Effectiveness of AI and rule-based conversational agents for depression, anxiety and stress: A meta-analysis Authors: Saeed Mokhtari Masoumi Alamdarloo, Ali Mirzakhani, Mohamad Azhdarloo, Elaheh Haghani-Samani, Abouzar Nazari Published: 2026 ABSTRACT AbstractDepression, anxiety, and stress are significant global health burdens worsened by restricted access to care. Conversational agents (CAs), encompassing both AI(Artificial Intelligence)- and rule-based systems, provide scalable mental health interventions. This systematic review and meta-analysis analyzed the efficacy of CA interventions in 48 randomized controlled trials covering 28,071 participants. It demonstrated small-to-moderate but significant effects in minimizing symptoms of depression (SMD: −0.27), anxiety (SMD: −0.20), and stress (SMD: −0.26). A Robust Variance Estimation (RVE) approach was used to address within-study dependencies, with results consistent with the random-effects model. Subgroup analyses investigated sources of heterogeneity and indicated greater effects in clinical populations and shorter-duration interventions. Meta-regression identified no appreciable effect of outcome timing. Risk of bias was largely low, although some studies had issues due to missing data or selective reporting. Publication bias was negligible. These results provide evidence for the utilization of conversational agents as effective interventions for lessening psychological symptoms. Additional research is required to evaluate their long-term effectiveness and how best to integrate them into mental healthcare systems.Subject terms: Diseases, Health care, Medical research, Signs and symptoms Introduction Depression, anxiety disorders, and stress-related disorders are the most disabling and prevalent mental disorders globally. As estimated by the World Health Organization (WHO), over 280 million individuals suffer from depression1 and around 301 million individuals suffer from anxiety disorders annually2, and stress-related disorders have an equally large percentage of the global burden of disease and disability-adjusted life years. These conditions, in addition to diminishing the well-being and functioning of the people, put substantial pressure on public health systems and national economies in terms of excess use of healthcare, absenteeism, and lost productivity3. Even though they exert significant impacts, provision of effective and timely mental health care is predominantly limited by structural factors such as a lack of trained staff, cost of treatment, geographical disparity, and persistent stigma against mental illness, particularly in low- and middle-income environments4. Responding to these challenges, there has been a massive advancement in creating and applying digital interventions for mental health. Amongst these, conversational agents (CAs) have garnered significant interest because of their potential to offer low-cost, accessible, and scalable psychological therapy. Conversational agents are software-based entities that engage users with the use of natural language interfaces, normally simulating human conversation. They have found application in numerous mental health treatments such as mood monitoring, Cognitive Behavioral Therapy (CBT) delivery, and psychoeducation5–7. The compounds can function on multiple platforms—cell phones, messaging apps, and the web—and are particularly appealing in that they can provide on-demand, anonymous support, thus circumventing some of the most recognized barriers to traditional care8,9. Interestingly, conversational agents may be typically characterized by their architecture and response generation mechanisms. At one end are AI-based agents, which implement natural language processing (NLP), machine learning (ML), or large language models (LLMs) to look at user inputs dynamically and generate context-apt responses. These may be chatbots like Woebot, which deliver interventions in the form of CBT through NLP, or more sophisticated agents employing affective computing to facilitate emotion detection and regulation of CBT responses10–12. At the other extreme are rule- or script-based systems, that follow pre-scripted decision trees or conversation flows and do not enable adaptive learning. While less personalized, they are simpler to put in place and have been used successfully in structured intervention formats13. Recent systematic reviews and meta-analyses have increasingly examined the role of AI-based conversational agents in mental health. For instance, Li et al. (2023)14 demonstrated the effectiveness of AI-driven conversational agents in improving mental health outcomes across diverse populations. Similarly, Feng et al. (2025)15 reported positive effects of AI-based chatbots in reducing symptoms of depression and anxiety among young people. More recent evidence, including Du et al. (2025)16, has begun to compare rule-based and large language model (LLM)-based systems, highlighting potential differences in adaptability and therapeutic engagement. In addition, emerging meta-analyses such as Sohn et al. (2026)17 further confirm the growing evidence base supporting chatbot interventions in mental health care. Despite these advances, most existing reviews focus either on AI-based systems alone or on specific populations or disorders, and few provide a comprehensive comparison between AI-based and rule-based conversational agents across multiple psychological outcomes. This gap underscores the need for updated and integrative evidence synthesis11. Despite the growing body of papers supporting the use of both AI-based vs rule-based Cas in mental health contexts, few have explicitly compared their relative effectiveness over well-used psychological outcomes such as depression, anxiety, and stress14,15,18,19. Most existing reviews either take a broad view of digital mental health interventions or focus exclusively on AI-driven systems14,20–22, often neglecting comparisons between different conversational agent architectures. This leaves an important gap in understanding how agent type—whether adaptive and AI-based or rule-based and scripted—may influence therapeutic outcomes23. This study aims to address this significant gap in knowledge. In addition, the accelerating pace of AI feature development—particularly for generative models—requires ongoing assessments of their relative advantages compared to more conventional digital tools16,17. Therefore, this study systematically evaluates the effectiveness of AI-based and rule-based conversational agents in reducing symptoms of depression, anxiety, and stress. Results Results of Systematic Review Database searches turned up 3976 records in all 2787 records were left for title and abstract screening after 693 records and 496 duplicates were eliminated using automated tools. Of these, 66 full-text articles were evaluated for eligibility after 2721 were eliminated. For various reasons, including non-RCT design, irrelevant outcomes, or ineligible interventions, 18 studies were eliminated. In the end, the systematic review contained 48 studies (Fig. 1). Additionally, the first 200 records retrieved from Google Scholar were screened, but no new articles were identified. A subset of these randomized controlled trials provided sufficient data for inclusion in the meta-analysis focused on depression, anxiety, and stress. Prisma flow diagram of included studies. Flow diagram illustrating the process of identification, screening, eligibility assessment, and inclusion of studies in the systematic review. The diagram shows the number of records identified through database searching and other sources, records after duplicates were removed, records screened, full-text articles assessed for eligibility, and studies included in the final analysis. Among the included studies, 41 trials assessed depression-related outcomes24–64, 34 reported on anxiety10,24–26,28–32,34–39,41,44–48,51–53,56,58,60,61,63–67, and 19 evaluated stress34,36,39,43,46,47,51–54,59,61,62,64,68–72 (Fig. 2). Summary of psychological outcomes evaluated in the studies. This figure presents the number of included randomized controlled trials assessing each psychological outcome, including depression, anxiety, and stress. It summarizes the frequency of studies reporting each outcome domain, showing the relative distribution of evidence across the three main psychological endpoints. The included studies had sample sizes ranging from 17 to more than 19,643 participants, and included 28,071 participants from 31 different countries. Given the quick rise in interest in AI and conversational agent (CA)-based mental health interventions, the earliest included study was published in 2015, and the most recent was in 2025. University students, medical professionals, community samples, and people with diagnosed mental health conditions were among the clinical and non-clinical populations included in the majority of studies, which were carried out in high-income nations. Conversational agents used in the included studies comprised different types of systems, including rule-based (scripted or decision-tree–based) agents (n = 18, 58.1%), retrieval-based or NLP/ML-based systems (n = 9, 29.0%), and generative AI-based models (n = 2, 6.5%), while a small number employed hybrid or unspecified approaches (n = 2, 6.5%). These agents were integrated into a variety of platforms, including mobile applications (n = 18, 58.1%), web-based systems (n = 8, 25.8%), instant messaging platforms (n = 5, 16.1%), and mixed or multi-platform delivery modes (n = 3, 9.7%). Cognitive behavioral therapy (CBT) was the most frequently used therapeutic approach (n = 20, 64.5%), and the majority of interventions incorporated psychotherapy or psychoeducational components (n = 26, 83.9%). Additional features such as gamification, mood tracking, personalization, and empathic responses were reported in several studies (n = 11, 35.5%). The duration of interventions varied substantially, ranging from single-session or brief interventions (n = 6, 19.4%) to multi-week programs lasting 12 weeks or longer (n = 12, 38.7%), with the remainder falling between these durations (n = 13, 41.9%). Most studies employed text-based interaction modes (n = 26, 83.9%), while a smaller subset used multimodal interaction including voice or embodied avatars (n = 5, 16.1%). Finally, a subset of studies incorporated adaptive dialogue techniques, personalization algorithms, or emotion recognition features such as sentiment analysis (n = 10, 32.3%). Regarding the timing of outcome assessment and follow-up periods, there was substantial variability across studies. Outcome measurements were most commonly conducted immediately post-intervention and at 4 weeks (both 29.2%, n = 14/48), followed by 8 weeks (16.7%, n = 8/48) and 2 weeks (14.6%, n = 7/48). Fewer studies assessed outcomes at 12 weeks (8.3%, n = 4/48) or 24 weeks (6.3%, n = 3/48), while assessments at 1, 6, 7, 16, 20, 22, and 28 weeks were rare. Follow-up durations showed a similar pattern, with the most common follow-up periods being 4 weeks (29.2%, n = 14/48), 8 weeks (16.7%, n = 8/48), and 2 weeks (14.6%, n = 7/48). In contrast, 18.8% (n = 9/48) of studies did not include any follow-up assessment. Overall, the evidence base is characterized by a predominance of short-term evaluations, with limited long-term follow-up data. Less than half of the studies specifically addressed safety procedures, human backup, or adverse event monitoring, despite the fact that safety and ethical issues are becoming more and more important in AI mental health applications. Furthermore, few studies systematically evaluated user engagement’s moderating role, despite the fact that it was frequently mentioned as a crucial element of intervention success. Table 1 offers a thorough synopsis of the study’s attributes. Table 1Major characteristics of studies included in the systematic reviewAuthor(year)CountryBlindingSample SizeSample typeAgeTherapeutic approachCA typeResponse generation approach, AI framework/ techniqueInteraction modeDelivery platformControl ConditionPsychological outcomesMeasurement ToolsOutcome Timing (week)Follow-up Duration (week)Main FindingsNicol (2022)41USAOpen-label17Nonclinical(Adolescents, primary care)14.7CBT (Woebot, W-GenZ)ChatbotNLP + ML (proprietary, Woebot)Text-basedSmartphoneWaitlist + usual careDepression, AnxietyPHQ-9, GAD-744Greater reduction in PHQ-9 and GAD-7 in IG compared to CG; large effect sizes but underpowered study.Ogawa (2022)42JapanOpen-label20clinical(Parkinson’s disease)NRHealth monitoring/chatChatbotScripted, NLP (IBM Watson, custom facial analysis)Text & videoiPad/web/telehealthWeekly teleconsultation onlydepressionBDI-II2020Improved smile and speech features in IG; no significant effect on depression or quality of life.Burton (2016)50Romania, Spain, UKOpen-label27Clinical (depression diagnosis)38.65CBT elementsEmbodied virtual agent (avatar)Rule-based, adaptive content via DSS and PMSText-basedLaptop computerUsual careDepressionBDI-II, QIDS-SR, PSS, SCL-90-R44TEO group showed greater reduction in depression and stress, but differences were not statistically significant; both groups had similar anxiety reductions.Ly (2017)69SwedenOpen-label28Nonclinical (university students/adults)26.2Positive Psychology + CBTChatbot (Shim)Rule-based, scriptedText-basediPhone appWaitlistStressPSS-1022No significant group × time interaction in ITT analysis; significant stress reduction observed only among completers.Nguyen (2024)40USAOpen-label29Nonclinical(Pregnant women or new mothers of color)31.7Health education / supportChatbotDense passage retrieval model (QA), NLP, RasaText-basedMobile (iOS/Android)Book club (monthly children’s book)DepressionPHQ-91212Significant reduction in depression in IG only (P = .008); no significant change in CG (P = .07).Stasinaki (2021)71SwitzerlandOpen-label31Clinical (adolescents, obesity)13.6Standard counseling + PathMate2 appChatbot (app, Anna/Lukas)Preprogrammed, gamified chatText-basedMobile appStandard BCI (no app)StressTICS2222Intervention improved fitness/body composition; only CG reduced BMI-SDS. No significant stress reduction in either group.Sadeh-Sharvit (2023)58USAOpen-label32Clinical(Outpatients, depression/anxiety)30.6CBT (clinician-led)AI therapist supportAI transcription + feedbackTherapist+patient, in-personClinicUsual careDepression, AnxietyPHQ-9, GAD-74, 84, 8Greater reductions in PHQ-9 and GAD-7 scores in the AI-supported group at 1 and 2 months.Greer (2019)32United StatesUnclear34Clinical(young adults with depression/anxiety symptoms; cancer survivors and non-cancer controls)37.56CBTChatbot (Woebot)Rule-basedText-basedMobile appeBook on depressionDepression, AnxietyPROMIS Emotional Distress-Depression44Significant reduction in depression and anxiety in Woebot vs. controlKocur (2021)54GermanyOpen-label34Clinical (inpatients)≈ 37CAT-DB (CBT + avatar)Embodied 3D avatarHuman-controlled (Faceware)Face-to-faceLaptop, Unreal EngineTAUDepressionBDI-IIImmediately, 22CAT-DB + TAU significantly reduced BDI-II scores more than TAU at post-intervention and 2-week follow-up.Loveys (2022)72New ZealandOpen-label38Nonclinical(community, stressed women)43.2CBSMVirtual Human (Soul Machines)Multimodal AI (emotion recog, dialogue management)Digital human + userPC/tabletE-manualStressVAS22Large, significant reduction in state stress immediately after intervention; no group differences.Reilly (2024)57USASingle-blind42Clinical (veterans, chronic pain)53.7Acceptance and Commitment Therapy (ACT)Embodied avatar (Coach Anne)Automated, preprogrammed dialog3D avatar + userWeb (PC)Waitlist + Usual careDepressionPHQ-977Significant reduction in PHQ-9 scores in VACT-CP group by week 7; no improvement in control group.Jang (2021)34South KoreaOpen-label46Nonclinical (university students)25.1CBT-based communication interventionChatbot (AI-based counselor)NLP, deep learning, unspecified AI architectureText-basedMobile AppWaitlist / Self-help bookDepression, Anxiety, StressQIDS-SR, SAS, PSS44Significant reductions in QIDS-SR, SAS, and PSS scores in the intervention group; control group worsened on anxiety.Gan (2025)31ChinaSingle-blind55Clinical (patients scheduled for TKA surgery)72.7Educational + informed consent supportChatGPT (GPT-4)LLM-based, real-time, default parametersText-basedChatGPT web interfaceTraditional informed consent (surgeon-led, oral + written)Depression, AnxietyHADS-D, HADS-A, VAS-AImmediately, 11ChatGPT significantly reduced anxiety levels both before and 5 days after surgery (HADS-A, PAS-7, VAS-A).Fitzpatrick (2017)28USAUnclear56Nonclinical (college students)22.3CBTChatbot (Woebot, fully automated)Decision tree, scripted CBT, some NLP for routingText-basedFacebook MessengerE-book psychoeducation (NIMH ebook)Depression, AnxietyPHQ-9, GAD-7ImmediatelyNASignificant reduction in depression (PHQ-9) in Woebot group (d = 0.44); both groups improved on anxiety (GAD-7), no significant difference.de Graaff (2025)26JordanSingle-blind60Nonclinical(Young people)19.7CBTChatbot(Salam)Rule-based (non-AI)Text-basedWebEnhanced care as usual (psychoeducation)Depression, AnxietyHSCL-2588No significant differences between groups at postassessment.Kampmann (2016)53NetherlandsOpen-label60Clinical (social anxiety disorder)36.9VRET (exposure only)Virtual humans (avatars)Therapist-controlled, scriptedFace-to-face in VRLab, VR headsetWaitlistDepression, Anxiety, StressDASS-21Immediately, 1212Significant reduction in stress and anxiety at post-treatment; no effect on depression. At 3 months, only stress remained significantly lower.Pinto (2015)55USASingle-blind60Nonclinical(young adults)22Avatar-based (serious gaming, SBAR3 comm. skills)Embodied avatarPreprogrammed/branchingDesktop-based, immersiveLaptopScreen-based health ed.DepressionHADS1212eSMART-MH was feasible and safe; depression (HADS) decreased from 8.08 to 6.5 over 12 weeks.Zheng (2025)60ChinaOpen-label60Nonclinical(Pregnant women, SCH)30Metaverse platform + LT4Metaverse virtual clinicVR + AI-based interactionText/VR/videoVR, app, online platformStandard care (in-person)Depression, AnxietySDS, SAS66Metaverse group had significantly lower depression and anxiety scores at delivery and 6 weeks postpartum (SDS, SAS).Gardiner (2017)62USAUnclear61Nonclinical(women)33–37Mindfulness, CBT elements, lifestyleEmbodied agent (avatar)Scripted, adaptive dialogueText-basedWeb browserPatient info sheets + meditation CD/MP3depression, StressPHQ-9, PSS-444No significant differences in PHQ-9 or PSS-4 at 4 weeks; ECA group showed greater reduction in alcohol use and increased fruit intake.MacNeill (2024)39CanadaOpen-label68clinical(Arthritis/diabetes)42.87CBT/DBT/MI (Wysa)Chatbot (Wysa)ProprietaryText-basedSmartphone appWaitlistDepression, Anxiety, StressPHQ-9, GAD-7, PSS-102, 42, 4Significant reductions in depression and anxiety in the intervention group only; no significant change in stress for either group.Fulmer (2018)30USASingle-blind74Nonclinical (college students)~23Integrative (CBT, ACT, MI, etc.)Chatbot (Tess)NLP rule-based + supervised MLText-basedMessenger, SMS, SlackInfo-only (NIMH eBook)Depression, AnxietyPHQ-9, GAD-72, 42, 4Significant reduction in depression (PHQ-9) and anxiety (GAD-7) in intervention groupsKarkosz (2024)35PolandOpen-label81Nonclinical (young adults, subclinical)25CBTChatbot (Fido)ML/NLP (Emplocity)Text-basedFacebook MessengerSelf-help book (Mind Over Mood)Depression, AnxietyPHQ-9, STAI (trait), CESD-R44Both groups improved; no significant difference between arms.Johnson (2024)52AustraliaOpen-label82Nonclinical (university students)20.7CBT for perfectionismAI chatbot (ChatGPT)LLM, Q&ASelf-guidedOnline, PDF + AISelf-help onlyDepression, Anxiety, StressDASS-21, GAD-7, PHQ-9Immediately, 44Both arms improved; no significant differenceLiu (2022)38ChinaOpen-label83Nonclinical (University students)23.08CBT-based (RASA-powered)Chatbot (XiaoNan)RASA, NLU, ML, template dialogue managementText/voiceWeChat appBibliotherapy (CBT-based self-help book)Depression, AnxietyPHQ-9, GAD-71616Chatbot group showed significantly greater reduction in PHQ-9 (large effect, d = 0.83) and GAD-7 (modest effect, d = 0.30).Lee (2025)68South KoreaSingle-blind85Nonclinical (university students with procrastination, without severe depression/anxiety)20.4CBT-based, semigenerative modelChatbot (Moa)Korean BERT + GPT2, custom NLPText-basedMobile (Android)App only (no chatbot)StressPSS4, 84, 8Significant reduction in stress (PSS) in chatbot group only (mean change −3.59, p < .001); no significant change in control.He (2022)33ChinaSingle-blind98Nonclinical (college students)18.78CBTChatbot (XiaoE)NLP + Deep Learning (RASA framework)Text, Image, VoiceWeChat Official AccountE-book (psychoeducation), general chatbot (Xiaoai)DepressionPHQ-91, 41, 4Significant reduction in PHQ-9 in XiaoE group at post (d = 0.51, moderate) and follow-up (d = 0.31, small).Kelly (2025)63UKOpen-label98Clinical (Type 1/2 diabetes)53Biopsychosocial, person-centered, AI-enabledAI triage & mapping platform (Spotlight-AQ)NLP + ML (BERT, Random Forest)Hybrid (self + clinician)Digital/remote + in-personUsual careDepression, AnxietyPHQ-2, GAD-72424PHQ-2, GAD-7 reduced in intervention, moderate effect, not significant.Chen (2025)24PolandUnclear103Nonclinical ((Community-based, nonclinical sample))NRGeneral support/informationChatbot (text-based)Natural language processing (NLP)Text-basedWhatsApp/Web-basedNurse hotlineDepression, AnxietyPHQ-9, GAD-7ImmediatelyNASignificant reduction in anxiety scores after chatbot use compared to baseline.Spytska (2025)65UkraineOpen-label104clinical (Women with anxiety disorder, war zone)18-60 yrsCBT, MI (crisis support)Chatbot (Friend)NLP, ML, sentiment analysis, deep learning (adaptive)Text-basedMobile app/TelegramConventional psychotherapyAnxietyHamilton Anxiety Scale, Beck Anxiety InventoryImmediatelyNABoth groups showed significant anxiety reduction; therapy more effective (45–50%) than chatbot (30–35%, p < .01), but chatbot still effective and scalable for crisis use.Bennion (2020)64United KingdomUnclear112Nonclinical ((Participants were older adults from the community))69.21Method of Levels (MYLO), Rogerian counseling (ELIZA)Chatbot (MYLO and ELIZA)Natural language processing (MYLO, ELIZA)Text-basedWeb-basedELIZA chatbot (digital comparator)Depression, anxiety, stressDASS-21, Problem distress scale (Likert)22Both agents significantly reduced all outcomes; MYLO showed significantly lower distress at follow-up.Chan (2024)25Hong KongSingle-blind129Nonclinical (general adult sample)34.09Cognitive Behavioral Therapy for Insomnia (CBTi)Chatbot-based coachingLogic-based preprogrammed conversational turnsText-basedMobile app (Sleep Sensei)Digital sleep hygiene and self-monitoring (dSH)Depression, AnxietyPHQ-9, GAD-7Immediately, 44Significant reduction in depressive and anxiety symptoms compared to control.Ulrich & Lienhard(2024)47SwitzerlandOpen-label140Nonclinical (University students)26.7CBT, mindfulness, psychoeducationChatbot (MISHA)Rule-based, MobileCoach platformText-basedSmartphone appWaitlistDepression, Anxiety, StressPHQ-9, GAD-7, PSS-10ImmediatelyNASignificant reductions in PHQ-9 (Δ = -1.95, p = .003, d = 0.50) and PSS-10 (Δ = -4.17, d = 0.60) in IG; no significant changes in CG. GAD-7 reduction in IG not significant.Yasukawa (2024)48JapanOpen-label143Nonclinical (Employed adults, subthreshold depression)41.4iCBT, adherence supportChatbotRule-based, progress-triggered, >300 msg databaseText-basediCBT app + LINEiCBT aloneDepression, AnxietyPHQ-9, GAD-788No significant differences in PHQ-9 and GAD-7 reductions. IG had higher completion rate (34.8% vs. 19.2%, p = .036).Suharwardy(2023)45USAOpen-label152Nonclinical (Postpartum women)34CBT, IPTChatbotWoebot (NLP, rules-based, proprietary, perinatal content)Text-basedSmartphone appUsual careDepression, AnxietyPHQ-9, GAD-766PHQ-9 reduction significantly greater in IG (p = .025); no significant group difference in GAD-7.Prochaska(2021)56USAOpen-label152Clinical (adults with substance misuse)40CBT, MI, mindfulnessChatbot (Woebot)Automated (AI), text-basedText-basedSmartphone appWaitlistDepression, AnxietyPHQ-8, GAD-788W-SUDs significantly reduced substance use; depression and anxiety improved similarly in both groups.Aguilera(2024)49USAOpen-label168clinical (Adults with diabetes & depression)49Motivational, behavioral activationSMS chatbotReinforcement learning, adaptive text (Thompson sampling bandit)Text-basedCustom mobile app + SMSControl (weekly mood) + random textDepressionPHQ-82424Adaptive arm showed greater increase in daily steps ( + 19%) vs. random ( + 1.6%); PHQ-8 scores reported at baseline only.Klos(2021)37ArgentinaUnclear181Nonclinical (college students)18–33CBT, emotion-focused, solution-focused, MIChatbot (Tess)Proprietary NLP (X2AI)Text-basedFacebook MessengerPsychoeducation e-bookDepression, AnxietyPHQ-9, GAD-788No between-group difference; experimental group showed significant anxiety reduction (d = 0.5), no change in depression; control showed no change.Ulrich(2024)46Switzerland, Germany, AustriaOpen-label198clinical (Adults with frequent headaches (migraine, TTH, etc.))38.7CBT, behavior change, mindfulnessChatbotMobileCoach (decision-tree, text-based, BCTs)Text-basedSmartphone appWaitlistDepression, Anxiety, StressPHQ-9, GAD-7, PSS-10ImmediatelyNAIG showed significantly greater improvements across all outcomes. Medium effect size for PHQ-9 and PSS-10 (d = 0.43), large for GAD-7 (d = 0.66).Six(2025)59USAOpen-label209Nonclinical (College students)19.97Brief CBTECA/chatbot (4 arms)Speech/text (AI), animationApp+speech/textMobile appAll arms active appDepression, StressPHQ-8, PSS-1022Slight reductions in PHQ-8 and PSS-10 in all groups; no significant group differences; animation and conversation had no added benefit.Chua(2025)61SingaporeSingle-blind234Nonclinical(pregnant women and their partners from the community)32.2Psychoeducation, mindfulness, peer supportChatbot (Parentbot, text-based, rule-based)Rule-based logicText-basedMobile app (PDA)Usual careDepression, Anxiety, StressEPDS, STAI, PSS-1088Chatbot use was linked to higher depression, but not significantly. Educational material lowered anxiety. No significant relationships between app use and depression, anxiety, or stress at 1 or 3 months.Peuters(2024)43BelgiumOpen-label245Nonclinical(Adolescents)13.6Healthy lifestyle (HAPA, BCTs, self-regulation, gamification)ChatbotScripted, rule-basedText-basedSmartphone (mHealth app + Fitbit)WaitlistDepression, StressKIDSCREEN, PROMISImmediatelyNAMood improved only in IG with normal in-school education; worsened with remote. No change in CG. No significant effect on depressive symptoms (β = 0.58, p = .50).Sabour(2023)44ChinaOpen-label247Nonclinical(Community adults)~31CBT + Emotional SupportChatbotTemplate-based, NLP/ML, Chinese LLMText-basedWeChatWaitlistDepression, AnxietyPHQ-9, GAD-7ImmediatelyNASignificant reduction in PHQ-9 in CBT-Bot group vs. control. Anxiety decreased over time in all groups; no significant between-group difference.Garety(2024)51UKSingle-blind345Clinical (psychosis)~39AVATAR therapyEmbodied avatarScripted/adaptive, therapist-guidedFace-to-face digitalCustom softwareUsual careDepression, Anxiety, StressDASS, BDI16, 2816, 28Both groups improved by week 16; AV-BRF had slightly stronger, more consistent effects. By week 28, only AV-BRF retained some benefits. Effect sizes were small to moderate and declined over time.Maciejewski(2023)70PolandOpen-label372Nonclinical(University students)20.98Self-efficacy-basedChatbot (Stressbot)Rule-based chatbotText-basedFacebook Messenger appWaitlistStressPSS-444Small reduction in stress at post-test in chatbot group vs. control (d = -0.33); effect not present at follow-up.Kleinau(2024)36MalawiSingle-blind511Nonclinical (health workers)Majority <30 yCBT + Positive PsychologyChatbot (Vitalk, avatar Viki)Rule-based, Ruby/Javascript (no NLU for Malawi version)Text-basedAndroid app (messenger-style)Web-based mental health resource links (waitlist, passive)Depression, Anxiety, StressPHQ-9, GAD-7, OLBI88Depression significantly reduced in chatbot group (DiD = -0.68). Anxiety and burnout also reduced, but less clearly (95% CI for anxiety includes 0).Fitzsimmons-Craft(2022)29USAUnclear700Nonclinical (high-risk women for ED)21.08CBT-based psychoeducationChatbot (Tessa), rule-based, scriptedRule-based, scripted contentText-basedSMS or Facebook MessengerWaitlistDepression, AnxietyPHQ-9, GAD-712, 2412, 24No significant group difference in depression and anxiety, but both groups improved from baseline.Maeda(2020)67JapanOpen-label927Nonclinical(women 20-34)28.8Educational/Health promotionChatbotScripted (Dialogflow, NLP)Text-basedWeb browserPDF (fertility booklet), PDF (unrelated topic)AnxietySTAIImmediatelyNAChatbot group had significantly lower post-test anxiety than both controls (all P < 0.001).Filippo(2023)27South AfricaUnclear19,643Nonclinical(general population (young women))21.2Behaviorally-informed psychoeducation (not classic CBT); edutainment, digital health promotionChatbot (gamified and narrative versions)Automated, scripted, user-adaptive feedback; not a generative AI but an interactive decision tree/dialogue systemText-basedMobile app(WhatsApp)Pure control (no chatbot); attention control (information-only chatbot)DepressionPHQ-9ImmediatelyNAIntervention group showed significantly lower depression (PHQ-9) scores compared to waitlist.Abbreviations for Sample type: IG – Intervention Group; CG – Control GroupAbbreviations for Therapeutic approach: CBT – Cognitive Behavioral Therapy; ACT – Acceptance and Commitment Therapy; DBT – Dialectical Behavior Therapy; MI – Motivational Interviewing; CBSM – Cognitive Behavioral Stress Management; BCT – Behavior Change Technique; BCI – Behavior Change Intervention; TAU – Treatment as UsualAbbreviations for CA type / AI framework: CA – Conversational Agent; AI – Artificial Intelligence; LLM – Large Language Model; ECA – Embodied Conversational Agent; DSS – Decision Support System; ML – Machine Learning; NLP – Natural Language Processing; SMS – Short Message Service; App – Application; VR – Virtual Reality; VRET – Virtual Reality Exposure TherapyAbbreviations for Delivery platform/Interaction mode: App – Application; SMS – Short Message Service; VR – Virtual Reality; TAU – Treatment as UsualAbbreviations for Measurement tools/Psychological outcomes: PHQ-9 – Patient Health Questionnaire-9; PHQ-8 – Patient Health Questionnaire-8; GAD-7 – Generalized Anxiety Disorder-7; QIDS-SR – Quick Inventory of Depressive Symptomatology – Self-Report; PSS – Perceived Stress Scale; HADS – Hospital Anxiety and Depression Scale; BDI-II – Beck Depression Inventory-II; DASS – Depression Anxiety Stress Scales; HSCL-25 – Hopkins Symptom Checklist-25; STAI – State-Trait Anxiety Inventory; VAS – Visual Analog Scale Major characteristics of studies included in the systematic review Abbreviations for Sample type: IG – Intervention Group; CG – Control Group Abbreviations for Therapeutic approach: CBT – Cognitive Behavioral Therapy; ACT – Acceptance and Commitment Therapy; DBT – Dialectical Behavior Therapy; MI – Motivational Interviewing; CBSM – Cognitive Behavioral Stress Management; BCT – Behavior Change Technique; BCI – Behavior Change Intervention; TAU – Treatment as Usual Abbreviations for CA type / AI framework: CA – Conversational Agent; AI – Artificial Intelligence; LLM – Large Language Model; ECA – Embodied Conversational Agent; DSS – Decision Support System; ML – Machine Learning; NLP – Natural Language Processing; SMS – Short Message Service; App – Application; VR – Virtual Reality; VRET – Virtual Reality Exposure Therapy Abbreviations for Delivery platform/Interaction mode: App – Application; SMS – Short Message Service; VR – Virtual Reality; TAU – Treatment as Usual Abbreviations for Measurement tools/Psychological outcomes: PHQ-9 – Patient Health Questionnaire-9; PHQ-8 – Patient Health Questionnaire-8; GAD-7 – Generalized Anxiety Disorder-7; QIDS-SR – Quick Inventory of Depressive Symptomatology – Self-Report; PSS – Perceived Stress Scale; HADS – Hospital Anxiety and Depression Scale; BDI-II – Beck Depression Inventory-II; DASS – Depression Anxiety Stress Scales; HSCL-25 – Hopkins Symptom Checklist-25; STAI – State-Trait Anxiety Inventory; VAS – Visual Analog Scale Risk of Bias Assessment The majority of included studies were assessed as having a low risk of bias in domains related to the randomization process (n = 34, 79.1%), deviations from intended interventions (n = 41, 95.3%), missing outcome data (n = 35, 81.4%), and measurement of the outcome (n = 40, 93.0%). In contrast, concerns were more frequently identified in the domain of selection of the reported result, where 12 studies (27.9%) were rated as having some concerns. Overall, 25 studies (58.1%) were judged to have a low risk of bias, 11 studies (25.6%) had some concerns, and 7 studies (16.3%) were rated as having a high overall risk of bias. Most high-risk judgments were driven by issues related to missing outcome data and selective reporting, often due to incomplete reporting or lack of pre-specified analysis plans. These findings indicate that while a substantial proportion of studies demonstrated acceptable methodological rigor, a non-trivial subset exhibited potential sources of bias, warranting cautious interpretation of the pooled estimates (Fig. 3 and Supplementary Fig. 1). Summary of Risk of Bias Assessments Across Included Randomized Controlled Trials (RoB 2 Tool). Summary plot presenting the risk of bias judgments for included randomized controlled trials based on the Cochrane Risk of Bias 2 (RoB 2) tool. Domains assessed include bias arising from the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Each study is categorized as low risk, some concerns, or high risk of bias across domains, with the overall risk of bias judgment presented for each trial. Certainty of Evidence Evidence certainty ranged from moderate for depression and stress outcomes to low for anxiety outcomes, based on the GRADE approach (Supplementary Table 1). Depression and stress showed consistent effects across studies with low heterogeneity, supporting a moderate level of confidence. However, the certainty for anxiety was downgraded due to potential publication bias and imprecision in some subgroup estimates. Results of meta-analysis A total of 31 RCTs(53 reports)24–29,31–36,38–41,44–48,50–59,66 assessed depression outcomes and were included in the meta-analysis. Conversational agent-based AI interventions showed a significant small-to-moderate effect in reducing depressive symptoms compared to controls (SMD = −0.27; 95% CI: −0.32 to −0.21; p < 0.001). Heterogeneity was low (I² = 11.2%), indicating consistent effects across diverse populations and settings (Fig. 4). Most studies used validated tools like PHQ-9, BDI-II, and DASS-21, with intervention durations ranging from brief to multi-week formats. Forest Plot of the Impact of Conversational Agent-Based Digital Interventions on Depression Symptoms. Forest plot displaying the pooled standardized mean differences (SMDs) and 95% confidence intervals (CIs) for the effect of conversational agent-based digital interventions on depressive symptoms across included studies. Individual study effects are represented by squares, with the size of each square reflecting study weight. Horizontal lines indicate 95% CIs, and the diamond represents the overall pooled effect estimate. A negative effect size indicates a reduction in depressive symptoms in the intervention group compared with controls. To assess robustness, a robust variance estimation (RVE) analysis was conducted. The pooled effect size remained consistent (SMD = −0.28; 95% CI: −0.38 to −0.19), confirming the stability of the main findings. A multilevel meta-analysis using restricted maximum likelihood (REML) accounted for dependency among multiple effect sizes and repeated measurements. The model included 53 effect sizes from 32 studies and showed a significant overall effect (SMD = −0.28; SE = 0.04; z = −6.79; p < 0.001; 95% CI [ − 0.36, −0.20]). Between-study heterogeneity was small (σ² = 0.0104), while within-study (time-level) variance was negligible, indicating limited contribution of repeated measures to overall variability. The multilevel model produced estimates highly consistent with the conventional random-effects model (SMD = −0.28 vs. −0.27), supporting robustness to dependency structure assumptions. Across different platforms (web, mobile) and conversational agent types (chatbots, avatars, and LLM-based systems), interventions showed consistent and statistically significant effects on mental health outcomes. The standardized mean differences (SMDs) ranged from -0.249 to -0.272 across studies, indicating small-to-moderate effects. Sensitivity analyses confirmed the robustness of these findings, as shown in Supplementary Fig. 2. Subgroup analyses were conducted to explore potential sources of heterogeneity. The results are presented in Supplementary Table 2. No statistically significant differences between subgroups were observed for most variables, including blinding, age, sample size, sample type, therapeutic approach, conversational agent type, and outcome timing (all p-values for subgroup differences > 0.05). However, subgroup differences by country were statistically significant (p = 0.772 overall heterogeneity test across categories indicates no meaningful between-group differences despite variation in effect sizes). Although effect sizes varied across countries (e.g., China: SMD = −0.45; USA: SMD = −0.11), the test for subgroup differences did not indicate statistically significant interaction effects. Meta-regression found that outcome timing was not a significant moderator (β = 0.079; p = 0.199), with residual heterogeneity (I²_res = 8.6%) and 21.9% of between-study variance explained (Supplementary Fig. 5). Egger’s test indicated no significant small-study effects (p = 0.217), and the funnel plot was symmetrical, suggesting low publication bias (Supplementary Fig. 8). A total of 23 RCTs (38 reports)24–26,29,31,34–39,41,44–47,51–53,56,58,66 assessed anxiety outcomes and were included in the meta-analysis. Conversational agent-based AI interventions demonstrated a significant small effect in reducing anxiety symptoms compared to control conditions (SMD = −0.20; 95% CI: −0.26 to −0.14; p < 0.001). Heterogeneity was low (I² = 2.8%), indicating consistent results across studies (Fig. 5). Validated instruments such as the GAD-7, STAI, and DASS-21 were commonly used, with intervention durations ranging from a few sessions to several weeks. Forest Plot of the Impact of Conversational Agent-Based Digital Interventions on Anxiety Symptoms. Forest plot presenting the pooled standardized mean differences (SMDs) and 95% confidence intervals (CIs) for the effect of conversational agent-based digital interventions on anxiety symptoms across included studies. Individual study effect sizes are shown as squares, with the size of each square reflecting the study weight. Horizontal lines indicate 95% CIs, and the diamond represents the overall pooled effect estimate. Negative effect sizes indicate a reduction in anxiety symptoms in the intervention group compared with controls. RVE analysis confirmed these findings, yielding a slightly attenuated but significant effect (SMD = −0.17; 95% CI: −0.26 to −0.09), consistent with the primary model. A multilevel meta-analysis (38 effect sizes from 23 studies) showed a significant overall effect (SMD = −0.21; SE = 0.03; z = −5.39; p < 0.001; 95% CI [ − 0.28, −0.13]). Variance components indicated low heterogeneity at both study (σ² = 0.0030) and time levels (σ² = 0.0023). The multilevel model closely aligned with the random-effects estimate (SMD = −0.21 vs. −0.20), confirming robustness. Effects were observed across delivery formats (e.g., mobile apps, chat interfaces, web platforms) and conversational agent types, including rule-based chatbots, embodied agents, and generative AI systems. Sensitivity analysis showed stability of the effect size (range: −0.185 to −0.213), confirming the robustness of the findings (Supplementary Fig. 3). Subgroup analyses were conducted to examine potential sources of heterogeneity in anxiety outcomes (Supplementary Table 3). Overall, the results did not indicate statistically significant differences between subgroups for most examined variables. No significant subgroup differences were observed for blinding status (p = 0.096), sample size (p > 0.45), age (p > 0.67), sample type (p > 0.76), therapeutic approach (p > 0.44), conversational agent type (p > 0.38), or outcome timing (p > 0.42). Similarly, no statistically significant differences were found across country subgroups (p > 0.23), although effect sizes varied across regions (e.g., China: WMD = −0.27; USA: WMD = −0.10). Overall, these findings suggest that the effectiveness of conversational agent-based interventions on anxiety is relatively consistent across the examined subgroups, with no evidence of statistically significant effect modification. However, meta-regression examining outcome timing revealed only a marginal trend toward weaker effects with longer follow-up (β = 0.007; p = 0.074). The model explained 27.3% of between-study variance, with negligible residual heterogeneity (I²_res = 0%) (Supplementary Fig. 6). Egger’s test indicated potential small-study effects (p = 0.012), supported by an asymmetric funnel plot, suggesting possible publication bias (Supplementary Fig. 9). A total of 14 RCTs (23 reports)34,36,39,46,47,51–54,59,66,68,69,71 assessed stress outcomes and were included in the meta-analysis. Conversational agent-based AI interventions were associated with a small but significant reduction in stress compared to controls (SMD = −0.26; 95% CI: −0.37 to −0.16; p < 0.001). Heterogeneity was moderate (I² = 31.4%), indicating some variability in effects across studies (Fig. 6). Validated instruments such as the Perceived Stress Scale (PSS), DASS-21, and OLBI were most commonly used, and intervention durations varied from single-session digital contacts to multi-week programs. Forest Plot of the Impact of Conversational Agent-Based Digital Interventions on Stress Symptoms. Forest plot presenting the pooled standardized mean differences (SMDs) and 95% confidence intervals (CIs) for the effect of conversational agent-based digital interventions on stress symptoms across included studies. Individual study effect sizes are displayed as squares, with the size of each square proportional to study weight. Horizontal lines indicate 95% CIs, and the diamond represents the overall pooled effect estimate. Negative effect sizes indicate a reduction in stress symptoms in the intervention group compared with controls. RVE analysis confirmed the robustness of findings (SMD = −0.25; 95% CI: −0.38 to −0.12), consistent with the primary model. A multilevel meta-analysis including 23 effect sizes from 15 studies showed a significant effect (SMD = −0.26; SE = 0.05; z = −4.56; p < 0.001; 95% CI [ − 0.36, −0.14]). Both between-study and within-study variances were small (σ² = 0.0077 for both levels), indicating modest contributions of clustering effects. The multilevel model matched the random-effects model (SMD = −0.26 vs. −0.26), supporting robustness. The intervention effect was consistent across different platforms and CA types (text-based chatbots, avatars, and hybrid models). Sensitivity analysis confirmed the robustness of findings, with effect sizes stable across studies (range: −0.24 to −0.28) (Supplementary Fig. 4). Subgroup analyses were conducted to explore potential sources of heterogeneity in stress outcomes (Supplementary Table 4). Overall, no statistically significant differences between subgroups were observed across the examined variables. The tests for subgroup differences were not statistically significant for blinding status (p values ranging from 0.264 to 0.096), sample size (p > 0.17), age (p > 0.29), sample type (p > 0.28), therapeutic approach (p > 0.11), conversational agent type (p > 0.10), or outcome timing (p > 0.13). Although some variation in effect sizes was observed across subgroups (e.g., larger effects in nonclinical populations and older age groups), these differences did not reach statistical significance. These findings suggest that the effectiveness of conversational agent-based interventions on stress is consistent across the examined subgroups, with no evidence of statistically significant effect modification. Egger’s test did not indicate significant small-study effects (p = 0.079), and the funnel plot appeared symmetrical, suggesting a low risk of publication bias (Supplementary Fig. 10). Discussion This meta-analysis and systematic review pooled data from randomized controlled trials that evaluated the effectiveness of conversational agent (CA) interventions in symptom reduction of depression, anxiety, and stress. In general, our evidence suggests that these digital interventions have small to moderate beneficial effects on psychological measures, with meta-analytic standardized mean differences (SMD) of −0.27 for depression, −0.20 for anxiety, and −0.26 for stress. Such effects were also continuously monitored using validated instruments such as the PHQ-9, GAD-7, and PSS, demonstrating a robust and clinically relevant pattern of psychological change. Our findings are consistent with previous meta-analyses of internet and mobile-delivered cognitive behavioral therapy (CBT) which observed similar effect sizes73,74. The present findings expand the existing evidence base by demonstrating that both AI-based and rule-based conversational agents serve as promising tools for alleviating common mental health symptoms. These results align with prior research highlighting the effectiveness of digital mental health interventions across diverse delivery formats14,15,75. Several mechanisms may explain the observed benefits. The majority of interventions employed evidence-based psychotherapeutic approaches, such as CBT and Acceptance and Commitment Therapy (ACT), which likely drive therapeutic success independent of the underlying software architecture. Moreover, conversational agents offer a scalable, accessible, and anonymous channel for support. This modality can enhance user engagement, decrease perceived stigma, and facilitate compliance—especially among demographic groups hesitant to seek face-to-face care7,10,28. The interactive, dialogical interface mimics therapeutic dialogue, and can further induce user reflection, the processing of emotions, and the modification of maladaptive behaviors. Subgroup analyses indicated some variation in effect sizes across population subgroups, with relatively larger point estimates observed in nonclinical groups and among older adults; however, these differences were not statistically significant. This pattern may partly reflect differences in baseline characteristics such as comorbidity burden and motivation for self-care, which previous literature suggests influence engagement with digital health interventions30,39,76. In addition, interventions incorporating mobile applications and more advanced AI-based systems—particularly those employing generative models—tended to show larger effect estimates compared to non-dynamic, rule-based systems. These systems may enable more adaptive, personalized, and human-like interactions, which could enhance user engagement, perceived usability, and the therapeutic alliance77. However, given that the subgroup differences did not reach statistical significance, these observations should be interpreted cautiously as exploratory rather than confirmatory. Clinically, the findings are consistent with the potential of conversational agents to expand the access of mental health services, especially in environments with workforce shortages or where stigma and access hurdles remain. CAs can be used as cost-efficient, low-intensity interventions for symptom management, early intervention, or adjunctive treatment—highly valuable in times of crisis such as the COVID-19 pandemic, when access to face-to-face services is restricted78. Clinically, the findings highlight the potential of conversational agents to expand access to mental health services, especially in environments with workforce shortages or where stigma and access hurdles persist. CAs can be utilized as cost-efficient, low-intensity interventions for symptom management, early intervention, or adjunctive treatment—highly valuable during periods of crisis, such as the COVID-19 pandemic, when access to conventional, in-person mental health services was restricted78. Several limitations must be considered when interpreting our findings. Our exclusion of dissertations, conference abstracts, unpublished manuscripts, and restriction to English-language publications may have introduced publication bias and limited the comprehensiveness of our review. While many studies demonstrated a low risk of bias, approximately one-third were rated as having some concerns or a high risk of bias, predominantly due to missing data, selective reporting, and a lack of participant or assessor blinding. Variability in conversational agent type, delivery platform, therapeutic content, and intervention duration limits the generalizability and precision of our effect estimates. Most studies relied on self-reported outcomes and lacked long-term follow-up assessments, reducing confidence in the sustainability of the therapeutic effects. Finally, few studies evaluated adverse events, long-term outcomes, or user trust, which are critical determinants of real-world viability and effectiveness. Importantly, even though we have examined both AI-based and non-AI-based interventions, the rapid evolution of generative AI necessitates further study of its relative impact, risks, and ethics. Future research must look beyond basic efficacy to bolster the evidence base and support practical application. The question of whether these conversational agents work in a vacuum is not the only one. To determine the feasibility of deploying these tools at scale—particularly in resource-constrained settings—researchers should conduct formal cost-effectiveness analyses. Furthermore, it is essential to investigate the human-AI interaction dynamics. Engagement and clinical outcomes are significantly influenced by trust, empathy, and user preferences regarding human involvement in care. Qualitative research is also needed to explore user experiences, expectations, and the factors that promote or inhibit sustained use. Ultimately, to ensure that conversational agents deliver lasting public health benefits, an integration of mixed methods and real-world perspectives is required. Otherwise, even the most promising tech risks falling flat outside the laboratory. In conclusion, both rule-based and AI-based conversational agent interventions show statistically significant reductions in depression, anxiety, and stress. Although these reductions are small to moderate and are not significantly different from routine face-to-face interventions (and in some cases show smaller effect sizes), they remain promising given the scalability and accessibility of these technologies. With thoughtful, evidence-based development and implementation, conversational agents can make a valuable contribution to global mental health efforts. Methods Search Strategy The PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines were followed in the conduct of this systematic review. Supplementary Table 5 contains a completed PRISMA checklist. Three significant electronic databases—PubMed, Scopus, and Web of Science—were thoroughly searched in order to find relevant studies. The initial search was performed on May 8, 2025 and Only randomized controlled trials (RCTs) published in English between January 2010 and May 2025 were included in the search, as this is when conversational agent and artificial intelligence technologies have become more prevalent in digital mental health interventions. The search strategy was created to find research assessing AI or conversational agent-based digital interventions that address stress, anxiety, or depression symptoms. Keywords and Medical Subject Headings (MeSH) terms pertaining to conversational agents, chatbots, artificial intelligence, digital mental health, and psychological outcomes were combined. Supplementary Table 6 provides comprehensive search strategies for each database. To identify additional relevant studies beyond those indexed in peer-reviewed databases, grey literature was systematically searched using Google Scholar. We screened the first 200 records returned by keyword searches sorted by relevance, using combinations of terms such as “chatbot”, “conversational agent”, “AI”, “mental health”, “depression”, “anxiety”, “stress”, and “RCT”. Filters were applied to include only English-language and full-text accessible studies. We also examined the reference lists of all included full-text articles and pertinent systematic reviews (retrieved from database searches) to identify additional eligible studies. This backward snowballing approach was complemented by forward citation tracking using Google Scholar, allowing us to include studies that cited any of the included articles. In addition, when required data for effect size calculation were missing or unclear, attempts were made to contact the corresponding authors via email (up to two attempts). All retrieved citations were imported into EndNote 20, and duplicates were removed prior to screening. The review protocol (ID: CRD420251059269) was registered in PROSPERO. Inclusion and Exclusion Criteria Studies were included in the systematic review and meta-analysis if they met the following criteria: Population: Included participants of any age, gender, or clinical status (clinical or nonclinical), without restrictions on diagnostic category or severity. Population: Included participants of any age, gender, or clinical status (clinical or nonclinical), without restrictions on diagnostic category or severity. Intervention: Evaluated a CA-based interventions delivering structured psychological support aimed at reducing symptoms of depression, anxiety, or stress, including:AI-based conversational agents, defined as systems utilizing artificial intelligence (e.g., natural language processing, machine learning, or large language models) to generate adaptive and context-aware responses;Non-AI/rule-based conversational agents, defined as pre-programmed systems following scripted or rule-based interaction flows without adaptive learning capabilities. Intervention: Evaluated a CA-based interventions delivering structured psychological support aimed at reducing symptoms of depression, anxiety, or stress, including:AI-based conversational agents, defined as systems utilizing artificial intelligence (e.g., natural language processing, machine learning, or large language models) to generate adaptive and context-aware responses;Non-AI/rule-based conversational agents, defined as pre-programmed systems following scripted or rule-based interaction flows without adaptive learning capabilities. AI-based conversational agents, defined as systems utilizing artificial intelligence (e.g., natural language processing, machine learning, or large language models) to generate adaptive and context-aware responses; AI-based conversational agents, defined as systems utilizing artificial intelligence (e.g., natural language processing, machine learning, or large language models) to generate adaptive and context-aware responses; Non-AI/rule-based conversational agents, defined as pre-programmed systems following scripted or rule-based interaction flows without adaptive learning capabilities. Non-AI/rule-based conversational agents, defined as pre-programmed systems following scripted or rule-based interaction flows without adaptive learning capabilities. Comparator: Employed a control or comparison group, such as treatment-as-usual, waitlist, or another active intervention. Comparator: Employed a control or comparison group, such as treatment-as-usual, waitlist, or another active intervention. Outcomes: Reported quantitative outcomes related to depression, anxiety, or stress, using validated psychological instruments, including the Patient Health Questionnaire-9 (PHQ-9), Generalized Anxiety Disorder-7 (GAD-7), Depression Anxiety Stress Scales-21 (DASS-21), and Perceived Stress Scale (PSS). Outcomes: Reported quantitative outcomes related to depression, anxiety, or stress, using validated psychological instruments, including the Patient Health Questionnaire-9 (PHQ-9), Generalized Anxiety Disorder-7 (GAD-7), Depression Anxiety Stress Scales-21 (DASS-21), and Perceived Stress Scale (PSS). Study Design: Used a RCT design. Study Design: Used a RCT design. Reporting: Provided sufficient data (e.g., pre- and post-intervention means and standard deviations, or equivalent statistics) to enable calculation of standardized effect sizes. Reporting: Provided sufficient data (e.g., pre- and post-intervention means and standard deviations, or equivalent statistics) to enable calculation of standardized effect sizes. Publication: Published in English in a peer-reviewed scientific journal. Publication: Published in English in a peer-reviewed scientific journal. The following criteria were used to exclude studies: (1) had a qualitative design, which is not appropriate for meta-analysis; (2) did not have a control or comparator group; (3) did not report sufficient outcome data or statistical information needed to estimate effect sizes; (4) were protocols, conference abstracts, reviews, or commentaries that did not include primary data; (5) involved interventions that did not use a conversational agent interface (e.g., video-only interventions or static psychoeducation). Selection of Studies All references retrieved from the database searches were imported into EndNote 20 for reference management and duplicate removal. Title and abstract screening was conducted independently by two reviewers (A.M. and M.A.) to identify potentially eligible studies. The full texts of selected articles were then retrieved and assessed independently by the same reviewers according to the predefined eligibility criteria. Discrepancies at any stage were resolved through discussion, and when consensus could not be reached, a third reviewer (A.N.) was consulted. All reviewers had prior experience in conducting systematic reviews and meta-analyses in the field of mental health research. Data Extraction Data extraction was performed independently by two reviewers (A.M. and M.A.) using a standardized data extraction form. The following study features were sourced from the data extracted: first author name, year of publication, country, sample size, type of population (clinical vs. nonclinical), participants’ mean and range of age, method of blinding, conversational agent type used (rule-based vs. AI-based), therapeutic approach type used (e.g., CBT, Acceptance and Commitment Therapy (ACT)), delivery platform (e.g., web, mobile app), method of response generation (i.e., the mechanism by which the conversational agent generates responses, including rule-based/scripted systems, retrieval-based approaches, and generative artificial intelligence–based models), control condition, outcome measures (e.g., GAD-7, PHQ-9, PSS), timing of post-intervention assessment, and findings reported for depression, anxiety, and stress. Means and standard deviations (SDs) were derived for baseline and post-intervention in both the intervention and control groups. Whenever studies had reported standard errors (SEs), confidence intervals (CIs), or interquartile ranges (IQRs) instead of SDs, respective conversion formulas were applied to calculate SDs so that there was uniformity in effect size calculation. Any discrepancies were resolved through discussion, and where necessary, by consultation with a third reviewer (A.N.). Risk of Bias Assessment The risk of bias of included studies was independently assessed by two reviewers(E.H.S. and M.A.) using the Cochrane Risk of Bias 2 (RoB-2) tool79. Random sequence generation, allocation concealment, participant and staff blinding, outcome assessment blinding, incomplete outcome data, and selective reporting are among the domains that are evaluated by this tool. “Low,” “high,” or “some concerns” were assigned to each domain. Any differences of opinion were discussed and decided by consensus. Disagreements were resolved through discussion or consultation with a third reviewer to ensure consistency and reliability of assessments. Certainty Assessment (GRADE) Two reviewers (A.N. and M.A.) independently applied the GRADE approach to assess the certainty of evidence for each pooled outcome, based on risk of bias, inconsistency, indirectness, imprecision, and publication bias. Any disagreements were resolved through discussion. Statistical Analysis Effect sizes were calculated as Standardized Mean Differences (SMDs) using mean change from baseline for both intervention and control groups80. This approach was specifically chosen to account for baseline imbalances observed in several included trials and to provide a more precise estimate of the therapeutic effect by removing between-person variation (Cochrane Handbook, Section 6.5.1.1)81. If S.D was not reported in each of individual study; it was calculated using following formula: S.Dchange = square root [(S.Dbaseline2 + S.Dfinal2)-(2 × R × S.Dbaseline × S.Dfinal)]. A correlation coefficient of 0.8 was considered as R-value of the above-mentioned formula82. Where standard error of mean (S.E.M) was only reported, standard deviation (S.D) was calculated by S.D = S.E.M × √n (n is the number of participants in each group). Finally, When data were only presented in graphical formats, figures were digitized using WebPlotDigitizer (version 4.6) to extract numerical values. The extraction was performed independently by two reviewers, with discrepancies resolved through consensus. For studies where data required for effect size calculation were missing or unclear, we attempted to contact the corresponding authors via email up to two times to request the necessary statistical information83,84. Estimates of effect sizes were expressed as standardized mean differences (SMDs) with 95% confidence intervals (CIs) using a random-effects model. A random-effects model was selected as the primary analytical approach to account for potential between-study variability arising from differences in populations, interventions, and measurement instruments across studies. This model assumes that the true effect size may vary between studies and incorporates both within-study and between-study variance in the weighting process. Studies were weighted using the inverse of their variance, with additional adjustment for between-study heterogeneity. In addition to the random-effects model, a Robust Variance Estimation (RVE) approach was also applied to account for potential dependencies within studies. The RVE method, performed using the robumeta command in Stata, considered the correlation between effect sizes within studies (with an assumed correlation of 0.8). This method provides more accurate standard errors for the estimated effect sizes by adjusting for dependencies between effect sizes within studies. The RVE analysis was conducted to assess the robustness of the results and confirm that the findings were consistent across different methodological approaches. Furthermore, to explicitly account for the hierarchical structure of the data and the dependency introduced by multiple time points within the same trial, we conducted a three-level multilevel meta-analysis using restricted maximum likelihood (REML) estimation. This model partitioned the total variance into between-study variance and within-study (time-level) variance, enabling us to assess the robustness of the pooled estimates and determine whether repeated measurements significantly impacted overall heterogeneity. Pre-defined subgroup analysis based on country, sample size, sample type, age, therapeutic approach, CA type, and outcome timing was conducted to detect potential sources of heterogeneity. The sensitivity analysis was conducted by the one-study remove (leave-one-out) approach, to explore the impact of each study on the pooled effect size. Meta-regression was performed to examine whether the timing of outcome assessment moderated the observed effect sizes. Timing of outcome assessment was defined as the duration (in weeks) between the end of the intervention and the measurement of outcomes. When multiple follow-up time points were reported within a study (e.g., 4 weeks, 8 weeks, 16 weeks), each time point was extracted and included as a separate effect size. Publication bias was evaluated using visual assessment of funnel plots and Egger’s weighted regression tests. All statistical analyses were performed using Stata software version 21 (StataCorp. College Station, Texas, USA). P < 0.05 was considered as statistically significant. Supplementary information Supplementary Information Acknowledgements Not applicable. The authors received no financial support for the research, authorship, or publication of this article. Author contributions A.N. and E.H.S. conceptualized the study and designed the search strategy. A.M. and M.A. screened studies for eligibility, extracted data, and assessed risk of bias. A.N. and S.M.M.A conducted the statistical analysis. A.M., E.H.S., and M.A. drafted the manuscript. A.N. critically revised the manuscript. All authors contributed to the final version and approved the submitted manuscript. Data Availability All data generated or analyzed during this study are included in this published article and its supplementary files. Additional data are available from the corresponding author upon reasonable request. Competing interests The authors declare no competing interests. Ethical Approval This study is a systematic review and meta-analysis of previously published studies. No new data involving human participants were collected, and therefore ethical approval and informed consent were not required. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41746-026-02820-1. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement All data generated or analyzed during this study are included in this published article and its supplementary files. Additional data are available from the corresponding author upon reasonable request.