this population [36].

The exploratory subgroup analyses revealed that students experiencing high financial stress showed significant improvements in depression and loneliness, while those with low financial stress demonstrated more modest changes. This pattern, though based on exploratory analyses given the marginally significant 3-way interaction, is consistent with emerging research on the relationship between financial stress and mental health interventions.

Students with high financial stress may face additional challenges, which make them more receptive to intervention. Longitudinal evidence shows that financial difficulties predict subsequent mental health problems in university students through increased stress and reduced coping resources [10]. Our chatbot specifically addressed financial concerns through dedicated modules on cognitive restructuring for financial worries and problem-solving for financial challenges, making it particularly relevant for financially stressed students.

Financial barriers often prevent students from accessing traditional mental health services. A review found that cost concerns were among the top barriers to help-seeking among university students, with those experiencing financial hardship particularly unlikely to access in-person services [61]. The chatbot’s accessibility may have been especially valuable for these students, offering support without a financial burden. This aligns with a recent work demonstrating that digital interventions can help bridge the treatment gap for economically disadvantaged students [62].

Furthermore, the anonymity of digital interventions may be particularly important for students concerned about the stigma of help-seeking. A meta-review found that stigma concerns were heightened among individuals from lower socioeconomic backgrounds, who reported greater fears of judgment when seeking mental health support [63]. Our chatbot’s private, anonymous nature may have circumvented these concerns.

This study found that financial stress–moderated intervention effectiveness adds to emerging evidence that economic factors play a crucial role in treatment response. For instance, a web-based CBT program (“Space from Money Worries”) improved mental health and financial well-being outcomes most in participants reporting severe money concerns [64]. This suggests that comprehensive approaches that consider socioeconomic factors may enhance intervention effectiveness for vulnerable populations.

This study has several theoretical and practical implications. From a theoretical perspective, it supports the feasibility of adapting evidence-based CBT principles for delivery through AI chatbots, extending the reach of these interventions to populations who may not access traditional services. The results align with the unified protocol model of transdiagnostic CBT [65], which suggests that core therapeutic elements can effectively address common underlying processes across different psychological symptoms.

The differential effectiveness for various mental health outcomes highlights the importance of considering the specific mechanisms of change for different psychological symptoms and adapting digital interventions accordingly. This idea echoes the behavioral-intervention-technology model, which emphasizes the need to match intervention components to specific therapeutic targets and mechanisms [66].

Practically, this study supports the integration of AI-driven interventions into existing university mental health support systems. The findings suggest that such interventions can be particularly beneficial for financially stressed students, providing a way to overcome barriers to traditional mental health services. Universities could consider implementing similar chatbot interventions as part of a stepped care approach, potentially using them as an initial intervention or as a supplement to traditional services, as recommended by Lattie et al [20] in their framework for digital mental health implementation in higher education settings.

Furthermore, the cultural adaptation process demonstrated in this study provides a framework for developing culturally sensitive digital mental health interventions. The incorporation of cultural elements, such as collectivistic values, indirect communication styles, and culturally relevant metaphors, may have enhanced the acceptability and effectiveness of the intervention for Chinese students. This approach aligns with the framework for cultural adaptation of psychological interventions and could be applied to adapt digital interventions for other cultural contexts, potentially improving their global applicability.

Our sample size (N=100) was adequate for detecting primary effects according to our a priori power analysis (which determined a minimum sample of 52 participants was needed), but limited the power of subgroup and moderation analyses. The exploratory nature of our subgroup analyses, with relatively small sample sizes in each financial stress category (27 high stress and 23 low stress in the intervention group), limits the robustness of our moderation findings. Future research with larger samples specifically powered for moderation analyses is needed to confirm the differential effectiveness of the intervention based on financial stress levels.

In addition, our exclusion criteria (current psychotherapy, psychiatric-medication use, or a history of severe mental illness) restrict generalizability to the wider student population, where complex mental health needs are common. Although these criteria were necessary to isolate intervention effects and protect participants in this initial evaluation, future work should test the chatbot as an adjunct to conventional treatments in more clinically diverse samples, in keeping with recommendations that digital tools complement rather than replace routine care [67].

The use of a waitlist control group, while providing a baseline for comparison, has inherent limitations recognized in digital mental health research [45]. Waitlist controls may not account for nonspecific factors such as expectancy effects or attention. Participants in waitlist conditions may also postpone seeking help or making changes while waiting for the intervention, potentially exaggerating between-group differences. Future studies should consider active control conditions, such as psychoeducational resources or attention-matched control interventions, to better isolate the specific effects of the chatbot’s therapeutic components.

Our reliance on self-report measures may have introduced reporting biases. Cultural factors might influence how Chinese students report mental health symptoms, potentially affecting the accuracy of our assessments. Chinese students, for example, tend to emphasize somatic rather than psychological complaints, which may have affected our measurement of mental health outcomes. Future studies should incorporate multiple assessment methods, including clinician-rated measures, behavioral indicators, or physiological markers, to provide a more comprehensive evaluation of intervention effects.

The relatively brief 7-day intervention period may have been insufficient to produce meaningful changes in anxiety symptoms or to establish lasting improvements in depression and loneliness. Meta-analytic evidence indicates stronger effects for digital CBT delivered over a long period [68]. Future studies should investigate longer intervention durations and include follow-up assessments to evaluate the sustainability of effects over time.

The high adherence rate observed in our study (47/50, 94% in the intervention group) exceeds typical rates reported in digital intervention research, where adherence rates of 40%‐60% are common. While this speaks to the engaging nature of the chatbot, it may have been influenced by the monetary compensation provided to participants, particularly relevant for financially stressed students. In real-world implementations without financial incentives, adherence rates might be lower. Future research should explore sustainable engagement strategies that do not rely on monetary compensation, such as gamification elements, personalized messaging, or integration with existing university systems, which have been proposed [69] and should be explored.

While our chatbot incorporated several advanced features, it still had technical limitations. The rule-based and template-based responses (comprising 90% of interactions) may not fully capture the nuanced needs of users compared to more sophisticated AI systems. As natural language processing technology advances, future iterations could incorporate more adaptive and personalized interaction capabilities while maintaining clinical safety and appropriateness.

Finally, despite substantial efforts we made to culturally adapt the chatbot, there may be deeper cultural nuances that were not fully addressed. Future research could use more comprehensive cultural adaptation frameworks and involve larger, more diverse groups of cultural consultants and end users in the development process to enhance cultural relevance and acceptability.

This study demonstrated that a culturally adapted, CBT-based AI chatbot can effectively reduce depression and loneliness symptoms among Chinese university students, with particularly strong effects for those experiencing high financial stress. These findings highlight the potential of digital mental health interventions to address mental health disparities related to economic factors and expand access to support for underserved populations.

The differential effectiveness across mental health domains, with significant improvements in depression and loneliness but not anxiety, underscores the importance of tailoring digital interventions to specific psychological mechanisms and cultural contexts. Future developments should incorporate stronger anxiety-specific components and longer intervention periods to potentially enhance effects across a broader range of mental health outcomes.

As universities worldwide face increasing demand for mental health services amid resource constraints, culturally adapted digital interventions offer a promising approach to extend support to students who might otherwise go untreated. The moderating effect of financial stress suggests that such interventions may be particularly valuable for economically disadvantaged students, potentially helping to address socioeconomic disparities in mental health treatment access and outcomes.

Future research should focus on refining these interventions to address anxiety more effectively, investigating their long-term effects, and exploring their implementation in diverse real-world settings. In addition, studies examining the specific mechanisms through which financial stress impacts mental health and how digital interventions can best address these mechanisms would further enhance our understanding and improve intervention design for vulnerable student populations.