Frontiers in Psychiatry 系统综述与元分析：AI 对话式干预对大学生抑郁与焦虑的效果

Abstract

Background:

University students experience substantial mental-health burdens, while access to timely support remains uneven. AI conversational agents may provide low-threshold support, but the evidence base has expanded rapidly and prior reviews have often combined different age groups, intervention types, outcome domains, and comparator conditions.

Objective:

To evaluate the effects of AI-mediated conversational interventions on depressive and anxiety symptoms in university students, while distinguishing outcome domains, comparator intensity, intervention architecture, and the evidential requirements for standardized post-intervention effect estimation.

Methods:

We updated the search in PubMed, Embase, PsycINFO, CENTRAL, and Web of Science to the revised search date and rebuilt an auditable cross-database record ledger. The updated workflow included 3,786 exported records, 451 DOI/PMID duplicates, 41 confirmed title/author duplicates, and 3,294 unique records screened. All title/abstract records completed a two-reviewer screening and consensus chain. Forty-six update-search records entered research-level/full-text assessment. Randomized controlled trials were eligible for causal quantitative synthesis; relevant feasibility, acceptability, usability, and non-comparable controlled studies were retained narratively. Standard post-intervention analyses used Hedges g, REML random-effects models, and Hartung-Knapp adjustment, and were restricted to attributable two-group post-intervention data.

Results:

In the final locked non-active/low-intensity comparator analysis, three depression comparisons favored conversational interventions (k=3; Hedges g=-0.413, 95% CI -0.557 to -0.268; I2 = 0.0%) and four anxiety comparisons also favored them (k=4; Hedges g=-0.435, 95% CI -0.603 to -0.268; I2 = 2.1%). In the separate active conversational-control analysis, two MYLO comparisons did not show a reliable incremental effect for depression (k=2; g=0.165, 95% CI -0.580 to 0.911) or anxiety (k=2; g=0.190, 95% CI -0.556 to 0.937). The evidence base remained small, clinically heterogeneous, and limited by attrition, self-reported outcomes, and uncertainty regarding generalizability.

Conclusions:

AI conversational interventions may improve depressive and anxiety symptoms relative to non-active or low-intensity comparators in university students, but the certainty of evidence is low and estimates are based on few comparisons. Evidence against active conversational controls is very uncertain. Outcome-domain differences and architecture-related claims should be treated as hypothesis-generating. Secure, consent-based, human-supervised implementation and objective outcome assessment are priorities for future research.

Systematic review registration:

https://www.crd.york.ac.uk/PROSPERO/view/CRD420261323382, identifier CRD420261323382.

1 Introduction

1.1 Background: the mental health crisis among university students

Globally, the university student population is facing an increasingly severe mental health crisis. A multitude of compounding factors—including academic pressure, social adaptation challenges, and financial burdens—have contributed to the high prevalence of psychological issues such as depression and anxiety (, ). Traditional mental health service models struggle to meet the massive and urgent needs of this demographic due to inherent limitations such as restricted resources, poor accessibility, and pervasive social stigma (, ).

1.2 The role of AI conversational agents in digital psychiatry

Against this backdrop, digital mental health interventions, particularly artificial intelligence (AI) conversational agents (chatbots), have emerged as a promising solution (). These technologies demonstrate substantial application potential owing to their anonymity, convenience, cost-effectiveness, and high scalability (). By delivering personalized support grounded in evidence-based psychotherapeutic frameworks (e.g., Cognitive Behavioral Therapy) (Daley et al., 2020), AI conversational agents transcend temporal and spatial constraints, providing students with immediate and confidential mental health care ().

1.3 Why this review remains needed in a rapidly evolving evidence base

The present review is not intended to replace earlier syntheses or to claim that previous reviews were unnecessary. Existing reviews address related but non-identical questions, and their differences in population, intervention definition, outcome selection, comparator, and study-design eligibility affect the interpretation of their pooled estimates ().

First, the population boundaries differ. Chen et al. () and Feng et al () examined adolescents and young adults, while Feng et al () examined young people aged 12–25 years; these populations include, but are not limited to, university students. Li et al. () likewise synthesized chatbot interventions among young people, and Li et al. () included participants ranging from children and adolescents to adults and older adults. Leung et al. () focused on people in Asia across age groups rather than university students specifically. These broader age and geographic scopes are valuable for understanding the field, but they may combine developmental stages, educational contexts, help-seeking patterns, and campus-service environments that are not interchangeable with those of university students.

Second, the intervention boundaries differ. Zhang et al. () specifically synthesized generative or hybrid generative chatbots. Hang et al. () restricted the intervention to CBT-based, NLP-enabled conversational agents. These focused definitions provide useful evidence about particular technical or therapeutic classes, but they do not represent the full range of rule-based, generative, and hybrid participant-facing systems considered in the present review. By contrast, Lau et al. () evaluated AI-based psychotherapeutic interventions more broadly, Ferrari et al. () evaluated digital psychological interventions among university students, and Lattie et al. () reviewed digital mental-health interventions for depression, anxiety, and psychological well-being among college students. The latter categories may include digital interventions in which a conversational agent is not the active therapeutic component. Lai et al. () addressed AI in undergraduate health-professions education, which is relevant to the wider educational-AI landscape but does not answer a mental-health intervention question. Amaro et al. () described a protocol for a broad review of mental-health-promoting intervention models in university students; it is therefore not a completed chatbot-specific effectiveness synthesis.

Third, the outcome and comparator estimands are not identical. Several reviews combine depression, anxiety, stress, distress, wellbeing, positive or negative affect, and health behaviors in a single evidence landscape or across related analyses (, , –, , ). Sohn et al. () focused specifically on depressive and anxiety symptoms and therefore represents a particularly close outcome-level comparison, but included a much broader population and 39 studies. He et al. () and Ferrari et al. () illustrate the importance of comparator and outcome selection by including different active, information, passive, or digital-intervention conditions. An effect relative to a waitlist or low-intensity information condition does not answer the same question as an incremental effect relative to another conversational agent. The present review therefore synthesizes depression and anxiety separately and stratifies non-active or low-intensity comparators from active conversational controls.

The closest population-and-intervention comparison is Nyakhar and Wang (), which focused on AI chatbots among college students. This review makes an important timely contribution, but it was explicitly designed as a rapid systematic review, searched four databases, included nine studies, and used single-reviewer title/abstract screening with dual verification. The present review extends that evidence base through an updated search to 24 August 2026, a two-reviewer screening and consensus workflow, explicit report-to-study linkage, and separation of qualitative eligibility from quantitative meta-analytic eligibility. It also distinguishes depression from anxiety, separates immediate post-intervention from follow-up outcomes, and audits whether each quantitative estimate is attributable to the conversational intervention and based on compatible two-group post-intervention data.

The closest outcome-level comparison is Sohn et al. (), which synthesized chatbot effects on depression and anxiety using 39 studies. Its broader scope is important, but the inclusion of diverse populations and control conditions means that its pooled estimates answer a general chatbot-effectiveness question rather than the university-student-specific question addressed here. Chen et al. () provides another close comparator because it restricted its synthesis to depression and anxiety and randomized trials, but its adolescent-and-young-adult population and earlier search date differ from the present review. Feng et al () similarly studied 12–25-year-olds and multiple mental-health outcomes, whereas the present review isolates university or college students and applies a stricter attribution rule for standard post-intervention SMD pooling. Zhang et al. () is closest in its explicit effect-size eligibility and generative/hybrid focus, but it does not cover the full architecture range or the university-student population of the present review.

Accordingly, the incremental contribution of this review is deliberately limited and specific: a current, university-student-focused, outcome-specific, comparator-stratified, and data-source-audited synthesis of participant-facing AI-mediated conversational interventions. The review does not claim that its narrower scope makes broader reviews less useful. Rather, it makes population, intervention architecture, comparator intensity, timepoint, report linkage, and effect-size provenance visible so that estimates from non-equivalent questions are not combined as though they represented the same intervention effect. A detailed comparison of these prior reviews with the present review is provided in Supplementary Table S1.

2 Methods

2.1 Study design and registration

This systematic review and meta-analysis was conducted in strict adherence to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (). The pre-specified study protocol was prospectively registered on the International Prospective Register of Systematic Reviews (PROSPERO) under the registration number CRD420261323382 ().

2.2 Search strategy

We searched PubMed, Embase, PsycINFO, CENTRAL, and Web of Science from each database’s inception, where supported by the database, to the updated search date documented in the supplementary search log. Search terms combined population terms (university students, college students, undergraduates, and related terms), intervention terms (artificial intelligence, chatbot, conversational agent, virtual coach, generative AI, and related terms), and mental-health outcome terms (depression, anxiety, psychological distress, mental health, and related terms). Reference lists of relevant systematic reviews and eligible full-text reports were also checked.

The updated search was undertaken because the manuscript had originally been completed and submitted approximately six months earlier, during a period in which AI conversational-agent research expanded substantially. The update was therefore a complete search and screening exercise rather than a limited citation refresh.

The primary quantitative evidence base was restricted to randomized controlled trials to reduce confounding in estimates of comparative efficacy. Relevant quasi-experimental, single-arm, feasibility, acceptability, usability, and mixed-methods studies were not silently discarded: they were retained for narrative or scope-limited synthesis when they addressed the target population and intervention. They were not pooled as randomized controlled effects.

No language restriction was applied at the database retrieval stage. Any language-related limitation during full-text assessment is reported in the supplementary screening log. Grey literature and trial registries were considered as sources for identifying potentially relevant studies and ongoing or completed records; records without outcome results were logged separately and were not treated as completed evidence. The search strategy was checked against known eligible reports and the reference lists of relevant reviews. Nevertheless, database coverage, indexing differences, language availability, and unpublished-study retrieval may leave residual publication and language bias.

2.3 Eligibility criteria and intervention definition

Eligible interventions were participant-facing AI-mediated conversational systems intended to provide mental-health support, prevention, or treatment. Systems could be rule-based, generative, or hybrid. Architecture was extracted as a descriptive characteristic: rule-based systems used predefined scripts or decision rules; generative systems generated responses through a large language model or comparable generative engine; and hybrid systems combined scripted therapeutic modules with generative or natural-language components. Hybrid systems were not forced into either the rule-based or generative category.

The functional intervention definition was used because therapeutic purpose and participant-facing conversational delivery were central to the review question, whereas architecture is an evolving technical property. Architecture type, therapeutic framework, duration, contact frequency, human support, and comparator were extracted separately. Because the number of comparable studies was small and intervention components were strongly confounded with year, country, framework, and comparator, architecture-specific subgroup meta-analysis was considered exploratory and was not used to make causal claims.

2.4 Data extraction and quality assessment

Two reviewers independently assessed records and resolved discrepancies by consensus, following the review methods prespecified in the PROSPERO record and reported in accordance with PRISMA 2020 (). Reports were linked at the study level using DOI, trial registration, author/sample characteristics, intervention identity, comparator structure, and other report-level information. Multiple reports from the same trial were not counted as independent studies. The primary Shoshani JAMA Network Open report and its associated KAI report were treated as reports from one research study (). Bird and Gaffney were treated as two independent MYLO student trials (, ).

Extracted variables included participant characteristics, country, recruitment context, intervention architecture, therapeutic framework, intervention duration, contact schedule, human involvement, comparator, outcome instrument, post-intervention and follow-up timepoints, sample size, mean, standard deviation, adjusted estimates, attrition, and analysis population.

Risk of bias was assessed with RoB 2 across randomization, deviations from intended interventions, missing outcome data, outcome measurement, and selection of the reported result (). Lack of participant blinding was considered in the interpretation of deviations and self-reported outcome measurement. Certainty of evidence was assessed using GRADE, taking risk of bias, inconsistency, indirectness, imprecision, and publication bias into account ().

2.5 Statistical analysis

Depression and anxiety were synthesized separately. Hedges g was selected as the standardized mean difference because it applies a small-sample correction that is relevant to the small student trials included in this review (). Negative values indicate lower symptoms in the conversational-agent group.

The standard post-intervention SMD analysis required an attributable conversational-agent intervention, an eligible comparator, and compatible two-group post-intervention mean, standard deviation, and sample-size data. Adjusted means accompanied by standard errors were not treated as raw standard deviations. Within-group change effects were not converted into between-group post-intervention SMDs. Studies with incompatible reporting were retained in the qualitative synthesis or presented as scope-limited evidence.

REML random-effects models with Hartung-Knapp adjustment were used because clinical and methodological heterogeneity was expected and the number of comparable studies was small (, ). Non-active or low-intensity comparators were analyzed separately from active conversational comparators. Immediate post-intervention and follow-up outcomes were not pooled. Leave-one-out analyses and risk/attrition-focused sensitivity analyses were conducted where the available data supported them. Architecture, duration, and therapeutic framework were summarized descriptively; the small number of studies did not support reliable meta-regression.

Funnel plots and formal small-study tests were not used when fewer than 10 comparisons contributed to an outcome, because such tests have limited interpretability in small meta-analyses.

3 Results

3.1 Literature search and study selection

The updated search produced 3,786 exported records. After removal of 451 DOI/PMID duplicates and 41 confirmed title/author duplicates, 3,294 unique records entered title/abstract screening. All 3,294 records completed the two-reviewer screening and consensus chain. A total of 3,248 records were excluded at title/abstract level. Forty-six update-search records entered research-level/full-text assessment.

Among these 46 records, 24 had verified full text, 12 were associated reports mapped to existing verified studies, 6 were excluded at the research-level stage because the abstract or bibliographic record was sufficient to establish ineligibility, and 4 were registry records or records without outcome results. These categories are mutually exclusive in the update-search log. Historical representative reports were added to the working study registry to ensure coverage of previously locked studies, but those five backfilled reports are not substituted for the 46 update-search records in the PRISMA update-search node (Figure 1) ().

Figure 1

3.2 Characteristics of included studies

The current working independent-study registry contains 23 study clusters, including randomized studies of Woebot, Tess, XiaoNan, MISHA, UP chatbot, MYLO, KAI, Psy-Bot, and other feasibility or scope-limited conversational interventions (–, –). Twenty-two clusters have a qualitative, effectiveness, feasibility, acceptability, or scope-limited synthesis disposition; one associated technical report does not constitute an independent clinical study. The evidence spans rule-based, generative, and hybrid systems and CBT, mindfulness, psychoeducation, problem-solving, and mixed therapeutic frameworks. Intervention periods range from approximately 7 days to 16 weeks, and comparators include waitlists, informational or psychoeducational materials, low-intensity digital comparators, and active conversational agents.

The diversity of interventions and contexts is clinically important but limits direct comparability. The study registry therefore reports architecture, therapeutic framework, duration, comparator, and outcome separately rather than treating them as interchangeable features of one intervention class (Table 1).

Table 1

Characteristics of primary randomized controlled trials of AI-mediated conversational interventions among university or college students.

3.3 Risk of bias assessment

Risk-of-bias judgments are presented in the RoB 2 table and traffic-light figure (Figure 2). The most recurrent concerns relate to inability to blind participants, self-reported outcomes, attrition, and incomplete or completer-based analyses. Klos et al. contributes anxiety data using week-8 completers and is considered high risk in the missing-outcome-data domain (). These limitations reduce confidence in the magnitude of pooled effects.

Figure 2

GRADE certainty was rated as low for the non-active or low-intensity comparator evidence and very low for active conversational-comparator evidence. Random allocation does not eliminate concerns arising from missing data, subjective outcome measurement, small sample sizes, indirectness, or the limited number of independent comparisons. The GRADE certainty-of-evidence assessment is summarized in Figure 3.

Figure 3

3.4 Effect of AI interventions on depressive symptoms

In the final locked non-active/low-intensity comparator analysis, three depression comparisons contributed compatible post-intervention data (, , ). Conversational-agent interventions favored lower depressive symptoms (k=3; Hedges g=-0.413, 95% CI -0.557 to -0.268; I2 = 0.0%). This estimate should be interpreted as a short-term average across a small number of heterogeneous interventions and not as evidence that all chatbot architectures or therapeutic frameworks have equivalent effects (Figure 4).

Figure 4

3.5 Effect of AI interventions on anxiety symptoms

Four anxiety comparisons contributed compatible post-intervention data in the final locked non-active/low-intensity comparator analysis (, , , ). The pooled estimate favored conversational-agent interventions (k=4; Hedges g=-0.435, 95% CI -0.603 to -0.268; I2 = 2.1%). The small number of comparisons and the presence of attrition and self-report limitations mean that the estimate remains uncertain. It should not be interpreted as evidence that conversational agents are sufficient for acute anxiety or crisis care.

In the separate active conversational-control analysis, two MYLO comparisons did not show a reliable incremental effect for depression (k=2; Hedges g=0.165, 95% CI -0.580 to 0.911; , ) or anxiety (k=2; Hedges g=0.190, 95% CI -0.556 to 0.937). These results emphasize that comparator intensity materially changes the interpretation of apparent efficacy (Figure 5).

Figure 5

4 Discussion

4.1 Principal findings and relation to previous reviews

The systematic review provides a focused synthesis of AI-mediated conversational interventions for university students. Relative to non-active or low-intensity comparators, the final locked estimates favor improvement in both depressive and anxiety symptoms. However, the evidence base is small, and the available active conversational comparisons did not demonstrate a reliable incremental benefit. This pattern is broadly compatible with the direction of findings reported in several previous reviews, but the estimates should not be assumed to be numerically interchangeable because those reviews differ in population, intervention scope, outcomes, comparators, and search dates (, , –, , ).

The closest population-and-intervention review, Nyakhar and Wang (), focused specifically on college students and AI chatbots and included nine studies. Its rapid-review design provided a useful and timely synthesis, but its four-database search and single-reviewer initial screening differ from the two-reviewer, study-level, and auditable workflow used in the present review. The present review differs from it in several respects, including the updated search to 24 August 2026, explicit report-to-study linkage, separation of qualitative from quantitative eligibility, comparator stratification, separate depression and anxiety analyses, separation of post-intervention from follow-up outcomes, and auditing of compatible two-group post-intervention effect-size inputs. The closest outcome-level synthesis, Sohn et al. (), included 39 studies of depressive and anxiety symptoms, but its broader population and comparator conditions address a general chatbot-effectiveness question. Chen et al. () and Feng et al. () are also close in their focus on younger populations, but neither is restricted to university students; Feng et al. () further included health behaviors and a broader set of mental-health outcomes. Zhang et al. () provided a complementary generative-AI-focused synthesis, while Hang et al. () provided a CBT/NLP-specific synthesis. Their conclusions cannot be directly generalized to all architectures included in the present review.

The principal added value of this review is therefore methodological and population-specific. It separates symptom domains, comparator intensity, timepoint, architecture, and effect-size provenance. It also makes explicit which reports are independent studies and which are associated publications. These distinctions matter because broader reviews may combine adolescents with university students, general digital interventions with chatbot interventions, or active and passive comparators in analyses that target different estimands (, , , ). The present synthesis does not eliminate the limitations of the underlying trials, but it makes the boundaries of each estimate more transparent.

4.2 Depression, anxiety, and clinical interpretation

The observed estimates for depression and anxiety should not be interpreted as definitive evidence of different therapeutic mechanisms. Depression and anxiety are related but distinct symptom domains, and differences in recruitment, baseline severity, measurement instruments, treatment framework, exposure, and attrition may contribute to apparent domain differences. The small number of studies prevents a reliable test of whether architecture or symptom domain modifies treatment response.

CBT-oriented conversational delivery may support structured reflection, psychoeducation, and behavioral activation, while acute anxiety may require rapid assessment, crisis-sensitive responses, and forms of support that cannot be provided safely by an unsupervised automated agent (, ). These are plausible clinical considerations, not causal explanations established by this meta-analysis.

4.3 Comparator, engagement, and expectancy effects

Effects relative to waitlists or low-intensity materials may include nonspecific attention, hope, novelty, expectancy, and increased monitoring, in addition to therapeutic mechanisms. Active conversational-control results are therefore particularly informative, although currently sparse and imprecise. User engagement may influence exposure and adherence, but engagement should not be assumed to mediate symptom change without prospective mediation analyses. Future trials should predefine engagement measures and distinguish therapeutic content, attention, social presence, and technical usability.

4.4 Heterogeneity, attrition, and generalizability

Intervention duration varied from approximately 7 days to 16 weeks, and therapeutic frameworks and architectures also differed. These differences limit the assumption that the pooled estimate represents a single standardized dose. The REML random-effects model and Hartung-Knapp adjustment account statistically for uncertainty but cannot remove clinical heterogeneity. With few studies, subgroup analyses and meta-regression are unstable; architecture and duration findings should therefore remain descriptive.

Attrition is a major limitation. Completer analyses may overrepresent participants who tolerated or benefited from the intervention. High attrition, differential dropout, and reliance on self-reported measures reduce certainty. The evidence also spans multiple countries and educational settings, but geographic diversity alone does not establish cultural generalizability. Language, stigma, digital access, campus services, socioeconomic conditions, and expectations about AI-mediated support may influence both engagement and outcomes.

4.5 Responsible scalability, digital phenotyping, and future research

Conversational agents may offer a scalable delivery channel, but scalability is not equivalent to proven population-level effectiveness. Implementation at university scale requires local language and cultural adaptation, accessibility, integration with existing services, clear human escalation pathways, quality monitoring, and safeguards against inequitable access.

We have deliberately reframed digital phenotyping as a future research possibility rather than a demonstrated clinical function. Interaction logs may contain potentially informative behavioral signals, but their validity, reliability, clinical utility, and fairness require prospective validation (–). Any secondary use of conversational data should require meaningful consent, data minimization, purpose limitation, secure storage and transmission, role-based access controls, transparent retention policies, human oversight, and a clear response plan for safety signals. Passive monitoring must never silently replace clinical assessment or informed participation.

Future randomized trials should preregister architecture, therapeutic framework, duration, engagement, safety, and missing-data plans, and should report AI interventions transparently using relevant reporting guidance (). In addition to validated self-report scales, investigators should consider clinician-rated outcomes, behavioral indicators, physiological measures where ethically and technically justified, and longer-term follow-up. Active conversational comparators and head-to-head architecture studies are particularly important for separating therapeutic effects from attention, expectancy, and novelty.

5 Conclusions

Among university students, AI-mediated conversational interventions may reduce depressive and anxiety symptoms relative to non-active or low-intensity comparators, but the certainty of evidence is low and the number of independent comparisons is limited. Evidence from active conversational comparisons is very uncertain and does not establish a reliable incremental benefit. The results should not be generalized to crisis intervention, unsupervised clinical care, or all chatbot architectures.

The field now requires transparent, preregistered, adequately powered trials with active comparators, complete follow-up, robust missing-data handling, culturally responsive design, and objective or clinician-rated outcomes alongside self-report. Conversational data may support future research on digital phenotyping only under explicit consent, strict privacy and security protections, and human clinical oversight.

Statements

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

Author contributions

CS: Project administration, Writing – original draft, Visualization, Methodology, Investigation, Conceptualization, Writing – review & editing, Software. XC: Supervision, Methodology, Software, Writing – original draft. SQ: Formal analysis, Writing – review & editing, Investigation. AL: Formal analysis, Data curation, Writing – review & editing. SL: Visualization, Writing – review & editing. YC: Writing – original draft, Investigation.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2026.1824378/full#supplementary-material

References

• 1

  DengJZhouFHouWSilverZWongCYChangOet al. The prevalence of depressive symptoms, anxiety symptoms and sleep disturbance in higher education students during the COVID-19 pandemic: A systematic review and meta-analysis. Psychiatry Res. (2021) 301:113863. doi: 10.1016/j.psychres.2021.113863

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  Ramón-ArbuésEGea-CaballeroVGranada-LópezJMJuárez-VelaRPellicer-GarcíaBAntón-SolanasI. The prevalence of depression, anxiety and stress and their associated factors in college students. Int J Environ Res Public Health. (2020) 17:7001. doi: 10.3390/ijerph17197001

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  KoutraKPantelaiouVMavroeidesG. Breaking barriers: Unraveling the connection between mental health literacy, attitudes towards mental illness, and self-stigma of psychological help-seeking in university students. Psychol Int. (2024) 6:590–602. doi: 10.3390/psycholint6020035

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  WaoHWaoMAOmolloSBMutuaCK. Availability, accessibility, and utilization of mental health services or support among university students in Africa: A mixed methods systematic review with meta-analysis and meta-synthesis. BMC Psychiatry. (2025) 25:1082. doi: 10.1186/s12888-025-07529-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  HeYYangLQianCLiTSuZZhangQet al. Conversational agent interventions for mental health problems: Systematic review and meta-analysis of randomized controlled trials. J Med Internet Res. (2023) 25:e43862. doi: 10.2196/43862

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  BoucherEMWardNRWardHEStoecklSEVargasJMinkel,Jet al. Artificially intelligent chatbots in digital mental health interventions: A review. Expert Review of Medical Devices. (2021) 18(sup1):37–49. doi: 10.1080/17434440.2021.2013200

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  FengXTianLHoGWKYorkeJHuiV. The effectiveness of AI chatbots in alleviating mental distress and promoting health behaviors among adolescents and young adults: Systematic review and meta-analysis. J Med Internet Res. (2025) 27:e79850. doi: 10.2196/79850

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  Abd-AlrazaqAARababehAAlajlaniMBewickBMHousehM. Effectiveness and safety of using chatbots to improve mental health: Systematic review and meta-analysis. J Med Internet Res. (2020) 22(7):e16021. doi: 10.2196/16021

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  ChenTHChuGPanRHMaWF. Effectiveness of mental health chatbots in depression and anxiety for adolescents and young adults: A meta-analysis of randomized controlled trials. Expert Rev Med Devices. (2025) 22:233–41. doi: 10.1080/17434440.2025.2466742

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  FengYHangYWuWSongXXiaoXDongFet al. Effectiveness of AI-driven conversational agents in improving mental health among young people: Systematic review and meta-analysis [Review. J Med Internet Res. (2025) 27:e69639. doi: 10.2196/69639

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  LiJLiYYorkeJHuYMaDCFMeiXet al. Chatbot-delivered interventions for improving mental health among young people: A systematic review and meta-analysis. Worldviews Evidence-Based Nurs. (2025) 22:70059. doi: 10.1111/wvn.70059

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  LiHZhangRLeeYCKrautREMohrDC. Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. NPJ Digital Med. (2023) 6:236. doi: 10.1038/s41746-023-00979-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  LeungWKCLamSCChanBCLChowJNLWongYYYNgFet al. Chatbot interventions for improving mental health among people in Asia: A systematic review and meta-analysis of randomised controlled trials. BMJ Health Care Inf. (2026) 33:e101479. doi: 10.1136/bmjhci-2025-101479

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  ZhangQZhangRXiongYSuiYTongCLinFH. Generative AI mental health chatbots as therapeutic tools: Systematic review and meta-analysis of their role in reducing mental health issues. J Med Internet Res. (2025) 27:e78238. doi: 10.2196/78238

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  HangYWuWFengYYanKLiuYXiaoXet al. The effectiveness of CBT-based NLP-enabled AI conversational agents for mental health intervention: A systematic review and meta-analysis. NPJ Digital Med. (2026). doi: 10.1038/s41746-026-02886-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  LauYAngWHDAngWWWongSHPangPCIChanKS. Artificial intelligence-based psychotherapeutic intervention on psychological outcomes: A meta-analysis and meta-regression. Depression Anxiety. (2025) 2025:8930012. doi: 10.1155/da/8930012

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  FerrariMAllanSArnoldCEleftheriadisDAlvarez-JimenezMGumleyAet al. Digital interventions for psychological well-being in university students: Systematic review and meta-analysis. J Med Internet Res. (2022) 24:e39686. doi: 10.2196/39686

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  LattieEGAdkinsECWinquistNStiles-ShieldsCWaffordQEGrahamAK. Digital mental health interventions for depression, anxiety, and enhancement of psychological well-being among college students: Systematic review. J Med Internet Res. (2019) 21:e12869. doi: 10.2196/12869

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  LaiNMLimYSWinMTBhargavaPThomasPOngQC. The effectiveness of artificial intelligence in undergraduate health professions education: Systematic review and meta-analysis of randomized controlled trials. JMIR Med Educ. (2026) 12:e88933. doi: 10.2196/88933

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  ZhongWJLuoJHZhangH. The therapeutic effectiveness of artificial intelligence-based chatbots in alleviation of depressive and anxiety symptoms in short-course treatments: A systematic review and meta-analysis. J Affect Disord. (2024) 356:459–69. doi: 10.1016/j.jad.2024.04.057

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  AmaroPFonsecaCPereiraAAfonsoABarrosMLSerraIet al. Mental health-promoting intervention models in university students: A systematic review and meta-analysis protocol. BMJ Open. (2025) 15, e091297. doi: 10.1136/bmjopen-2024-091297

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  SohnJSHaBGParkSKimJJLeeEOhHet al. Systematic review and meta-analysis of chatbots in the management of depressive and anxiety symptoms. NPJ Digital Med. (2026) 9:377. doi: 10.1038/s41746-026-02566-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  NyakharSWangH. Effectiveness of artificial intelligence chatbots on mental health & well-being in college students: A rapid systematic review. Front Psychiatry. (2025) 16:1621768. doi: 10.3389/fpsyt.2025.1621768

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  PageMJMcKenzieJEBossuytPMBoutronIHoffmannTCMulrowCDet al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. Bmj. (2021) 372:n71. doi: 10.1136/bmj.n71

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  PieperDRombeyT. Where to prospectively register a systematic review. Systematic Rev. (2022) 11:8. doi: 10.1186/s13643-021-01877-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  ShoshaniAGurfinkelBKorABen-HaimYKanarekOSegevRet al. Efficacy of a conversational AI agent for psychiatric symptoms and digital therapeutic alliance: A randomized clinical trial. JAMA Netw Open. (2026) 9:e266713. doi: 10.1001/jamanetworkopen.2026.6713

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  BirdTMansellWWrightJGaffneyHTaiS. Manage your life online: A web-based randomized controlled trial evaluating the effectiveness of a problem-solving intervention in a student sample. Behav Cogn Psychother. (2018) 46:570–82. doi: 10.1017/S1352465817000820

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  GaffneyHMansellWEdwardsRWrightJ. Manage your life online (MYLO): A pilot trial of a conversational computer-based intervention for problem solving in a student sample. Behav Cogn Psychother. (2014) 42:731–46. doi: 10.1017/S135246581300060X

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  SterneJACSavovićJPageMJElbersRGBlencoweNSBoutronIet al. RoB 2: A revised tool for assessing risk of bias in randomised trials. Bmj. (2019) 366:l4898. doi: 10.1136/bmj.l4898

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  SchünemannHBrożekJGuyattGOxmanAThe GRADE Working Group. GRADE handbook for grading quality of evidence and strength of recommendations (2013). Available online at: https://gdt.gradepro.org/app/handbook/handbook.html (Accessed August 28, 2026).

  • Google Scholar
  • View reference in article

• 31

  BorensteinMHedgesLVHigginsJPTRothsteinHR. Introduction to Meta-Analysis. Chichester: John Wiley & Sons. (2021).

  • Google Scholar
  • View reference in article

• 32

  ViechtbauerW. Conducting meta-analyses in R with the metafor package. J Stat Software. (2010) 36:1–48. doi: 10.18637/jss.v036.i03

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  FitzpatrickKKDarcyAVierhileM. Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Ment Health. (2017) 4(2):e19. doi: 10.2196/mental.7785

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  KlosMCEscoredoMJoerinALemosVNRauwsMBungeEL. Artificial intelligence-based chatbot for anxiety and depression in university students: Pilot randomized controlled trial. JMIR Formative Res. (2021) 5(8):e20678. doi: 10.2196/20678

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  LiuHPengHMSongXYXuCZZhangM. Using AI chatbots to provide self-help depression interventions for university students: A randomized trial of effectiveness. Internet Interventions-The Appl Inf Technol Ment Behav Health. (2022) 27:100495. doi: 10.1016/j.invent.2022.100495

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  UlrichSLienhardNKünzliHKowatschT. A chatbot-delivered stress management coaching for students (MISHA App): Pilot randomized controlled trial. JMIR Mhealth Uhealth. (2024) 12:e54945. doi: 10.2196/54945

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  YokotaniKItoMIharaNShigeedaY. A unified protocol chatbot reduces anxiety by encouraging university students' negative emotional expressions: A randomized controlled trial. Comput Hum Behav Rep. (2025) 19:100770. doi: 10.1016/j.chbr.2025.100770

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  WangYHLiXHZhangQCYeungDNWuYH. Effect of a cognitive behavioral therapy-based AI chatbot on depression and loneliness in Chinese university students: Randomized controlled trial with financial stress moderation. JMIR Mhealth Uhealth. (2025) 13:e63806. doi: 10.2196/63806

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  HeYYangLZhuXWuBZhangSQianC. Mental health chatbot for young adults with depressive symptoms during the COVID-19 pandemic: Single-blind, three-arm randomized controlled trial. J Med Internet Res. (2022) 24:e40719. doi: 10.2196/40719

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  TorousJBucciSBellIHKessingLFaurholt-JepsenMWhelanPet al. The growing field of digital psychiatry: Current evidence and the future of apps, social media, chatbots, and virtual reality. World Psychiatry. (2021) 20:318–35. doi: 10.1002/wps.20883

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  MohrDCZhangMSchuellerSM. Personal sensing: Understanding mental health using ubiquitous sensors and machine learning. Annu Rev Clin Psychol. (2017) 13:23–47. doi: 10.1146/annurev-clinpsy-032816-044949

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  MalgaroliMSchultebraucksKMyrickKJLochAAOspina-PinillosLChoudhuryTet al. Large language models for the mental health community: Framework for translating code to care. Lancet Digital Health. (2025) 7:e282–5. doi: 10.1016/S2589-7500(24)00255-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  LiuXXRiveraSCMoherDCalvertMJDennistonAKGrpSP-AC-AW. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. Lancet Digital Health. (2020) 2:E537–48. doi: 10.1016/S2589-7500(20)30218-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

Keywords

anxiety, artificial intelligence, chatbot, depression, digital phenotyping, mental health, university students

Citation

Shao C, Chen X, Qiao S, Li A, Liu S and Chen Y (2026) Efficacy of artificial intelligence-driven conversational agents (chatbots) for mental health promotion among university students: a systematic review and meta-analysis. Front. Psychiatry 17:1824378. doi: 10.3389/fpsyt.2026.1824378

Received

06 March 2026

Revised

02 September 2026

Accepted

20 September 2026

Published

09 October 2026

Volume

17 - 2026

Edited by

Wulf Rössler, Charité University Medicine Berlin, Germany

Updates

Copyright

© 2026 Shao, Chen, Qiao, Li, Liu and Chen.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Chen Shao, shaochen@sdmu.edu.cn

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.