Effectiveness of AI-delivered Cognitive Behavioral Therapy Interventions for Anxiety and Depressive Symptoms: A Systematic Review

Effectiveness of AI-delivered Cognitive Behavioral Therapy Interventions for Anxiety and Depressive Symptoms: A Systematic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Effectiveness of AI-delivered Cognitive Behavioral Therapy Interventions for Anxiety and Depressive Symptoms: A Systematic Review Wing Lam Tiffany Yip, Ya Yang, Zilu (Lucia) Wang, David Stuckler This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6625905/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Depression and anxiety significantly contribute to the global mental health crisis. AI-delivered CBT (AI CBT) offers a scalable solution, yet existing evidence is fragmented and lacks diversity. This systematic review aims to address this gap and provide an updated and rigorous assessment of AI CBTs. We conducted a systematic review using Web of Science, PubMed, and PsycINFO, including RCTs from 2016 onward that evaluated AI CBT for anxiety or depressive symptoms. Of 208 studies screened, 16 met the criteria and were thematically synthesized. AI CBT showed limited effectiveness for anxiety, with only 1 of 14 studies reporting significant improvements. Third-wave CBT and interventions targeting youths appeared more promising in reducing anxiety. For depressive symptoms, only 4 of 16 studies showed significant improvements. Across both conditions, interventions were largely ineffective for older adults. Additionally, personalization and human support did not reliably enhance outcomes. There were concerns due to study quality. Health sciences/Medical research/Outcomes research Scientific community and society/Social sciences/Psychology Biological sciences/Computational biology and bioinformatics/Software Health sciences/Health care Artificial Intelligence AI personalization anxiety depression Cognitive Behavioral Therapy CBT chatbot mental health systematic review Introduction Depression and anxiety are two of the leading contributors to the global burden of mental health disorders 1 . Depressive symptoms are characterized by ongoing profound sadness, loss of interest in daily activities, and chronic fatigue, whereas anxiety is defined by the anticipation of imagined threats, resulting in emotional distress and physiological tension 2 . Over the past few decades, the incidence of both conditions has escalated significantly, with the COVID-19 pandemic further exacerbating these challenges 1 . Despite the availability of evidence-based treatments such as Cognitive Behavioral Therapy (CBT), a large proportion of individuals do not receive adequate care 3 . Contributing factors include a shortage of trained professionals, financial and systemic barriers, and cultural differences in the acceptability of mental health interventions 4 . To address these service limitations, Artificial Intelligence (AI)-enabled interventions have emerged as a potentially transformative approach to mental healthcare. Although AI lacks a universally accepted definition, it is generally described as a system capable of interpreting external data, learning from it, and applying this learning to achieve specific goals with adaptive flexibility 5 . AI's capacity to process large-scale data and identify complex behavioral patterns and generate personalized responses has made it an appealing tool for mental health interventions. In AI-delivered cognitive behavioral therapy (CBT), the level of personalization in user interactions depends critically on the technological sophistication of the AI system. Rule-based AI uses fixed algorithms and pre-scripted dialogue trees, offering limited personalization 6 . Keyword-based Natural Language Processing (NLP) triggers predefined responses by matching user-input words, slightly improving adaptability but remaining formulaic; Predictive machine learning–based NLP analyzes language patterns to infer intent or emotion, enabling moderate personalization through learned behavioral correlations 6 . Generative AI, powered by large language models (LLMs), dynamically generates context-aware responses in real time, achieving the highest degree of personalization through open-ended, human-like dialogue 6 . More technologically advanced features enable AI to deliver more tailored and personalized interactions for mental health interventions. While the potential of personalization for AI CBT interventions is frequently praised, research findings have been mixed. Some studies have reported favorable outcomes in feasibility, user satisfaction, and symptom reduction, whereas others have noted that users find AI CBT chatbots repetitive or lacking in emotional authenticity 7 – 10 . Past systematic reviews, while rigorous and high-quality, have evaluated AI CBT interventions primarily through the lens of earlier technologies, such as rule-based and simpler NLP systems 11 , 12 . This may overlook the growing potential of AI-based CBT interventions in light of the rapid evolution of generative AI, particularly its implications for personalization, user engagement, and therapeutic efficacy. Furthermore, while early research tended to focus on Western populations and clinical depression or anxiety as primary outcomes, recent studies have expanded geographically across Asia, Europe, and the Americas, and broadened their scope to include other health conditions—such as substance use, insomnia, and chronic illness—with depression and anxiety as secondary outcomes 13 – 16 . These findings highlight the need for more updated research that captures these diverse and emerging outcomes. Additionally, while AI CBT is frequently described as a promising solution by major healthcare providers, there is currently no high-quality systematic review examining rigorously designed studies to substantiate these claims of its effectiveness. Engagement has been noted to increase with the integration of human support into AI interventions, but questions remain about whether this increased engagement translates to improved clinical effectiveness 16 . Moreover, older studies cite novelty effects as drivers of mood improvement, further complicating interpretations of the success of AI CBT interventions 17 . Given these developments, a new systematic review is warranted to assess the effectiveness of AI-delivered CBT interventions in the current technological and clinical landscape. We must assess not only the outcomes of recent studies using advanced generative models, but also the design rigor, contextual implementation, and real-world feasibility of these interventions. To address gaps in the current literature, our study conducted a systematic review with a thematic analysis, offering a timely and rigorous examination of AI-delivered Cognitive Behavioral Therapy (AI CBT) interventions. Unlike previous reviews, which often focused on outdated forms of AI (like rule-based or predictive systems) or lacked methodological rigor, our review synthesized only high-quality, peer-reviewed randomized controlled trials (RCTs), ensuring a robust and credible evidence base. We captured the most recent and relevant literature, with most included studies published within the past few years, thereby reflecting the rapidly evolving landscape of AI applications in mental health care. Furthermore, we included studies across a diverse range of populations and clinical contexts to capture the full picture of its effectiveness. We also noted down the technological aspects of the AI CBT interventions, and how it relates to personalization. Our review also offers guidance for the design, implementation, and optimization of future AI CBT interventions and studies. Methods Design and Protocol Registration We performed our search following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines 18 . Our PRISMA Flow Chart can be found in Appendix 1 . We used the protocol published in the PROSPERO repository (CRD42024615340). Search Strategy We searched Web of Science, PubMed, and PsycINFO (via OVID) on 6th November 2024. To ensure the inclusion of newly published literature, the search was updated on 13th April 2025 using the same strategy. Our search strategy built upon previous systematic reviews 12 , 19 , operationalizing different permutations of each keyword: “AI”, “CBT”, “anxiety” and “depression”. The transformer architecture, developed in 2016, served as the foundation for various AI models, including GPT-3, which was subsequently employed in interventions 20 . Hence, the scope of this search was limited to publications from 2016 onward. The search strategies can be found in Appendix 2 . Our initial search (6th November 2024) identified a total of 282 articles: 167 articles in Web of Science, 78 articles in PubMed, and 37 articles in PsycINFO. Our updated search (13th April 2025) identified a total of 43 articles: 25 articles in Web of Science, 15 articles in PubMed, and 3 articles in PsycINFO. These articles were then imported into Zotero. 116 duplicate records and 1 retracted record were identified and removed, which left a total of 208 articles for screening and eligibility stages. Inclusion/ Exclusion Criteria We established a series of inclusion and exclusion criteria using the PICOTS framework, as follows: (1) Population: all demographic groups were considered eligible; (2) Intervention: only studies utilizing AI-delivered CBT interventions, or variations of CBT, were included. In this study, AI-delivered CBT interventions were defined as software-based agents or bots using AI models to administer and/or conduct the therapy, or components of the therapy, in order to achieve therapeutic outcomes for participants.; (3) Comparator: studies with any form of comparison were included, whether that was standard care or an alternative CBT intervention; (4) Outcome: any outcomes related to anxiety or depressive symptoms qualified, whether measured by self-reported surveys, objective indicators (e.g., passive sensing systems like audio or visual data), or third-party assessments; (5) Time: Only studies published on or after 2016-01-01 till now were included; (6) Study Design: only Randomized Controlled Trials (RCTs) were included. Only peer-reviewed articles were included. This review also excluded pilot studies to ensure high quality of data. A detailed description of the eligibility criteria based on PICO, along with examples, is provided in Appendix 3 . We then screened the remaining citations and abstracts in two stages: first by title and abstract, then through a full-text review. Out of the 208 articles screened, we excluded 32 studies because they were not randomized controlled trials (RCTs) and 99 because they did not investigate the intervention (AI CBT) relevant to our research question. All remaining records were retrieved for a full-text review, resulting in 77 articles for the full-text review stage. Upon reviewing the full texts, we excluded 3 articles for not assessing anxiety or depressive symptoms, 9 articles for not involving an AI-delivered CBT (or variations of CBT) intervention, 18 articles for not being a randomized controlled trial (RCT), 4 articles for not being peer-reviewed, 16 study protocols, and 11 pilot studies. This left us with 16 articles for the final review. Reviewer Consensus Two reviewers (YWLT, WZL) assessed all articles for inclusion in the study. Each reviewer independently determined whether the articles were relevant, and their decisions were compared to check for consistency. If the two reviewers disagreed on whether an article should be included, a third reviewer (YY) was consulted to make the final decision. For both abstract and full-text review stages, we used Cohen’s kappa to assess interrater reliability 21 . In the abstract screening stage, there was “almost perfect agreement” between the two reviewers (κ = 0.865, 95% CI: 0.788–0.942). In the full-text review stage, there was “perfect agreement” between the two reviewers (κ = 1.000, 95% CI: 1.000–1.000). Results Descriptive statistics Our systematic review identified 16 articles published between 2018 and 2025 that examined Artificial Intelligence–delivered Cognitive Behavioral Therapy (AI CBT) interventions using a randomized controlled trial (RCT) design, targeting anxiety and/or depressive symptoms. These studies covered 2,873 participants across 10 regions, with most studies from the United States (n = 5). For anxiety symptom measures, the most used tool was the Generalized Anxiety Disorder-7 (GAD-7; n = 7). For depressive symptom measures, the most used tool was the Patient Health Questionnaire-9 (PHQ-9; n = 8). Most studies utilized AI interventions with second-wave CBT components (restructuring negative thoughts and behaviors), with only 3 studies using third-wave CBT components (Acceptance and Commitment Therapy, mindfulness). There were multiple modes of delivery for AI CBT interventions: Instant messenger app (n = 6), mobile app (n = 5), and online website platform (n = 5). The AI CBT interventions used a mixture of rule-based AI (n = 16), keyword-based NLP (n = 5), predictive machine learning-NLP (n = 4), machine-learning based voice-recognition (n = 1), and generative AI (n = 1). Control groups mostly consisted of a non-CBT AI (n = 4) or waitlist control conditions (n = 5), and others utilized bibliotherapy (n = 3), a self-guided intervention (n = 3), or no intervention (n = 3). The full extraction sheet and quality assessment can be found in Appendix 4 . It is important to note that not all studies used an intention-to-treat analysis when comparing the AI CBT intervention and control groups, which analyze all randomized participants 22 . Some studies only used a completer or per-protocol analysis, which only included individuals who completed the intervention and/or satisfied all pre-defined criteria when completing the intervention 22 . Some studies provided a completer or per-protocol analysis as supplementary information. Impact of AI CBT interventions on anxiety AI CBT interventions were mostly ineffective in reducing anxiety symptoms compared to control. Using an intention-to-treat analysis, only 1 out of 14 AI CBT interventions were effective in reducing anxiety symptoms compared to the control group. In this effective study, a univariate ANCOVA revealed that the chatbot-enhanced self-help intervention showed a significant advantage to bibliotherapy control in terms of the reduction of anxiety with a low effect size (d = 0.30) 23 . When we include completer or per protocol analyses, 5 out of 14 AI CBT interventions were effective in reducing anxiety symptoms compared to the control group. AI CBT interventions with third-wave CBT components (Acceptance and Commitment Therapy, Mindfulness) were generally showed more effectiveness in reducing anxiety than those with second-wave CBT components (restructuring negative thoughts and behaviors). For example, when we include completer or per protocol analyses, all AI CBT interventions utilizing third-wave CBT components were effective in reducing anxiety symptoms, which was 3 out of 5 studies using a per protocol analysis 24 – 26 . AI CBT interventions were more effective in reducing anxiety among youth populations than older populations. Among the 5 effective studies, 2 were conducted on a university population, 2 focused on adolescents, and 1 study had a mean participant age of 31 years, which is relatively young. On the other hand, all AI CBT interventions designed for older populations aged 50 or above were ineffective in reducing anxiety symptoms. For example, the AI CBT intervention TEO (Therapy Empowerment Opportunity) was ineffective in reducing anxiety among older active workers in both Generalized Anxiety Disorder (GAD-7) and Symptom Checklist-90 (SCL-90) scales 27 . AI CBT interventions using technological designs with higher levels of personalization (free-form text input/ predictive machine-learning NLP/ Generative AI/ data-driven feedback) were not more effective in reducing anxiety outcomes compared to those with lower levels of personalization (structured input/ keyword-based NLP/ Rule-based AI). Among the 5 effective studies, 2 used rule-based AI with structured input, allowing only a low level of personalization and a high degree of rigidity in user interaction. Interestingly, despite less personalization, these AI-delivered CBT interventions yielded more effective outcomes than those using more advanced personalization features. 25 , 26 . AI CBT interventions with added human input did not improve outcomes either. In 5 out of 14 studies, the AI CBT intervention was supplemented by human input, including support from trained students, research assistants, health coaches, therapists, or integration with traditional therapy. However, no studies showed that added human support led to more effectiveness in anxiety reduction. For example, adding human coaching to the Sleep Sensei intervention was not more effective in reducing anxiety compared to using this AI CBT intervention alone 16 w. Impact of AI CBT interventions on depressive symptoms There were mixed results for AI CBT interventions in reducing depressive symptoms. Using intention-to-treat analyses, 4 out of 16 AI CBT interventions were effective in reducing depressive symptoms compared to the control group. For example, in one study, there was a significant Group X Time interaction effect that stemmed from decreases in both Emohaa AI CBT intervention groups from pre-test to post-test, but there was an increase of depression in the control group from pre-test to post-test 28 . When we include completer or per protocol analyses, 7 out of 16 AI CBT interventions were effective in reducing depressive symptoms compared to the control group. AI CBT interventions were not effective in reducing depressive symptoms among older populations compared to control. 2 out of 16 AI CBT interventions were designed for older populations (50 or above), and all of them were ineffective in reducing depressive symptoms. For example, a study examining the effects of the AI-CBT intervention MYLO on older adults with emotional distress found no significant differences in depressive symptoms between the intervention and control groups 29 . AI CBT interventions with higher levels of personalization, powered by more technologically advanced models, were not more effective in reducing depressive symptoms. Both effective and ineffective groups of studies had a mixture of more personalized (utilized Machine Learning, Natural Language Processing AI) and less personalized (rule-based AI) AI CBT interventions. AI CBT interventions with added human input did not improve outcomes either. In 5 out of 16 studies, the AI CBT intervention was supplemented by human input, including support from trained students, research assistants, health coaches, therapists, or integration with traditional therapy. However, no studies showing effectiveness in reducing depressive symptoms had human input using an intention-to-treat analysis, and only 1 study had a trained psychology student’s support when we include per protocol or completer analyses. For example, in an AI CBT intervention for insomnia, there was no significant difference in depressive symptoms between all guided dCBTi (AI groups), which included groups with human support, and dCBTi-unguided (control group) 16 . There was also no significant difference in depressive symptoms between AI groups with and without human support 16 . Risk of Bias Assessment We used the Cochrane Risk of Bias Tool 2 in our quality assessment of the papers. The full quality assessment is shown in Appendix 4 . All studies were rated as having some concerns or high risk of bias, largely due to very high attrition rates. Additionally, several studies had methodological limitations, such as the absence of intention-to-treat analyses, the lack of pre-registered protocols or analysis plans, and lack of full blinding. Discussion Our systematic review highlights a series of important findings. Overall, while there were signs of potential for both anxiety and depression alleviation, the results were mostly null. For anxiety, only 1 out of 14 AI-delivered CBT interventions showed effectiveness when analyzed using intention-to-treat. For depressive symptoms, only 4 out of 16 AI CBT interventions reported significant improvements when analyzed using intention-to-treat. Of note, neither greater personalization nor added human input appeared to enhance AI CBT interventions for anxiety or depression. AI CBT interventions also appeared to be ineffective on older populations. Our quality assessment revealed that all papers had ‘some concerns’ or ‘high risk of bias’ due to the high attrition rate of participants. This suggests that we may not be able to draw reliable conclusions from the study. One critical finding of this review is that AI-delivered third-wave CBT was more effective than second-wave CBT in reducing anxiety symptoms. However, recent meta-analyses found that second- and third-wave CBT approaches are equally effective for anxiety 30 , 31 . This suggests that the difference we observed is likely due to how well each approach fits with AI delivery, rather than the therapy itself. Second-wave CBT is content-driven and individually tailored to the client, focusing closely on challenging and changing a client's specific thoughts through nuanced, personalized dialogue 32 . In contrast, third-wave CBT is process-driven and more structured, emphasizing general techniques like mindfulness or acceptance of thoughts that apply widely across diverse situations 32 . Thus, third-wave CBT may be simpler for AI to deliver through structured and empathetic responses. Future research should directly compare how second- and third-wave CBT components interact with AI to determine which therapeutic strategies are most amenable to automation. Perhaps surprisingly, greater personalization from more advanced AIs did not appear to alter the effectiveness of AI CBT interventions, even though earlier work has linked it to higher engagement 33 – 35 . This may be due to the way advanced technology was implemented. For instance, some features may have felt surface-level or not truly responsive to users' emotional needs 36 . It is also possible that the type of AI used was not well-suited to delivering the psychological intervention effectively. While personalization may enhance initial engagement, it does not necessarily lead to better clinical outcomes if the therapeutic content itself is not aligned with users’ needs. Future research could look into why current personalization strategies may not consistently improve anxiety and depressive symptoms and explore which types of AI work best with different therapy approaches. While human input may help boost engagement by encouraging participants to complete the intervention 35 , this does not necessarily translate into improved effectiveness. In some cases, users may view human support more as a form of accountability than as something that enhances the therapeutic experience. However, it is important to note that the intensity and type of human coaching in our reviewed studies varied widely—from structured weekly sessions to minimal contact via one or two brief calls. Future studies should standardize or explicitly compare these different levels of human involvement to clarify whether and how AI can work with human support to enhance therapeutic outcomes. Our review also highlights that AI CBT interventions showed limited effectiveness among older adults. This may be linked to factors such as lower digital literacy, reduced familiarity or comfort with AI technologies, or a lack of trust in automated mental health tools. One study suggested that voice-based assistants may offer a more intuitive and accessible modality for older adults, potentially improving engagement and outcomes in this demographic 37 . As with all systematic reviews, our study is subject to several important limitations. First, it was not possible to conduct a meta-analysis for this study. This is because most studies had at least some risk of bias, owing to extremely high attrition rates and poor statistical design. Combining low-quality studies in meta-analysis could create an artificial sense of certainty. Further, although all interventions were deploying AI and CBT, their content and delivery varied markedly, which a meta-analytic approach would conflate as though it were one standardized intervention. Second, AI for mental health is rapidly evolving, drawing upon new technologies and advances in AI. While we were unable to evaluate heterogeneity based on AI-model type (e.g. ChatGPT vs. DeepSeek or BERT vs. RASA), it is possible that performance may continue to improve with technological advances. Limitations also emerged from existing studies. One is that many studies were of low methodological quality. All studies included had some concerns or high risk of bias due to attrition rates, poor statistical design, and lack of pre-registration among others. Additionally, some studies only used a per protocol analysis, which would lose the benefits of randomization and lead to an overestimation of intervention effectiveness. Second, most of the AI CBT interventions had high attrition rates and were not more effective compared to control conditions. This raises questions about the engagement and usability of these interventions in real-world settings. Additionally, not all studies provided sufficiently detailed information on the AI CBT intervention, such as its technological design, sample dialogues, user interface and features to enable direct comparisons (see point about meta-analysis above). Greater transparency and standardization are needed in reporting this new generation of AI interventions. Despite these limitations, our study has some key strengths. First, we included a diverse range of studies from various countries examining different populations, which was a key limitation in previous reviews. The AI CBT interventions also varied widely in terms of the type of AI used, the level of personalization, the mode of delivery (e.g., website, mobile app), and the specific form of CBT applied. This diversity highlights the comprehensiveness of our findings. Second, this review only included Randomized Controlled Trials (RCTs) and excluded pilot studies and those that were not peer-reviewed, to support stronger causal inference and enhance the potential for more reliable and valid conclusions 38 . Despite the growing enthusiasm around AI-delivered CBT and the widespread push from healthcare providers for its adoption, this review highlights that these interventions still require significant refinement. The promise of revolutionizing mental health care remains largely unmet at this stage. Simply increasing personalization through advanced large language models or supplementing AI with human support does not automatically lead to better outcomes. More research is urgently needed to understand the reasons behind high attrition rates, determine which forms of CBT are best suited for AI delivery, and explore how these technologies can be optimized to genuinely enhance mental health care. Additionally, modifications could be made to enhance therapeutic outcomes in AI CBT interventions. For example, one study suggested that voice assistant-based AI CBT was more user-friendly for older adults than chatbot or text-style AI CBT 37 . Additionally, having gamification elements may increase effectiveness and engagement, and including social support features may be more attractive to women 39 . Our systematic review highlights several promising avenues for future research. First, there is a need to explore a broader range of CBT modalities delivered via different types of AI to understand which approaches are most effective in different contexts. Second, future studies should prioritize methodological rigor, including standardizing reporting of AI CBT interventions, improving research design and adopting strategies to reduce attrition rates. Third, future studies evaluating the effectiveness of the newest AI-based CBT interventions for anxiety and depression are needed, with particular attention to user-centered design principles to improve real-world adoption. Finally, we also need to further investigate how AI CBT performs across different age groups. Future studies should examine which populations benefit most, whether adaptations are needed for specific age cohorts, and whether the effectiveness of AI CBT interventions is truly consistent across diverse demographic groups. This study also highlights several key implications for mental health practice and policy. Considering the findings of this study, it is important to develop a further understanding of the application of AI CBT interventions in real-world settings before and during widespread adoption in clinical contexts. In policymaking, it is also crucial to consider whether AI CBT, despite its promises for personalization, could be adopted for all populations. For instance, at least at present, this intervention does not appear effective for older people. Policymakers should prioritize rigorous, user-centered testing of AI CBT interventions to generate evidence that guides their development and integration. This approach can help prevent them from falling into the trap of techno-determinism and ensure that these tools are genuinely beneficial across diverse contexts. Declarations Acknowledgements No funding went into this manuscript. Competing Interests The first author is a Founder of a start-up company Komorebi AI Ltd., an artificial intelligence company that develops various AI services, including for mental health. There are no other competing interests to be declared. Author contributions YWLT decided on the search criteria, screened all the articles, conducted data extraction, and wrote the main manuscript text. YY acted as the third reviewer to determine whether articles should be included and was responsible for reviewing the full manuscript. WZL served as the second reviewer, screening all the articles. SD provided supervision and reviewed the manuscript. References Santomauro, D. F. et al. Global prevalence and burden of depressive and anxiety disorders in 204 countries and territories in 2020 due to the COVID-19 pandemic. The Lancet 398, 1700–1712 (2021). American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders . (2013). Patel, V. et al. Treatment and prevention of mental disorders in low-income and middle-income countries. The Lancet 370, 991–1005 (2007). Patel, V., Chowdhary, N., Rahman, A. & Verdeli, H. Improving access to psychological treatments: Lessons from developing countries. Behav. Res. Ther. 49, 523–528 (2011). Haenlein, M. & Kaplan, A. A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. Calif. Manage. Rev. 61, 5–14 (2019). Data science and AI glossary. The Alan Turing Institute https://www.turing.ac.uk/news/data-science-and-ai-glossary (2025). Oh, J., Jang, S., Kim, H. & Kim, J. Efficacy of mobile app-based interactive cognitive behavioral therapy using a chatbot for panic disorder. Int. J. Med. Inf. 140, (2020). Jang, S. et al. Mobile app-based chatbot to deliver cognitive behavioral therapy and psychoeducation for adults with attention deficit: A development and feasibility/usability study. Int. J. Med. Inf. 150, (2021). Egan, S. et al. A pilot study of the perceptions and acceptability of guidance using artificial intelligence in internet cognitive behaviour therapy for perfectionism in young people. INTERNET Interv.- Appl. Inf. Technol. Ment. Behav. Health 35, (2024). Keinert, M. et al. Facing depression: evaluating the efficacy of the EmpkinS-EKSpression reappraisal training augmented with facial expressions - protocol of a randomized controlled trial. BMC PSYCHIATRY 24, (2024). Lim, S., Shiau, C., Cheng, L. & Lau, Y. Chatbot-Delivered Psychotherapy for Adults With Depressive and Anxiety Symptoms: A Systematic Review and Meta-Regression. Behav. Ther. 53, 334–347 (2022). Li, H., Zhang, R., Lee, Y.-C., Kraut, R. E. & Mohr, D. C. Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. Npj Digit. Med. 6, 1–14 (2023). Hunt, M., Miguez, S., Dukas, B., Onwude, O. & White, S. Efficacy of Zemedy, a Mobile Digital Therapeutic for the Self-management of Irritable Bowel Syndrome: Crossover Randomized Controlled Trial. JMIR MHealth UHealth 9, (2021). Prochaska, J. et al. A Therapeutic Relational Agent for Reducing Problematic Substance Use (Woebot): Development and Usability Study. J. Med. INTERNET Res. 23, (2021). Cheng, A. et al. Behavioral Mechanisms That Mediate Mental and Physical Health Improvements in People With Chronic Pain Who Receive a Digital Health Intervention: Prospective Cohort Pilot Study. JMIR Form. Res. 7, (2023). Chan, W. et al. Assessing the Short-Term Efficacy of Digital Cognitive BehavioralTherapy for Insomnia With Different Types of Coaching:Randomized Controlled Comparative Trial. JMIR Ment. Health 11, (2024). Herbener, A., Klincewicz, M. & Damholdt, M. A narrative review of the active ingredients in psychotherapy delivered by conversational agents. Comput. Hum. Behav. Rep. 14, (2024). Page, M. J. et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372, n71 (2021). Villarreal-Zegarra, D. et al. Self-Administered Interventions Based on Natural Language Processing Models for Reducing Depressive and Anxiety Symptoms: Systematic Review and Meta-Analysis. JMIR Ment. Health 11, e59560 (2024). Cao, Y. et al. A Comprehensive Survey of AI-Generated Content (AIGC): A History of Generative AI from GAN to ChatGPT. Preprint at https://doi.org/10.48550/arXiv.2303.04226 (2023). Landis, J. R. & Koch, G. G. The Measurement of Observer Agreement for Categorical Data. Biometrics 33, 159–174 (1977). Andrade, C. Intent-to-Treat (ITT) vs Completer or Per-Protocol Analysis in Randomized Controlled Trials. Indian J. Psychol. Med. 44, 416–418 (2022). Liu, H., Peng, H., Song, X., Xu, C. & Zhang, M. Using AI chatbots to provide self-help depression interventions for university students: A randomized trial of effectiveness. Internet Interv. 27, 100495 (2022). Fulmer, R., Joerin, A., Gentile, B., Lakerink, L. & Rauws, M. Using Psychological Artificial Intelligence (Tess) to Relieve Symptoms of Depression and Anxiety: Randomized Controlled Trial. JMIR Ment. Health 5, (2018). Lappalainen, P. et al. In the shadow of COVID-19: A randomized controlled online ACT trial promoting adolescent psychological flexibility and self-compassion. J. Context. Behav. Sci. 27, 34–44 (2023). Keinonen, K. et al. Brief online ACT intervention to improve adolescents’ well-being: Effectiveness among adolescents with depressive symptoms during COVID-19. Child. Youth Serv. Rev. 161, (2024). Danieli, M. et al. Assessing the Impact of Conversational Artificial Intelligence in the Treatment of Stress and Anxiety in Aging Adults: Randomized Controlled Trial. JMIR Ment. Health 9, (2022). Sabour, S. et al. A chatbot for mental health support: exploring the impact of Emohaa on reducing mental distress in China. Front. Digit. Health 5, (2023). Bennion, M., Hardy, G., Moore, R., Kellett, S. & Millings, A. Usability, Acceptability, and Effectiveness of Web-Based Conversational Agents to Facilitate Problem Solving in Older Adults: Controlled Study. J. Med. Internet Res. 22, (2020). Papola, D. et al. Psychotherapies for Generalized Anxiety Disorder in Adults. JAMA Psychiatry 81, 250–259 (2024). Hancock, K. M. et al. Acceptance and Commitment Therapy versus Cognitive Behavior Therapy for Children With Anxiety: Outcomes of a Randomized Controlled Trial. J. Clin. Child Adolesc. Psychol. Off. J. Soc. Clin. Child Adolesc. Psychol. Am. Psychol. Assoc. Div. 53 47, 296–311 (2018). Brown, L. A., Gaudiano, B. A. & Miller, I. W. Investigating the Similarities and Differences between Practitioners of Second and Third Wave Cognitive-Behavioral Therapies. Behav. Modif. 35, 187–200 (2011). Kim, M., Yang, J., Ahn, W. & Choi, H. Machine Learning Analysis to Identify Digital Behavioral Phenotypes for Engagement and Health Outcome Efficacy of an mHealth Intervention for Obesity: Randomized Controlled Trial. J. Med. INTERNET Res. 23, (2021). Rajeshkumar, L. et al. Computer-assisted Cognitive Behavior Therapy and Mobile Apps for Depression and Anxiety: Evidence-based Digital Tools for Clinical Practice. J. Psychiatr. Pract. 30, 389–399 (2024). Tabares, M., Landa, E., Alvarez, C., Gallego, P. & Salcedo, J. Can Chatbots Alleviate Depression? Results of a Systematic Review. Adv. Artif. Intell. Mach. Learn. 4, 2665–2686 (2024). Loveys, K., Hiko, C., Sagar, M., Zhang, X. & Broadbent, E. “I felt her company”: A qualitative study on factors affecting closeness and emotional support seeking with an embodied conversational agent. Int. J. Hum.-Comput. Stud. 160, 102771 (2022). Striegl, J., Gotthardt, M., Loitsch, C. & Weber, G. Investigating the Usability of Voice Assistant-Based CBT for Age-Related Depression. in (eds. Miesenberger, K. et al.) 432–441 (2022). doi: 10.1007/978-3-031-08648-9_50 . Burns, P. B., Rohrich, R. J. & Chung, K. C. The Levels of Evidence and their role in Evidence-Based Medicine. Plast. Reconstr. Surg. 128, 305–310 (2011). Bruijniks, S. et al. The relation between therapy quality, therapy processes and outcomes and identifying for whom therapy quality matters in CBT and IPT for depression. Behav. Res. Ther. 139, (2021). Additional Declarations Competing interest reported. The first author is a Founder of a start-up company Komorebi AI Ltd., an artificial intelligence company that develops various AI services, including for mental health. There are no other competing interests to be declared. Supplementary Files ExtractiontableandQualityAssessment.xlsx Appendix.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 08 Sep, 2025 Reviews received at journal 01 Sep, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviews received at journal 29 May, 2025 Reviewers agreed at journal 26 May, 2025 Reviewers agreed at journal 15 May, 2025 Reviewers invited by journal 15 May, 2025 Editor assigned by journal 14 May, 2025 Submission checks completed at journal 14 May, 2025 First submitted to journal 09 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6625905","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":457484295,"identity":"d5e3f15d-a507-403c-8a27-6e123e0b7d63","order_by":0,"name":"Wing Lam Tiffany Yip","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/UlEQVRIiWNgGAWjYDACZjApwcAPZ0OBAUEtkg1Ea4GrOECsFoPjvIdfMLZZ5BnfPnxMuqDiDgO/RALjhx8Mh41xajnMl2bB2CZRbHYuLU16xplnDJI9B5glexgOm+HSYnaYx8wAqCVx2xkeM2netsNAexsYpBkYDtsQ1LK5h/+bNO8/oJbDDMy/CWgxfgDSsoGHh02atwFsCxvIFpwOswfawpBwTqJY4gybsfWMY4d5JHsOtln2GKTj9L5k/xnjDx/K6vL4e5gf3i6oOSzHL5F8+MaPCmvDBlx6GBjYJBLZGBJgPB4GBsYGQhHJ/IHhD0LLKBgFo2AUjAIMAACtVU0mKBPdGwAAAABJRU5ErkJggg==","orcid":"","institution":"University of Oxford","correspondingAuthor":true,"prefix":"","firstName":"Wing","middleName":"Lam Tiffany","lastName":"Yip","suffix":""},{"id":457484297,"identity":"1b787fb2-eb49-44c9-8974-148ea36cb7d6","order_by":1,"name":"Ya Yang","email":"","orcid":"","institution":"University of California, Davis","correspondingAuthor":false,"prefix":"","firstName":"Ya","middleName":"","lastName":"Yang","suffix":""},{"id":457484299,"identity":"7e001ee5-067e-42fc-a097-f9a3cfa2e898","order_by":2,"name":"Zilu (Lucia) Wang","email":"","orcid":"","institution":"University of Oxford","correspondingAuthor":false,"prefix":"","firstName":"Zilu","middleName":"(Lucia)","lastName":"Wang","suffix":""},{"id":457484305,"identity":"41abf769-438e-4f09-bdd9-4e262cca5090","order_by":3,"name":"David Stuckler","email":"","orcid":"","institution":"Bocconi University","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Stuckler","suffix":""}],"badges":[],"createdAt":"2025-05-09 07:08:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6625905/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6625905/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83101863,"identity":"b60e769d-287b-4286-9a1e-3b8c158a2bd1","added_by":"auto","created_at":"2025-05-20 05:10:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":470015,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6625905/v1/ef37277a-6021-4920-8d56-c97035b1be1f.pdf"},{"id":83101048,"identity":"85679bdd-444c-40a1-a521-cd9c0d7564ae","added_by":"auto","created_at":"2025-05-20 04:54:14","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":22288,"visible":true,"origin":"","legend":"","description":"","filename":"ExtractiontableandQualityAssessment.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6625905/v1/95b25a8ccabde3c0d594f054.xlsx"},{"id":83101054,"identity":"cef0dd50-66be-4e52-8a36-fd5875af82b3","added_by":"auto","created_at":"2025-05-20 04:54:14","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":104674,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix.docx","url":"https://assets-eu.researchsquare.com/files/rs-6625905/v1/9dc5197811ba07be0f99e23d.docx"}],"financialInterests":"Competing interest reported. The first author is a Founder of a start-up company Komorebi AI Ltd., an artificial intelligence company that develops various AI services, including for mental health. There are no other competing interests to be declared.","formattedTitle":"Effectiveness of AI-delivered Cognitive Behavioral Therapy Interventions for Anxiety and Depressive Symptoms: A Systematic Review","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDepression and anxiety are two of the leading contributors to the global burden of mental health disorders\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Depressive symptoms are characterized by ongoing profound sadness, loss of interest in daily activities, and chronic fatigue, whereas anxiety is defined by the anticipation of imagined threats, resulting in emotional distress and physiological tension \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Over the past few decades, the incidence of both conditions has escalated significantly, with the COVID-19 pandemic further exacerbating these challenges \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Despite the availability of evidence-based treatments such as Cognitive Behavioral Therapy (CBT), a large proportion of individuals do not receive adequate care \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Contributing factors include a shortage of trained professionals, financial and systemic barriers, and cultural differences in the acceptability of mental health interventions \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo address these service limitations, Artificial Intelligence (AI)-enabled interventions have emerged as a potentially transformative approach to mental healthcare. Although AI lacks a universally accepted definition, it is generally described as a system capable of interpreting external data, learning from it, and applying this learning to achieve specific goals with adaptive flexibility \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. AI's capacity to process large-scale data and identify complex behavioral patterns and generate personalized responses has made it an appealing tool for mental health interventions. In AI-delivered cognitive behavioral therapy (CBT), the level of personalization in user interactions depends critically on the technological sophistication of the AI system. Rule-based AI uses fixed algorithms and pre-scripted dialogue trees, offering limited personalization \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Keyword-based Natural Language Processing (NLP) triggers predefined responses by matching user-input words, slightly improving adaptability but remaining formulaic; Predictive machine learning\u0026ndash;based NLP analyzes language patterns to infer intent or emotion, enabling moderate personalization through learned behavioral correlations \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Generative AI, powered by large language models (LLMs), dynamically generates context-aware responses in real time, achieving the highest degree of personalization through open-ended, human-like dialogue \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. More technologically advanced features enable AI to deliver more tailored and personalized interactions for mental health interventions.\u003c/p\u003e \u003cp\u003eWhile the potential of personalization for AI CBT interventions is frequently praised, research findings have been mixed. Some studies have reported favorable outcomes in feasibility, user satisfaction, and symptom reduction, whereas others have noted that users find AI CBT chatbots repetitive or lacking in emotional authenticity \u003csup\u003e\u003cspan additionalcitationids=\"CR8 CR9\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Past systematic reviews, while rigorous and high-quality, have evaluated AI CBT interventions primarily through the lens of earlier technologies, such as rule-based and simpler NLP systems \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. This may overlook the growing potential of AI-based CBT interventions in light of the rapid evolution of generative AI, particularly its implications for personalization, user engagement, and therapeutic efficacy. Furthermore, while early research tended to focus on Western populations and clinical depression or anxiety as primary outcomes, recent studies have expanded geographically across Asia, Europe, and the Americas, and broadened their scope to include other health conditions\u0026mdash;such as substance use, insomnia, and chronic illness\u0026mdash;with depression and anxiety as secondary outcomes \u003csup\u003e\u003cspan additionalcitationids=\"CR14 CR15\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. These findings highlight the need for more updated research that captures these diverse and emerging outcomes.\u003c/p\u003e \u003cp\u003eAdditionally, while AI CBT is frequently described as a promising solution by major healthcare providers, there is currently no high-quality systematic review examining rigorously designed studies to substantiate these claims of its effectiveness. Engagement has been noted to increase with the integration of human support into AI interventions, but questions remain about whether this increased engagement translates to improved clinical effectiveness \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Moreover, older studies cite novelty effects as drivers of mood improvement, further complicating interpretations of the success of AI CBT interventions \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eGiven these developments, a new systematic review is warranted to assess the effectiveness of AI-delivered CBT interventions in the current technological and clinical landscape. We must assess not only the outcomes of recent studies using advanced generative models, but also the design rigor, contextual implementation, and real-world feasibility of these interventions.\u003c/p\u003e \u003cp\u003e To address gaps in the current literature, our study conducted a systematic review with a thematic analysis, offering a timely and rigorous examination of AI-delivered Cognitive Behavioral Therapy (AI CBT) interventions. Unlike previous reviews, which often focused on outdated forms of AI (like rule-based or predictive systems) or lacked methodological rigor, our review synthesized only high-quality, peer-reviewed randomized controlled trials (RCTs), ensuring a robust and credible evidence base. We captured the most recent and relevant literature, with most included studies published within the past few years, thereby reflecting the rapidly evolving landscape of AI applications in mental health care. Furthermore, we included studies across a diverse range of populations and clinical contexts to capture the full picture of its effectiveness. We also noted down the technological aspects of the AI CBT interventions, and how it relates to personalization. Our review also offers guidance for the design, implementation, and optimization of future AI CBT interventions and studies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDesign and Protocol Registration\u003c/h2\u003e \u003cp\u003eWe performed our search following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Our PRISMA Flow Chart can be found in \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eAppendix 1\u003c/span\u003e. We used the protocol published in the PROSPERO repository (CRD42024615340).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSearch Strategy\u003c/h3\u003e\n\u003cp\u003eWe searched Web of Science, PubMed, and PsycINFO (via OVID) on 6th November 2024. To ensure the inclusion of newly published literature, the search was updated on 13th April 2025 using the same strategy. Our search strategy built upon previous systematic reviews\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, operationalizing different permutations of each keyword: \u0026ldquo;AI\u0026rdquo;, \u0026ldquo;CBT\u0026rdquo;, \u0026ldquo;anxiety\u0026rdquo; and \u0026ldquo;depression\u0026rdquo;. The transformer architecture, developed in 2016, served as the foundation for various AI models, including GPT-3, which was subsequently employed in interventions \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Hence, the scope of this search was limited to publications from 2016 onward. The search strategies can be found in \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eAppendix 2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eOur initial search (6th November 2024) identified a total of 282 articles: 167 articles in Web of Science, 78 articles in PubMed, and 37 articles in PsycINFO. Our updated search (13th April 2025) identified a total of 43 articles: 25 articles in Web of Science, 15 articles in PubMed, and 3 articles in PsycINFO. These articles were then imported into Zotero. 116 duplicate records and 1 retracted record were identified and removed, which left a total of 208 articles for screening and eligibility stages.\u003c/p\u003e\n\u003ch3\u003eInclusion/ Exclusion Criteria\u003c/h3\u003e\n\u003cp\u003eWe established a series of inclusion and exclusion criteria using the PICOTS framework, as follows: (1) Population: all demographic groups were considered eligible; (2) Intervention: only studies utilizing AI-delivered CBT interventions, or variations of CBT, were included. In this study, AI-delivered CBT interventions were defined as software-based agents or bots using AI models to administer and/or conduct the therapy, or components of the therapy, in order to achieve therapeutic outcomes for participants.; (3) Comparator: studies with any form of comparison were included, whether that was standard care or an alternative CBT intervention; (4) Outcome: any outcomes related to anxiety or depressive symptoms qualified, whether measured by self-reported surveys, objective indicators (e.g., passive sensing systems like audio or visual data), or third-party assessments; (5) Time: Only studies published on or after 2016-01-01 till now were included; (6) Study Design: only Randomized Controlled Trials (RCTs) were included. Only peer-reviewed articles were included. This review also excluded pilot studies to ensure high quality of data. A detailed description of the eligibility criteria based on PICO, along with examples, is provided in \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eAppendix 3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWe then screened the remaining citations and abstracts in two stages: first by title and abstract, then through a full-text review. Out of the 208 articles screened, we excluded 32 studies because they were not randomized controlled trials (RCTs) and 99 because they did not investigate the intervention (AI CBT) relevant to our research question. All remaining records were retrieved for a full-text review, resulting in 77 articles for the full-text review stage. Upon reviewing the full texts, we excluded 3 articles for not assessing anxiety or depressive symptoms, 9 articles for not involving an AI-delivered CBT (or variations of CBT) intervention, 18 articles for not being a randomized controlled trial (RCT), 4 articles for not being peer-reviewed, 16 study protocols, and 11 pilot studies. This left us with 16 articles for the final review.\u003c/p\u003e\n\u003ch3\u003eReviewer Consensus\u003c/h3\u003e\n\u003cp\u003eTwo reviewers (YWLT, WZL) assessed all articles for inclusion in the study. Each reviewer independently determined whether the articles were relevant, and their decisions were compared to check for consistency. If the two reviewers disagreed on whether an article should be included, a third reviewer (YY) was consulted to make the final decision.\u003c/p\u003e \u003cp\u003eFor both abstract and full-text review stages, we used Cohen\u0026rsquo;s kappa to assess interrater reliability\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. In the abstract screening stage, there was \u0026ldquo;almost perfect agreement\u0026rdquo; between the two reviewers (κ\u0026thinsp;=\u0026thinsp;0.865, 95% CI: 0.788\u0026ndash;0.942). In the full-text review stage, there was \u0026ldquo;perfect agreement\u0026rdquo; between the two reviewers (κ\u0026thinsp;=\u0026thinsp;1.000, 95% CI: 1.000\u0026ndash;1.000).\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDescriptive statistics\u003c/h2\u003e \u003cp\u003eOur systematic review identified 16 articles published between 2018 and 2025 that examined Artificial Intelligence\u0026ndash;delivered Cognitive Behavioral Therapy (AI CBT) interventions using a randomized controlled trial (RCT) design, targeting anxiety and/or depressive symptoms. These studies covered 2,873 participants across 10 regions, with most studies from the United States (n\u0026thinsp;=\u0026thinsp;5). For anxiety symptom measures, the most used tool was the Generalized Anxiety Disorder-7 (GAD-7; n\u0026thinsp;=\u0026thinsp;7). For depressive symptom measures, the most used tool was the Patient Health Questionnaire-9 (PHQ-9; n\u0026thinsp;=\u0026thinsp;8). Most studies utilized AI interventions with second-wave CBT components (restructuring negative thoughts and behaviors), with only 3 studies using third-wave CBT components (Acceptance and Commitment Therapy, mindfulness). There were multiple modes of delivery for AI CBT interventions: Instant messenger app (n\u0026thinsp;=\u0026thinsp;6), mobile app (n\u0026thinsp;=\u0026thinsp;5), and online website platform (n\u0026thinsp;=\u0026thinsp;5). The AI CBT interventions used a mixture of rule-based AI (n\u0026thinsp;=\u0026thinsp;16), keyword-based NLP (n\u0026thinsp;=\u0026thinsp;5), predictive machine learning-NLP (n\u0026thinsp;=\u0026thinsp;4), machine-learning based voice-recognition (n\u0026thinsp;=\u0026thinsp;1), and generative AI (n\u0026thinsp;=\u0026thinsp;1). Control groups mostly consisted of a non-CBT AI (n\u0026thinsp;=\u0026thinsp;4) or waitlist control conditions (n\u0026thinsp;=\u0026thinsp;5), and others utilized bibliotherapy (n\u0026thinsp;=\u0026thinsp;3), a self-guided intervention (n\u0026thinsp;=\u0026thinsp;3), or no intervention (n\u0026thinsp;=\u0026thinsp;3). The full extraction sheet and quality assessment can be found in \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eAppendix 4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIt is important to note that not all studies used an intention-to-treat analysis when comparing the AI CBT intervention and control groups, which analyze all randomized participants\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Some studies only used a completer or per-protocol analysis, which only included individuals who completed the intervention and/or satisfied all pre-defined criteria when completing the intervention\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Some studies provided a completer or per-protocol analysis as supplementary information.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eImpact of AI CBT interventions on anxiety\u003c/h3\u003e\n\u003cp\u003eAI CBT interventions were mostly ineffective in reducing anxiety symptoms compared to control. Using an intention-to-treat analysis, only 1 out of 14 AI CBT interventions were effective in reducing anxiety symptoms compared to the control group. In this effective study, a univariate ANCOVA revealed that the chatbot-enhanced self-help intervention showed a significant advantage to bibliotherapy control in terms of the reduction of anxiety with a low effect size (d\u0026thinsp;=\u0026thinsp;0.30)\u003csup\u003e23\u003c/sup\u003e. When we include completer or per protocol analyses, 5 out of 14 AI CBT interventions were effective in reducing anxiety symptoms compared to the control group.\u003c/p\u003e \u003cp\u003eAI CBT interventions with third-wave CBT components (Acceptance and Commitment Therapy, Mindfulness) were generally showed more effectiveness in reducing anxiety than those with second-wave CBT components (restructuring negative thoughts and behaviors). For example, when we include completer or per protocol analyses, all AI CBT interventions utilizing third-wave CBT components were effective in reducing anxiety symptoms, which was 3 out of 5 studies using a per protocol analysis \u003csup\u003e\u003cspan additionalcitationids=\"CR25\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAI CBT interventions were more effective in reducing anxiety among youth populations than older populations. Among the 5 effective studies, 2 were conducted on a university population, 2 focused on adolescents, and 1 study had a mean participant age of 31 years, which is relatively young. On the other hand, all AI CBT interventions designed for older populations aged 50 or above were ineffective in reducing anxiety symptoms. For example, the AI CBT intervention TEO (Therapy Empowerment Opportunity) was ineffective in reducing anxiety among older active workers in both Generalized Anxiety Disorder (GAD-7) and Symptom Checklist-90 (SCL-90) scales \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAI CBT interventions using technological designs with higher levels of personalization (free-form text input/ predictive machine-learning NLP/ Generative AI/ data-driven feedback) were not more effective in reducing anxiety outcomes compared to those with lower levels of personalization (structured input/ keyword-based NLP/ Rule-based AI). Among the 5 effective studies, 2 used rule-based AI with structured input, allowing only a low level of personalization and a high degree of rigidity in user interaction. Interestingly, despite less personalization, these AI-delivered CBT interventions yielded more effective outcomes than those using more advanced personalization features. \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAI CBT interventions with added human input did not improve outcomes either. In 5 out of 14 studies, the AI CBT intervention was supplemented by human input, including support from trained students, research assistants, health coaches, therapists, or integration with traditional therapy. However, no studies showed that added human support led to more effectiveness in anxiety reduction. For example, adding human coaching to the Sleep Sensei intervention was not more effective in reducing anxiety compared to using this AI CBT intervention alone\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003ew.\u003c/p\u003e\n\u003ch3\u003eImpact of AI CBT interventions on depressive symptoms\u003c/h3\u003e\n\u003cp\u003eThere were mixed results for AI CBT interventions in reducing depressive symptoms. Using intention-to-treat analyses, 4 out of 16 AI CBT interventions were effective in reducing depressive symptoms compared to the control group. For example, in one study, there was a significant Group X Time interaction effect that stemmed from decreases in both Emohaa AI CBT intervention groups from pre-test to post-test, but there was an increase of depression in the control group from pre-test to post-test \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. When we include completer or per protocol analyses, 7 out of 16 AI CBT interventions were effective in reducing depressive symptoms compared to the control group.\u003c/p\u003e \u003cp\u003eAI CBT interventions were not effective in reducing depressive symptoms among older populations compared to control. 2 out of 16 AI CBT interventions were designed for older populations (50 or above), and all of them were ineffective in reducing depressive symptoms. For example, a study examining the effects of the AI-CBT intervention MYLO on older adults with emotional distress found no significant differences in depressive symptoms between the intervention and control groups \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAI CBT interventions with higher levels of personalization, powered by more technologically advanced models, were not more effective in reducing depressive symptoms. Both effective and ineffective groups of studies had a mixture of more personalized (utilized Machine Learning, Natural Language Processing AI) and less personalized (rule-based AI) AI CBT interventions.\u003c/p\u003e \u003cp\u003eAI CBT interventions with added human input did not improve outcomes either. In 5 out of 16 studies, the AI CBT intervention was supplemented by human input, including support from trained students, research assistants, health coaches, therapists, or integration with traditional therapy. However, no studies showing effectiveness in reducing depressive symptoms had human input using an intention-to-treat analysis, and only 1 study had a trained psychology student\u0026rsquo;s support when we include per protocol or completer analyses. For example, in an AI CBT intervention for insomnia, there was no significant difference in depressive symptoms between all guided dCBTi (AI groups), which included groups with human support, and dCBTi-unguided (control group)\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. There was also no significant difference in depressive symptoms between AI groups with and without human support \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eRisk of Bias Assessment\u003c/h2\u003e \u003cp\u003eWe used the Cochrane Risk of Bias Tool 2 in our quality assessment of the papers. The full quality assessment is shown in \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eAppendix 4\u003c/span\u003e. All studies were rated as having some concerns or high risk of bias, largely due to very high attrition rates. Additionally, several studies had methodological limitations, such as the absence of intention-to-treat analyses, the lack of pre-registered protocols or analysis plans, and lack of full blinding.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003e Our systematic review highlights a series of important findings. Overall, while there were signs of potential for both anxiety and depression alleviation, the results were mostly null. For anxiety, only 1 out of 14 AI-delivered CBT interventions showed effectiveness when analyzed using intention-to-treat. For depressive symptoms, only 4 out of 16 AI CBT interventions reported significant improvements when analyzed using intention-to-treat. Of note, neither greater personalization nor added human input appeared to enhance AI CBT interventions for anxiety or depression. AI CBT interventions also appeared to be ineffective on older populations. Our quality assessment revealed that all papers had \u0026lsquo;some concerns\u0026rsquo; or \u0026lsquo;high risk of bias\u0026rsquo; due to the high attrition rate of participants. This suggests that we may not be able to draw reliable conclusions from the study.\u003c/p\u003e \u003cp\u003eOne critical finding of this review is that AI-delivered third-wave CBT was more effective than second-wave CBT in reducing anxiety symptoms. However, recent meta-analyses found that second- and third-wave CBT approaches are equally effective for anxiety\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. This suggests that the difference we observed is likely due to how well each approach fits with AI delivery, rather than the therapy itself. Second-wave CBT is content-driven and individually tailored to the client, focusing closely on challenging and changing a client's specific thoughts through nuanced, personalized dialogue\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. In contrast, third-wave CBT is process-driven and more structured, emphasizing general techniques like mindfulness or acceptance of thoughts that apply widely across diverse situations\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. Thus, third-wave CBT may be simpler for AI to deliver through structured and empathetic responses. Future research should directly compare how second- and third-wave CBT components interact with AI to determine which therapeutic strategies are most amenable to automation.\u003c/p\u003e \u003cp\u003ePerhaps surprisingly, greater personalization from more advanced AIs did not appear to alter the effectiveness of AI CBT interventions, even though earlier work has linked it to higher engagement \u003csup\u003e\u003cspan additionalcitationids=\"CR34\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. This may be due to the way advanced technology was implemented. For instance, some features may have felt surface-level or not truly responsive to users' emotional needs\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. It is also possible that the type of AI used was not well-suited to delivering the psychological intervention effectively. While personalization may enhance initial engagement, it does not necessarily lead to better clinical outcomes if the therapeutic content itself is not aligned with users\u0026rsquo; needs. Future research could look into why current personalization strategies may not consistently improve anxiety and depressive symptoms and explore which types of AI work best with different therapy approaches.\u003c/p\u003e \u003cp\u003eWhile human input may help boost engagement by encouraging participants to complete the intervention\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e, this does not necessarily translate into improved effectiveness. In some cases, users may view human support more as a form of accountability than as something that enhances the therapeutic experience. However, it is important to note that the intensity and type of human coaching in our reviewed studies varied widely\u0026mdash;from structured weekly sessions to minimal contact via one or two brief calls. Future studies should standardize or explicitly compare these different levels of human involvement to clarify whether and how AI can work with human support to enhance therapeutic outcomes.\u003c/p\u003e \u003cp\u003eOur review also highlights that AI CBT interventions showed limited effectiveness among older adults. This may be linked to factors such as lower digital literacy, reduced familiarity or comfort with AI technologies, or a lack of trust in automated mental health tools. One study suggested that voice-based assistants may offer a more intuitive and accessible modality for older adults, potentially improving engagement and outcomes in this demographic \u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAs with all systematic reviews, our study is subject to several important limitations. First, it was not possible to conduct a meta-analysis for this study. This is because most studies had at least some risk of bias, owing to extremely high attrition rates and poor statistical design. Combining low-quality studies in meta-analysis could create an artificial sense of certainty. Further, although all interventions were deploying AI and CBT, their content and delivery varied markedly, which a meta-analytic approach would conflate as though it were one standardized intervention. Second, AI for mental health is rapidly evolving, drawing upon new technologies and advances in AI. While we were unable to evaluate heterogeneity based on AI-model type (e.g. ChatGPT vs. DeepSeek or BERT vs. RASA), it is possible that performance may continue to improve with technological advances.\u003c/p\u003e \u003cp\u003eLimitations also emerged from existing studies. One is that many studies were of low methodological quality. All studies included had some concerns or high risk of bias due to attrition rates, poor statistical design, and lack of pre-registration among others. Additionally, some studies only used a per protocol analysis, which would lose the benefits of randomization and lead to an overestimation of intervention effectiveness. Second, most of the AI CBT interventions had high attrition rates and were not more effective compared to control conditions. This raises questions about the engagement and usability of these interventions in real-world settings. Additionally, not all studies provided sufficiently detailed information on the AI CBT intervention, such as its technological design, sample dialogues, user interface and features to enable direct comparisons (see point about meta-analysis above). Greater transparency and standardization are needed in reporting this new generation of AI interventions.\u003c/p\u003e \u003cp\u003eDespite these limitations, our study has some key strengths. First, we included a diverse range of studies from various countries examining different populations, which was a key limitation in previous reviews. The AI CBT interventions also varied widely in terms of the type of AI used, the level of personalization, the mode of delivery (e.g., website, mobile app), and the specific form of CBT applied. This diversity highlights the comprehensiveness of our findings. Second, this review only included Randomized Controlled Trials (RCTs) and excluded pilot studies and those that were not peer-reviewed, to support stronger causal inference and enhance the potential for more reliable and valid conclusions \u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDespite the growing enthusiasm around AI-delivered CBT and the widespread push from healthcare providers for its adoption, this review highlights that these interventions still require significant refinement. The promise of revolutionizing mental health care remains largely unmet at this stage. Simply increasing personalization through advanced large language models or supplementing AI with human support does not automatically lead to better outcomes. More research is urgently needed to understand the reasons behind high attrition rates, determine which forms of CBT are best suited for AI delivery, and explore how these technologies can be optimized to genuinely enhance mental health care. Additionally, modifications could be made to enhance therapeutic outcomes in AI CBT interventions. For example, one study suggested that voice assistant-based AI CBT was more user-friendly for older adults than chatbot or text-style AI CBT \u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Additionally, having gamification elements may increase effectiveness and engagement, and including social support features may be more attractive to women \u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e Our systematic review highlights several promising avenues for future research. First, there is a need to explore a broader range of CBT modalities delivered via different types of AI to understand which approaches are most effective in different contexts. Second, future studies should prioritize methodological rigor, including standardizing reporting of AI CBT interventions, improving research design and adopting strategies to reduce attrition rates. Third, future studies evaluating the effectiveness of the newest AI-based CBT interventions for anxiety and depression are needed, with particular attention to user-centered design principles to improve real-world adoption. Finally, we also need to further investigate how AI CBT performs across different age groups. Future studies should examine which populations benefit most, whether adaptations are needed for specific age cohorts, and whether the effectiveness of AI CBT interventions is truly consistent across diverse demographic groups.\u003c/p\u003e \u003cp\u003eThis study also highlights several key implications for mental health practice and policy. Considering the findings of this study, it is important to develop a further understanding of the application of AI CBT interventions in real-world settings before and during widespread adoption in clinical contexts. In policymaking, it is also crucial to consider whether AI CBT, despite its promises for personalization, could be adopted for all populations. For instance, at least at present, this intervention does not appear effective for older people. Policymakers should prioritize rigorous, user-centered testing of AI CBT interventions to generate evidence that guides their development and integration. This approach can help prevent them from falling into the trap of techno-determinism and ensure that these tools are genuinely beneficial across diverse contexts.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding went into this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe first author is a Founder of a start-up company Komorebi AI Ltd., an artificial intelligence company that develops various AI services, including for mental health. There are no other competing interests to be declared.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYWLT decided on the search criteria, screened all the articles, conducted data extraction, and wrote the main manuscript text. YY acted as the third reviewer to determine whether articles should be included and was responsible for reviewing the full manuscript. WZL served as the second reviewer, screening all the articles. SD provided supervision and reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSantomauro, D. F. \u003cem\u003eet al.\u003c/em\u003e Global prevalence and burden of depressive and anxiety disorders in 204 countries and territories in 2020 due to the COVID-19 pandemic. The Lancet 398, 1700\u0026ndash;1712 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmerican Psychiatric Association. \u003cem\u003eDiagnostic and Statistical Manual of Mental Disorders\u003c/em\u003e. (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel, V. \u003cem\u003eet al.\u003c/em\u003e Treatment and prevention of mental disorders in low-income and middle-income countries. The Lancet 370, 991\u0026ndash;1005 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel, V., Chowdhary, N., Rahman, A. \u0026amp; Verdeli, H. Improving access to psychological treatments: Lessons from developing countries. Behav. Res. Ther. 49, 523\u0026ndash;528 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaenlein, M. \u0026amp; Kaplan, A. A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. Calif. Manage. Rev. 61, 5\u0026ndash;14 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eData science and AI glossary. The Alan Turing Institute \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.turing.ac.uk/news/data-science-and-ai-glossary\u003c/span\u003e\u003cspan address=\"https://www.turing.ac.uk/news/data-science-and-ai-glossary\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOh, J., Jang, S., Kim, H. \u0026amp; Kim, J. Efficacy of mobile app-based interactive cognitive behavioral therapy using a chatbot for panic disorder. Int. J. Med. Inf. 140, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJang, S. \u003cem\u003eet al.\u003c/em\u003e Mobile app-based chatbot to deliver cognitive behavioral therapy and psychoeducation for adults with attention deficit: A development and feasibility/usability study. Int. J. Med. Inf. 150, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEgan, S. \u003cem\u003eet al.\u003c/em\u003e A pilot study of the perceptions and acceptability of guidance using artificial intelligence in internet cognitive behaviour therapy for perfectionism in young people. INTERNET Interv.- Appl. Inf. Technol. Ment. Behav. Health 35, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKeinert, M. \u003cem\u003eet al.\u003c/em\u003e Facing depression: evaluating the efficacy of the EmpkinS-EKSpression reappraisal training augmented with facial expressions - protocol of a randomized controlled trial. BMC PSYCHIATRY 24, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLim, S., Shiau, C., Cheng, L. \u0026amp; Lau, Y. Chatbot-Delivered Psychotherapy for Adults With Depressive and Anxiety Symptoms: A Systematic Review and Meta-Regression. Behav. Ther. 53, 334\u0026ndash;347 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, H., Zhang, R., Lee, Y.-C., Kraut, R. E. \u0026amp; Mohr, D. C. Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. Npj Digit. Med. 6, 1\u0026ndash;14 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHunt, M., Miguez, S., Dukas, B., Onwude, O. \u0026amp; White, S. Efficacy of Zemedy, a Mobile Digital Therapeutic for the Self-management of Irritable Bowel Syndrome: Crossover Randomized Controlled Trial. JMIR MHealth UHealth 9, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProchaska, J. \u003cem\u003eet al.\u003c/em\u003e A Therapeutic Relational Agent for Reducing Problematic Substance Use (Woebot): Development and Usability Study. J. Med. INTERNET Res. 23, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng, A. \u003cem\u003eet al.\u003c/em\u003e Behavioral Mechanisms That Mediate Mental and Physical Health Improvements in People With Chronic Pain Who Receive a Digital Health Intervention: Prospective Cohort Pilot Study. JMIR Form. Res. 7, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan, W. \u003cem\u003eet al.\u003c/em\u003e Assessing the Short-Term Efficacy of Digital Cognitive BehavioralTherapy for Insomnia With Different Types of Coaching:Randomized Controlled Comparative Trial. JMIR Ment. Health 11, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHerbener, A., Klincewicz, M. \u0026amp; Damholdt, M. A narrative review of the active ingredients in psychotherapy delivered by conversational agents. Comput. Hum. Behav. Rep. 14, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePage, M. J. \u003cem\u003eet al.\u003c/em\u003e The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372, n71 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVillarreal-Zegarra, D. \u003cem\u003eet al.\u003c/em\u003e Self-Administered Interventions Based on Natural Language Processing Models for Reducing Depressive and Anxiety Symptoms: Systematic Review and Meta-Analysis. JMIR Ment. Health 11, e59560 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao, Y. \u003cem\u003eet al.\u003c/em\u003e A Comprehensive Survey of AI-Generated Content (AIGC): A History of Generative AI from GAN to ChatGPT. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2303.04226\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2303.04226\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLandis, J. R. \u0026amp; Koch, G. G. The Measurement of Observer Agreement for Categorical Data. Biometrics 33, 159\u0026ndash;174 (1977).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAndrade, C. Intent-to-Treat (ITT) vs Completer or Per-Protocol Analysis in Randomized Controlled Trials. Indian J. Psychol. Med. 44, 416\u0026ndash;418 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, H., Peng, H., Song, X., Xu, C. \u0026amp; Zhang, M. Using AI chatbots to provide self-help depression interventions for university students: A randomized trial of effectiveness. Internet Interv. 27, 100495 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFulmer, R., Joerin, A., Gentile, B., Lakerink, L. \u0026amp; Rauws, M. Using Psychological Artificial Intelligence (Tess) to Relieve Symptoms of Depression and Anxiety: Randomized Controlled Trial. JMIR Ment. Health 5, (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLappalainen, P. \u003cem\u003eet al.\u003c/em\u003e In the shadow of COVID-19: A randomized controlled online ACT trial promoting adolescent psychological flexibility and self-compassion. J. Context. Behav. Sci. 27, 34\u0026ndash;44 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKeinonen, K. \u003cem\u003eet al.\u003c/em\u003e Brief online ACT intervention to improve adolescents\u0026rsquo; well-being: Effectiveness among adolescents with depressive symptoms during COVID-19. Child. Youth Serv. Rev. 161, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDanieli, M. \u003cem\u003eet al.\u003c/em\u003e Assessing the Impact of Conversational Artificial Intelligence in the Treatment of Stress and Anxiety in Aging Adults: Randomized Controlled Trial. JMIR Ment. Health 9, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSabour, S. \u003cem\u003eet al.\u003c/em\u003e A chatbot for mental health support: exploring the impact of Emohaa on reducing mental distress in China. Front. Digit. Health 5, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBennion, M., Hardy, G., Moore, R., Kellett, S. \u0026amp; Millings, A. Usability, Acceptability, and Effectiveness of Web-Based Conversational Agents to Facilitate Problem Solving in Older Adults: Controlled Study. J. Med. Internet Res. 22, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePapola, D. \u003cem\u003eet al.\u003c/em\u003e Psychotherapies for Generalized Anxiety Disorder in Adults. JAMA Psychiatry 81, 250\u0026ndash;259 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHancock, K. M. \u003cem\u003eet al.\u003c/em\u003e Acceptance and Commitment Therapy versus Cognitive Behavior Therapy for Children With Anxiety: Outcomes of a Randomized Controlled Trial. J. Clin. Child Adolesc. Psychol. Off. J. Soc. Clin. Child Adolesc. Psychol. Am. Psychol. Assoc. Div. \u003cem\u003e53\u003c/em\u003e 47, 296\u0026ndash;311 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown, L. A., Gaudiano, B. A. \u0026amp; Miller, I. W. Investigating the Similarities and Differences between Practitioners of Second and Third Wave Cognitive-Behavioral Therapies. Behav. Modif. 35, 187\u0026ndash;200 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, M., Yang, J., Ahn, W. \u0026amp; Choi, H. Machine Learning Analysis to Identify Digital Behavioral Phenotypes for Engagement and Health Outcome Efficacy of an mHealth Intervention for Obesity: Randomized Controlled Trial. J. Med. INTERNET Res. 23, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajeshkumar, L. \u003cem\u003eet al.\u003c/em\u003e Computer-assisted Cognitive Behavior Therapy and Mobile Apps for Depression and Anxiety: Evidence-based Digital Tools for Clinical Practice. J. Psychiatr. Pract. 30, 389\u0026ndash;399 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTabares, M., Landa, E., Alvarez, C., Gallego, P. \u0026amp; Salcedo, J. Can Chatbots Alleviate Depression? Results of a Systematic Review. Adv. Artif. Intell. Mach. Learn. 4, 2665\u0026ndash;2686 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLoveys, K., Hiko, C., Sagar, M., Zhang, X. \u0026amp; Broadbent, E. \u0026ldquo;I felt her company\u0026rdquo;: A qualitative study on factors affecting closeness and emotional support seeking with an embodied conversational agent. Int. J. Hum.-Comput. Stud. 160, 102771 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStriegl, J., Gotthardt, M., Loitsch, C. \u0026amp; Weber, G. Investigating the Usability of Voice Assistant-Based CBT for Age-Related Depression. in (eds. Miesenberger, K. et al.) 432\u0026ndash;441 (2022). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-031-08648-9_50\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-08648-9_50\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBurns, P. B., Rohrich, R. J. \u0026amp; Chung, K. C. The Levels of Evidence and their role in Evidence-Based Medicine. Plast. Reconstr. Surg. 128, 305\u0026ndash;310 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBruijniks, S. \u003cem\u003eet al.\u003c/em\u003e The relation between therapy quality, therapy processes and outcomes and identifying for whom therapy quality matters in CBT and IPT for depression. Behav. Res. Ther. 139, (2021).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial Intelligence, AI, personalization, anxiety, depression, Cognitive Behavioral Therapy, CBT, chatbot, mental health, systematic review","lastPublishedDoi":"10.21203/rs.3.rs-6625905/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6625905/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDepression and anxiety significantly contribute to the global mental health crisis. AI-delivered CBT (AI CBT) offers a scalable solution, yet existing evidence is fragmented and lacks diversity. This systematic review aims to address this gap and provide an updated and rigorous assessment of AI CBTs. We conducted a systematic review using Web of Science, PubMed, and PsycINFO, including RCTs from 2016 onward that evaluated AI CBT for anxiety or depressive symptoms. Of 208 studies screened, 16 met the criteria and were thematically synthesized. AI CBT showed limited effectiveness for anxiety, with only 1 of 14 studies reporting significant improvements. Third-wave CBT and interventions targeting youths appeared more promising in reducing anxiety. For depressive symptoms, only 4 of 16 studies showed significant improvements. Across both conditions, interventions were largely ineffective for older adults. Additionally, personalization and human support did not reliably enhance outcomes. There were concerns due to study quality.\u003c/p\u003e","manuscriptTitle":"Effectiveness of AI-delivered Cognitive Behavioral Therapy Interventions for Anxiety and Depressive Symptoms: A Systematic Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-20 04:54:09","doi":"10.21203/rs.3.rs-6625905/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-08T14:36:53+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-01T17:14:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"933526329752274582778719169459485926","date":"2025-08-11T19:20:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-29T16:54:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"248451916293924192544847607627850855806","date":"2025-05-26T17:06:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"327834334600396143038463046786334419468","date":"2025-05-15T12:48:09+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-15T12:24:23+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-15T00:28:52+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-14T17:49:06+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2025-05-09T06:59:34+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a67ac393-d380-482e-aac6-9c8319bca1e7","owner":[],"postedDate":"May 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":48621020,"name":"Health sciences/Medical research/Outcomes research"},{"id":48621021,"name":"Scientific community and society/Social sciences/Psychology"},{"id":48621022,"name":"Biological sciences/Computational biology and bioinformatics/Software"},{"id":48621023,"name":"Health sciences/Health care"}],"tags":[],"updatedAt":"2026-05-02T15:38:49+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-20 04:54:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6625905","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6625905","identity":"rs-6625905","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}