Artificial intelligence and academic anxiety? a meta-analysis Authors: Weihao Wang, Long Diao, Jingting Yu Published: 2026 ABSTRACT AbstractBackgroundAcademic anxiety negatively impacts students' performance and well-being. Although AI technologies are increasingly used in education, findings on their effectiveness in reducing academic anxiety remain mixed.ObjectiveThis meta-analysis quantified the overall effect of AI interventions on academic anxiety and identified potential moderators.MethodsA systematic search identified 22 studies with 30 independent effect sizes. Robust variance estimation with random-effects models was used. Subgroup analyses and Q-tests were conducted.ResultsAI interventions showed a significant moderate overall effect on reducing academic anxiety (g = − 0.65, 95% CI [− 1.07, − 0.22], p < .01). Moderator analyses revealed: (a) test anxiety and general academic anxiety had significant effects, whereas mathematics, foreign language, and learning anxiety did not; (b) intelligent tutoring systems and robotics produced significant effects, but chatbots and virtual technology did not; (c) primary and middle school effects were marginally significant, university effects non‑significant; (d) Eastern cultural context showed a significant effect, Western and Islamic contexts did not; (e) effects sustained from immediate to delayed posttests; (f) post‑2022 studies yielded larger effects than earlier studies (marginally significant difference); (g) smaller‑sample studies tended to produce larger effect sizes. However, Q‑tests for most moderators were non‑significant (p > 0.05), indicating observed differences are descriptive rather than statistically robust.ConclusionAI interventions significantly reduce academic anxiety with a moderate effect size. Although efficacy varies descriptively across several factors, most subgroup differences are not statistically significant. Future research should employ rigorous experimental designs, longer follow‑ups, and meta‑analytic structural equation modeling to test underlying mechanisms.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40359-026-05018-y.Keywords: Artificial Intelligence, Virtual Technology, Chatbot, Intelligent Tutoring Systems, Academic Anxiety Introduction Artificial intelligence and academic anxiety Academic anxiety represents a pervasive challenge for students globally, exerting significant negative effects on academic performance, psychological health, and overall well-being [2, 41, 52, 66]. The rapid advancement of artificial intelligence (AI) has introduced transformative opportunities within educational contexts. Through sophisticated human–computer interaction capabilities, AI has fundamentally restructured pedagogical approaches and curricular frameworks [35, 44, 55, 75]. By providing personalized instructional guidance, immediate feedback mechanisms, and adaptive learning resources, AI applications enable students to adapt to learning environments and master content more effectively, thereby enhancing academic achievement and reducing academic anxiety [47, 89]. Despite preliminary evidence supporting AI's potential for academic anxiety reduction, inconsistencies across research findings necessitate further investigation into its effectiveness and underlying mechanisms. The present meta-analysis The present investigation employs meta-analysis methodology to systematically synthesize empirical research on AI's role in alleviating academic anxiety, establishing a more robust evidential foundation. This systematic evaluation quantifies the effect size of AI interventions on academic anxiety reduction while identifying potential moderating variables. Furthermore, the analysis examines variations in AI's effectiveness across diverse educational contexts and student populations. Through comprehensive analytical procedures, this research aims to provide empirical support for AI implementation in educational settings and establish directional guidance for future pedagogical practices and scholarly inquiry. Literature review AI AI has become integrated into daily life, fundamentally transforming human interaction patterns, learning methodologies, and lifestyle habits [16]. The conceptual foundations of AI emerged from computer science and engineering disciplines, while drawing significant influence from philosophy, cognitive science, neuroscience, and economics [96]. The interdisciplinary character of this field has resulted in a lack of consensus regarding the precise definition of AI among researchers. The term AI carries different connotations across disciplinary contexts, warranting a brief survey of how it has been operationalized in recent scholarship. Chassignol et al. [11] characterized AI as a computer science branch designed to enable machines to address complex problems associated with human intelligence, including content learning, problem-solving, and pattern recognition. According to their framework, AI encompasses numerous technical domains such as computer vision, speech processing, big data analytics, and natural language processing. Chiu et al. [16] defined AI as the capacity of digital systems to simulate intelligent entities in task performance, incorporating multiple technological branches including computer vision, speech processing, machine learning, big data analytics, and natural language processing. Jiao et al. [43] conceptualized AI as a discipline focused on developing theoretical frameworks and application systems that simulate and extend human intelligence, with the objective of enabling machines to replicate intelligent human behaviors such as learning, reasoning, thinking, and planning, thereby equipping machines with intelligent capabilities for complex task execution. Integrating these perspectives, the present study conceptualizes AI as an advanced computational technology with cross-disciplinary applications spanning machine learning, natural language processing, computer vision, and related fields. The fundamental objective of AI involves simulating, extending, and augmenting human cognitive functions through computational systems, enabling computers to execute sophisticated tasks requiring understanding, reasoning, learning, planning, self-correction, and environmental perception. AI for education Education represents one of the domains most significantly impacted by AI applications [31]. AI technologies have fundamentally transformed pedagogical approaches employed by educators and learning methodologies adopted by students. Through the implementation of personalized learning trajectories, automated assessment mechanisms, and interactive instructional resources, AI innovations significantly augment the capabilities of both educators and learners throughout the educational process [82]. A structured understanding of AI's deployment in educational settings requires a taxonomic overview of its modalities, as different application types—personalized tutoring, collaborative platforms, immersive simulations—carry distinct implications for intervention design. During its nascent stages, AI within educational contexts predominantly comprised intelligent assistance systems designed to enhance user performance through automated mechanisms [32]. Contemporary implementations, however, exhibit considerably greater diversification across educational environments. From a functional perspective, AI educational applications can be categorized into three distinct modalities: personalized tutoring systems providing individualized instructional guidance; intelligent collaborative learning platforms facilitating effective communication and cooperation among students; and immersive virtual reality technologies creating authentic learning environments through sophisticated simulation techniques [53]. Through comprehensive content analysis of 125 scholarly publications, Wang et al. [82, 83] proposed a four-category taxonomy of educational AI applications: adaptive learning and personalized tutoring tools; intelligent assessment and management systems; profiling and prediction frameworks; and emerging technological innovations. When classified according to intended beneficiaries, AI in education can be segmented into three categories: learner-oriented tools, teacher-oriented tools, and system-oriented tools, each addressing the specific requirements of different educational stakeholders [4]. The foregoing taxonomies further reveal a critical, yet underexplored dimension—the potential for different AI modalities to exert distinct psychological effects on learners. Tool categories vary fundamentally in their interaction paradigms, ranging from the cognitive scaffolding of Intelligent Tutoring Systems to the conversational dynamics of Chatbots, the physical co-presence of Robotics, and the immersive simulation of Virtual Technologies [16, 82, 83]. Consequently, different AI modalities are likely to shape learners' experiences across multiple psychological dimensions—including cognitive, motivational, and affective outcomes—through distinct mechanisms and with varying magnitudes, a pattern that holds particular relevance for domain-specific concerns such as academic anxiety. Nevertheless, relevant differences among AI modalities have not been fully examined in current review literature. Given the diversity of AI modalities and their potentially distinct psychological effects, it is instructive to examine how prior review literature has approached the landscape of AI in education. This body of work falls into two broad categories: comprehensive overviews charting the field's evolution and challenges, and focused reviews targeting specific educational sectors or technological modalities—with the gap between them pointing to the need for the present study. A substantial body of research has systematically analyzed the evolutionary trajectory, contemporary landscape, and persistent challenges associated with AI applications across broad educational domains [11, 16, 32, 82, 98]. In contrast, another stream of scholarly inquiry has concentrated on AI implementations within specific educational sectors or disciplines, as well as the efficacy of diverse technological interventions in educational settings. These specialized investigations encompass higher education environments [37, 96], STEM education initiatives [92], conversational agent technologies [62], intelligent tutoring systems [46, 55], and virtual reality applications [83]. Meta-analytic findings on AI in education present a nuanced picture. While AI interventions consistently yield positive effects, the magnitude varies widely across developmental indicators and demographic groups, underscoring the need to identify the sources of this heterogeneity. Wu and Yu [89] conducted a comprehensive analysis of 24 studies, revealing a substantial overall effect size (d = 0.964) of AI chatbots on students' learning outcomes, indicating pronounced educational benefits. In contrast, Hwang's [40] meta-analytical investigation specifically examining AI applications in mathematics education demonstrated a considerably smaller effect size on primary school students' mathematical achievement. Similarly, Zheng et al. [101] synthesized findings from 24 empirical studies and identified a modest effect size of 0.208 for AI interventions on learning perception. Y. Wang et al. [83] determined that virtual reality technology exerts a moderate positive influence on students' creative capabilities. Collectively, these empirical findings suggest that while AI generally contributes positively to various dimensions of student development, the magnitude of these effects exhibits considerable variability across specific developmental indicators and demographic populations. Despite the growing body of meta-analytic evidence on AI in education, no review has systematically examined AI interventions specifically targeting academic anxiety—the central concern of the present study. A thorough summary and analysis of existing studies are needed to understand the magnitude of these effects and the factors that might influence them. Academic anxiety Academic anxiety refers to feelings of nervousness and worry specifically tied to academic contexts [22, 56], encompassing multiple interrelated domains such as math anxiety, test anxiety, and writing anxiety [65]. Unlike generalized anxiety, it is contextually bound to educational settings and manifests through cognitive, affective, and physiological responses to performance expectations and evaluation situations [9]. The construct has been operationalized through validated instruments including the Academic Emotions Questionnaire anxiety subscales, which assess anxiety in class, during learning, and in test situations [93], and the Academic Anxiety Inventory, which captures domain-specific academic anxieties [65]. Academic anxiety is prevalent globally, with particularly elevated rates among Asian student populations [88], and demonstrates robust negative associations with achievement, self-efficacy, and intrinsic motivation [22, 59]. Notably, academic anxiety manifests across diverse academic situations including homework, examinations, and major assignments [57]. In the present study, academic anxiety is operationalized as self-reported anxious affect directly linked to academic performance contexts, encompassing both subject-specific forms and general academic worry, consistent with the outcome measures across the reviewed intervention studies. Anxiety carries profound clinical and educational significance, making it a pressing target for intervention research. Anxiety represents the most prevalent mental health disorder affecting adolescent populations [71]. In severe cases, untreated anxiety may culminate in suicidal ideation or behavior [15, 29, 77]. Accordingly, while documenting anxiety's deleterious impact is essential, identifying efficacious anxiety-reduction strategies for this population constitutes a more pressing imperative—one that the present study directly addresses. Several studies have examined country- or region-level differences in the effects of AI on students (e.g., [87, 95, 99]). Although some research has identified cross-national variation (e.g., [25, 99], the large number of countries involved—most of which contributed only one or two studies—makes it difficult to systematically explain the sources of heterogeneity in effect sizes across nations using the available data. Cultural background may offer a broader and more theoretically meaningful lens for examining this variation. Previous research has indicated that Confucian heritage cultures students report higher absolute levels of academic anxiety than students from other cultural backgrounds [26, 73]. However, whether cultural background moderates the effectiveness of AI interventions—that is, whether students from different cultural contexts show differential reductions in academic anxiety following AI-based interventions—remains an open question. It should be noted that the present meta-analysis does not aim to compare absolute levels of academic anxiety across cultural backgrounds. Rather, it examines whether cultural background moderates the effect size of AI interventions on academic anxiety. This distinction is important: even if baseline levels of anxiety differ across cultural contexts, the effects of AI interventions on reducing anxiety may be similar or may diverge. Given the severity of academic anxiety documented above, it is encouraging that cumulative meta-analytic evidence shows that targeted interventions effectively alleviate anxiety in academic and clinical groups, laying a solid empirical basis for this study. Zhang and Aslan [98] conducted a meta-analysis synthesizing 32 experimental studies, demonstrating that physical exercise interventions significantly attenuate anxiety symptoms among university students. Similarly, Ersozlu [28] established that game-based learning approaches effectively reduce math anxiety in educational contexts. Furthermore, Xiong et al. [91] reported in their meta-analysis that targeted language learning strategy interventions produce significant reductions in foreign language anxiety, yielding a moderate effect size. However, existing meta-analytic reviews have predominantly focused on conventional intervention modalities—physical exercise, game-based learning, and language learning strategies—without systematically examining the growing body of research on technology-driven approaches to anxiety reduction. This gap warrants particular attention in light of the rapid proliferation of AI-based tools in educational practice. AI for academic anxiety Although many studies have looked at ways to reduce anxiety, there are very few systematic reviews that focus on how AI can help with academic anxiety, even as AI technology is rapidly advancing. Because of this, we need a thorough evaluation that combines findings on the effectiveness of AI interventions and the factors that might influence them. This would help improve evidence-based methods for supporting students' mental health in schools. Three distinct perspectives emerge from the empirical literature on AI and academic anxiety, collectively demonstrating that the relationship is neither uniformly positive nor negative but contingent on contextual and methodological factors. The predominant view, supported by numerous empirical investigations [5, 23,30*, 34*], suggests that AI interventions significantly reduce academic anxiety levels. In a notable study, Tai and Chen [76*] implemented a two-week Google Assistant-mediated language learning protocol with 112 eighth-grade students. Their findings revealed a substantial reduction in foreign language anxiety accompanied by enhanced communication willingness among participants. Similarly, Vanbecelaere [79*] conducted a three-week experimental investigation with 84 participants, demonstrating that AI-based adaptive gaming platforms effectively diminished math anxiety among primary school students while concurrently enhancing learning efficiency. These adaptive systems facilitated individualized learning experiences calibrated to students' proficiency levels, thereby fostering increased motivation and self-efficacy. However, the evidence is not unidirectional, and acknowledging contradictory findings is essential for a rigorous assessment of AI's potential—particularly for identifying boundary conditions that the present study must consider. Shazly [72*] executed an 8-week experimental protocol with 48 Egyptian university students, documenting that while AI technologies demonstrated potential for enhancing linguistic competencies, they simultaneously induced slight elevations in foreign language anxiety. Cakmak [8*] examined 90 university students, requiring the experimental cohort to engage in chatbot-facilitated oral practice over a 12-week period. Despite exhibiting superior oral performance compared to their counterparts who participated in face-to-face practice sessions, the experimental group manifested significantly heightened foreign language anxiety levels. Beyond this binary picture, a third line of inquiry suggests that AI's effects on academic anxiety are dimension-dependent rather than uniform, underscoring the need for multidimensional outcome measures. Zhang [97*] employed chatbot interventions juxtaposed against traditional teacher support as a control condition, wherein both groups participated in five 45-min reading sessions. The longitudinal analysis revealed that chatbot interventions significantly reduced overall foreign language reading anxiety within the experimental group from pre-test to post-test measurements. However, between-group comparisons at post-test revealed no significant differences in foreign language reading anxiety. While many studies have explored ways to reduce anxiety, few have focused on how AI can help with academic anxiety, even as AI technology quickly improves. A detailed review is needed to understand how effective AI tools are and what factors affect their success. This would help improve proven methods for supporting students' mental health in schools. Method This study adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [63]. The methodological procedures were conducted in accordance with the recommendations outlined in the Cochrane Handbook for Systematic Reviews of Interventions [36], and were informed by methodological approaches from similar previously published meta-analyses [18, 85, 100, 103*]. Literature review procedure Studies for inclusion in this meta-analysis were identified through systematic literature searches across three major electronic databases: Web of Science, Scopus, and ProQuest. For the Web of Science database, we employed the following advanced Boolean search string: "(TI = (AI) OR TI = (artificial neural network*) OR TI = (machine intelligence) OR TI = (robot*) OR TI = (machine learning OR deep learning*) OR TI = (artificial intelligence) OR TI = (deep learning) OR TI = (chat*) OR TI = (intelligent tutor*) OR TI = (automated tutor*) OR TI = (neural network*) OR TI = (expert system) OR TI = (intelligent system) OR TI = (virtual learning) OR TI = (natural language processing)) AND (TI = (anxiety))". After completing the database search and screening, we identified nine directly relevant studies. However, we found that the number of empirical studies was still limited. To further improve literature coverage, we adopted a two-stage backward snowballing method. All studies identified through the snowballing method were retrieved and obtained in full text via Google Scholar. Duplicate checks were performed at each stage against the already included studies to ensure no redundant inclusion. The first stage was a meta-analysis-based snowballing method. We reviewed published meta-analyses focusing on the topic of "artificial intelligence and student development" (e.g., [83]). From the reference lists of these meta-analyses, we extracted original empirical studies related to artificial intelligence and academic anxiety, yielding a total of six studies. The second stage was a snowballing method based on close reading of the literature. For the nine studies obtained from the database search and the six studies obtained from the first-stage snowballing, we carefully examined the reference list of each study and further screened out seven additional studies that met the inclusion criteria. Ultimately, a total of 22 studies were included in this meta-analysis, nine directly from the database search and 13 supplemented by the snowballing method. Criteria for inclusion Studies were required to meet all of the following criteria for inclusion in this meta-analysis: (1) The study must be empirical research, utilizing rigorous research designs to collect and analyze quantitative data. (2) The research must be conducted within formal educational settings. (3) AI technology must be explicitly identified and implemented as an intervention component in the study. (4) The study must include validated measures of academic anxiety or its specific forms (math anxiety, test anxiety, foreign language anxiety, etc.). (5) Full-text articles must be accessible through electronic databases or university library systems. (6) The study must employ either controlled experimental designs, quasi-experimental designs, or single-arm intervention studies. (7) The study must report sufficient statistical information to enable effect size calculation. (8) Participants must be enrolled students at any Learning environment. Following the application of these criteria, 22 peer-reviewed journal articles reporting 30 effect sizes were included in the meta-analysis. The systematic screening process is illustrated in Fig. 1. Flow chart of the inclusion and exclusion of studies Coding For the 22 studies included in this meta-analysis, we implemented a systematic coding protocol with two independent researchers who established coding reliability through collaborative online consensus meetings. Each researcher independently examined and coded all articles, documenting extracted data in a standardized electronic spreadsheet. The comprehensive coding framework captured multiple variables: study identification, effect size metrics, publication year, sample characteristics, methodological attributes, AI type, anxiety domain specificity, learning environment, single-session intervention duration, post-test assessment timing, country, cultural background, and intervention duration measured in weeks. The brief coding structure and extracted data for all included studies are presented in Table 1. Full details including primary purpose of the intervention, interaction modality, anxiety measure and multiple effect sizes are provided in Appendix A. Table 1The descriptive information of individual studiesStudy NamePYSS(E;C)AIRDATLECIDSIDPATCBAlemi et al. (2015) [1*]201546(30;16)RBRFAMIR590IISBeketov et al. (2024) [5]2024246(123;123)ITSRGAURU8120IWEÇakmak (2022) [8*]202289CBSFAUTU1210IISChen &Hwang (2022) [12*]−1202249(28;21)VTQFAUJP8100DEAChen &Hwang (2022) [12*]−2202244(26;18)VTQFAUJP8100DEAChen &Hwang (2022) [12*]−3202293(54;39)VTQFAUJP8100DEAChen (2019) [13*]−1201941(22;19)VTRMAPCN640IEAChen (2019) [13*]−2201941(18;23)VTRMAPCN640IEAChu et al. (2021) [17*]2021116(58;58)ITSQMAPCN180IEACobbett &Snelgrove (2016) [20*]201655(27;28)VTRFAUCA45IWEDarmawansah & Hwang (2024) [23]202458(28;30)RBQLAPCN4DEAGao (2024) [30*]2024160(80;80)CBRTAUCN16240DEAHan (2020) [34*]202044(22;22)CBQFAMSK1020IEAHwang et al. (2020) [39*]−12020109(53;56)ITSQMAPCN1120IEAHwang et al. (2020) [39*]−22020109(53;56)ITSQMAPCN1120IEAKlos et al. (2021) [45*]2021181(99;82)CBRGAUAR8WELiu et al. (2022) [51*]202263(33;30)CBRGAUCN16EARezai (2023) [67*]-1202354(28;26)ITSRFAUIR960IISRezai (2023) [67*]-2202354(28;26)ITSRFAUIR960DISRodero &Larrea (2022) [68*]−12022100(50;50)VTRFAUESIWERodero &Larrea (2022) [68*]−22022100(50;50)VTRFAUESIWERodero &Larrea (2022) [68*]−32022100(50;50)VTRFAUESIWERong et al. (2022) [69*]202231VTSTAMCN420IEAShazly (2021) [72*]202138CBSFAUEG8100DISTai & Chen (2023) [76*]2023112ITSSFAMCN245IEAVanbecelaere et al. (2021) [79*]202166(33;33)ITSRMAPBE315IWEWang et al. (2021) [84*]−1202166(31;35)VTQLAUCN320IEAWang et al. (2021) [84*]−2202166(31;35)VTQLAUCN320IEAZhang et al. (2024) [97*]2024131(131;130)ITSRFAUCN650DEAZheng (2024) [102]202484(42;42)CBQFAMCN45IEAPY Publication Year, SS(E;C) Sample Size(Experimental Group; Control Group), AI AITypology, ITS Intelligent Tutoring Systems, CB Chatbots, RB Robotics, VT Virtual Technology, RD Research Design, R Experimental Study, Q Quasi-Experimental Study, S Single-Arm Study AT Types of Anxiety, MA Math Anxiety, LA Learning Anxiety, FA Foreign Language Anxiety, TA Test Anxiety, GA Generalized Anxiety, LE Learning Environment, P Primary School, M Middle School, U University, C Country, CN China, SK South Korea, CN China, IR Iran, RU Russia, AR Argentina, BE Belgium, TU Turkey, EG Egypt, CA Canada, ES Spain, JP Japan, ID Intervention Duration, SID Single-session Intervention Duration, PAT Post-test Assessment Timing, I Immediate Post-intervention Assessment, D Delayed Follow-up Assessment, CB Cultural Background, EA Eastern Cultural Background, WE Western Cultural Background, IS Islamic Cultural Background The descriptive information of individual studies PY Publication Year, SS(E;C) Sample Size(Experimental Group; Control Group), AI AITypology, ITS Intelligent Tutoring Systems, CB Chatbots, RB Robotics, VT Virtual Technology, RD Research Design, R Experimental Study, Q Quasi-Experimental Study, S Single-Arm Study AT Types of Anxiety, MA Math Anxiety, LA Learning Anxiety, FA Foreign Language Anxiety, TA Test Anxiety, GA Generalized Anxiety, LE Learning Environment, P Primary School, M Middle School, U University, C Country, CN China, SK South Korea, CN China, IR Iran, RU Russia, AR Argentina, BE Belgium, TU Turkey, EG Egypt, CA Canada, ES Spain, JP Japan, ID Intervention Duration, SID Single-session Intervention Duration, PAT Post-test Assessment Timing, I Immediate Post-intervention Assessment, D Delayed Follow-up Assessment, CB Cultural Background, EA Eastern Cultural Background, WE Western Cultural Background, IS Islamic Cultural Background Following initial coding, we assessed interrater reliability between the two coders using Cohen's Kappa coefficient [19]. The interrater reliability was high across all categorical variables: k = 0.90 for research design, k = 0.84 for academic anxiety type, and k = 1.00 for Learning environment. For continuous variables, we calculated Cronbach's α, which yielded perfect reliability (α = 1.00) for publication year, total sample size, experimental group sample size, and control group sample size, and near-perfect reliability (α = 0.99) for intervention duration. Any coding discrepancies between researchers were resolved through online discussion until consensus was reached. Risk of bias assessment To assess the risk of bias in the included randomized controlled trials, we applied the revised Cochrane Risk of Bias 2.0 (RoB 2.0) tool. This tool evaluates each study across five core domains: (1) randomization process, (2) deviations from intended interventions, (3) missing outcome data, (4) measurement of the outcome, and (5) selection of the reported result. Each domain was independently rated by two reviewers as "low risk", "some concerns", or "high risk", with discrepancies resolved through discussion or consultation with a third reviewer. An overall risk-of-bias judgment for each study was then derived, where "low risk" required all domains to be rated low risk, "some concerns" indicated at least one domain with some concerns but no high-risk ratings, and "high risk" was assigned if any domain was judged high risk. The overall domain-level and study-level risk-of-bias ratings for the 22 included studies are presented in Figs. 2 and 3. According to the summary plot, the domain of missing outcome data exhibited the highest methodological quality, with 86.4% of studies judged as low risk. In contrast, the randomization process showed the lowest proportion of low-risk judgments (50.0%), with 45.5% of studies rated as some concerns and 4.5% as high risk. The remaining domains yielded generally satisfactory quality: 77.3% low risk for deviations from intended interventions, 77.3% low risk for outcome measurement, and 68.2% low risk for selection of reported results. Risk of bias summary Risk of bias graph In terms of the overall study-level bias assessment, 40.9% (9/22) of the included studies were classified as low risk, 40.9% (9/22) as some concerns, and 18.2% (4/22) as high risk. Collectively, these results demonstrated that most eligible studies achieved acceptable methodological rigor, whereas a considerable proportion of trials presented notable limitations, primarily regarding inadequate reporting of randomization procedures and selective outcome reporting. Statistical analyses Data analysis was performed using the R statistical environment with the metafor and dmetar packages. A three-level meta-analytic model was employed to synthesize quantitative findings across independent studies, systematically investigating the effects of AI-assisted instruction on students' academic anxiety, while further identifying potential moderators and evaluating result robustness and publication bias. Hedges' g was adopted as the standardized effect size metric, representing the standardized mean difference in academic anxiety levels between groups. Negative effect sizes indicated lower academic anxiety in the AI intervention group relative to the control group. According to conventional benchmarks, Hedges' g values of approximately 0.2, 0.5, and 0.8 correspond to small, moderate, and large intervention effects, respectively [21]. A three-level random-effects meta-analytic model was fitted to accommodate the nested data structure in which multiple effect sizes were extracted from a single study. Model parameters were estimated using restricted maximum likelihood, and robust variance estimation was applied to correct standard errors, thereby enhancing the statistical robustness of the pooled estimates. To systematically explore sources of heterogeneity in the anxiety-reducing effects of AI, subgroup moderator analyses were conducted. The following variables were examined: AI type, anxiety type, learning environment, cultural background, research design, post-test assessment timing, publication period, sample size, total intervention duration, and single-session intervention duration. Between-group heterogeneity tests were used to determine whether effect sizes differed significantly across subgroups, clarifying the influence of each variable on intervention outcomes. To assess the robustness of the findings, multiple sensitivity analyses were performed. First, stratified sensitivity analysis based on ROB2 risk-of-bias ratings compared pooled effect sizes across three conditions: the full sample, after excluding high-risk studies, and including only low-risk studies. Second, additional sensitivity analyses were conducted by excluding single-arm studies and removing extreme outlier effect sizes, with between-group comparisons evaluating the extent to which extreme cases and low-quality studies influenced the core conclusions. Finally, a comprehensive set of established methods was employed to evaluate publication bias. These included RVE-corrected Egger's regression, Duval and Tweedie's trim-and-fill procedure, Rosenthal's fail-safe N, and the Vevea and Hedges weight-function model. This multi-method approach assessed publication bias from multiple perspectives—bias presence, number of missing studies, result tolerance, and model-based adjustment—ensuring the reliability and rigor of the meta-analytic conclusions. 3.6 Computation of effect sizes Effect sizes were extracted from the included studies. For experimental and quasi-experimental studies, the effect size was expressed as Cohen's d, calculated as the mean difference between the experimental and control groups divided by the pooled standard deviation (Pooled SD), as shown in Eq. (1). and denote the post-test mean scores of the target variable (e.g., academic anxiety, math anxiety) for the experimental and control groups, respectively. and represent the post-test standard deviations of the experimental and control groups, and and indicate the numbers of participants in each group. For studies employing a single-arm pre-post design, the effect size was transformed into a standardized mean difference (Cohen's d) compatible with independent-group meta-analyses. First, the paired standardized mean difference (StdPairedDiff) was computed based on the paired t-test statistic () and the paired sample size (n), as shown in in Eq. (2). Following the default convention in CMA 3.0, a correlation coefficient of R = 0.50 was applied to convert the paired effect size into a Cohen's d comparable to independent-group designs, as shown in Eq. (3). Following the default convention in CMA 3.0, a correlation coefficient of R = 0.50 was applied to convert the paired effect size into a Cohen's d comparable to independent-group designs, as shown in Eq. (3). Finally, following [7*], Hedges' g was adopted to correct for small-sample bias. Cohen's d values were accordingly converted to Hedges' g using Eq. (4). Result This meta-analysis examined 30 independent effect sizes from 22 studies that met the final inclusion criteria to assess the efficacy of AI in alleviating academic anxiety. Although the literature search spanned 2015–2024, the majority of included studies were published post-2020, with only four studies published prior to 2020. The methodological distribution comprised eleven experimental studies, seven quasi-experimental studies, and four single-arm studies. The research originated from eleven countries across five continents, with China contributing the largest proportion (n = 11), followed by studies from various countries in Asia, Europe, Africa, North America, and South America. Sample sizes ranged from 31 to 246 participants. Regarding learning environments, five studies were conducted in primary schools, five in middle schools, and twelve in university settings. For single-session interventions, durations ranged from 15 to 240 min. Foreign language anxiety emerged as the most frequently investigated domain, featuring in eleven studies. Analysis of effect sizes revealed 25 negative effects (indicating anxiety reduction) and 5 positive effects. The magnitude of effects ranged from a maximum positive effect of 1.13 to a minimum positive effect of 0.04, and from a maximum negative effect of −4.11 to a minimum negative effect of −0.06. The overall effect of AI on academic anxiety Results were reported based on a random-effects model, as illustrated in Fig. 4. After adjustment with robust variance estimation, the overall effect of AI-assisted instruction on academic anxiety was statistically significant (g = − 0.65, 95% CI [− 1.07, − 0.22], p < 0.01), indicating a moderate effect of AI-assisted instruction in alleviating students' academic anxiety. Forest plot A three-level meta-analytic model was employed to accommodate included studies that contributed multiple effect sizes. The heterogeneity decomposition revealed that the total variance in effect sizes primarily originated from between-study variance (Level 3: σ2 = 0.78, I2 = 79.5%), followed by within-study variance (Level 2: σ2 = 0.13, I2 = 13.8%), with sampling error contributing relatively little (Level 1: σ2 = 0.07). This indicates substantial between-study heterogeneity in the distribution of effect sizes across included studies, underscoring the need for further moderator analyses to explain the sources of this variability. The likelihood ratio test indicated that the three-level model did not provide a significantly better fit than the two-level model (χ2 = 2.18, df = 1, p = 0.140). Nevertheless, given that between-study heterogeneity accounted for 79.5% of the total variance, the three-level model was retained for subsequent analyses to more accurately disentangle the sources of variance across different levels. To examine the robustness of the results, a sensitivity analysis was conducted by excluding outlier effect sizes (|g|> 2.5; e.g., effect size id: [13*]–1; [76*]. After excluding these outliers, the overall pooled effect size changed from g = − 0.65 to g = − 0.47 (95% CI [− 0.78, − 0.16]), with no change in either the direction or statistical significance of the effect. The difference between the two estimates was not statistically significant (difference = 0.18, Z = 0.50, p > 0.05). These results indicate that the core conclusions of this meta-analysis were not driven by extreme values, and the anxiety-reducing effect of AI-assisted instruction is robust and reliable. Moderator analyses Moderator analyses were conducted based on AI typology, anxiety type, learning environment, cultural background, research design, post-test assessment timing, publication period, sample size category, total intervention duration, and single-session intervention duration. The results of subgroup moderator analyses are presented in Table 2. Table 2Moderator effectsVariableNgSE95CI_LL95CI_ULQpPublication Year4.070.07 Before 202222−0.360.21−0.770.040.08 After 20238−1.250.40−2.03−0.470.00AI Typology0.590.66 Chatbots7−0.330.33−0.980.330.33 Intelligent Tutoring Systems9−1.080.42−1.90−0.260.01 Robotics2−0.760.20−1.16−0.360.00 Virtual Technology12−0.500.43−1.340.350.25Research Design0.480.64 Quasi-Experimental Study11−0.420.15−0.72−0.110.01 Experimental Study15−0.810.29−1.37−0.250.00 Single-Arm Study4−0.630.92−2.431.160.49Types of Anxiety0.130.96 Foreign Language Anxiety16−0.590.38−1.330.140.12 Generalized Anxiety3−0.600.18−0.95−0.250.00 Learning Anxiety3−0.350.38−1.090.380.35 Math Anxiety6−0.970.63−2.200.260.12 Test Anxiety2−0.980.38−1.72−0.230.01Learning Environment1.090.38 Middle School5−1.200.55−2.27−0.130.03 Primary School7−0.930.50−1.920.060.06 University18−0.350.25−0.840.130.15Post-test Assessment Timing0.020.88 Delayed Follow-up Assessment8−0.810.28−1.35−0.260.00 Immediate Post-intervention Assessment20−0.700.30−1.28−0.110.02Cultural Background1.150.37 Eastern Cultural Background18−0.880.27−1.40−0.350.00 Islamic Cultural Background5−0.390.68−1.710.940.57 Western Cultural Background7−0.270.28−0.820.290.35Intervention Duration1.590.28 Less than 4 weeks9−0.790.41−1.590.010.05 5 to 8 weeks10−0.850.34−1.52−0.180.01 More than 8 weeks6−0.780.56−1.880.320.16Single-session Intervention Duration0.400.76 0 to 45 min11−0.740.49−1.710.220.13 46 to 90 min5−0.870.41−1.66−0.070.03 More than 90 min8−0.600.24−1.07−0.130.01Sample Size0.760.50 0 to 50 participants8−1.020.44−1.89−0.160.02 51 to 100 participants14−0.250.28−0.790.300.37 More than 100 participants8−0.980.39−1.74−0.220.01 Moderator effects By AI typology, intelligent tutoring systems (g = − 1.08, p = 0.01) and robotics (g = − 0.76, p < 0.01) significantly alleviated students’ academic anxiety, whereas no significant effects were observed for virtual technology (g = − 0.50, p = 0.25) or chatbots (g = − 0.33, p = 0.33). The heterogeneity test indicated no significant difference across AI typologies (Q = 0.59, p = 0.66). By anxiety type, AI interventions significantly reduced test anxiety (g = − 0.98, p = 0.01) and generalized anxiety (g = − 0.60, p < 0.01), while no significant effects were observed for math anxiety (g = − 0.97, p = 0.12), foreign language anxiety (g = − 0.59, p = 0.12), or learning anxiety (g = − 0.35, p = 0.35). The heterogeneity test revealed no significant difference across anxiety types (Q = 0.13, p = 0.96). By learning environment, the anxiety-reducing effects of AI were marginally significant for middle school students (g = − 1.20, p = 0.06) and primary school students (g = − 0.93, p = 0.06), but not significant for university students (g = − 0.35, p = 0.15). The heterogeneity test showed no significant difference across learning environments (Q = 1.09, p = 0.38). By cultural background, AI significantly reduced academic anxiety among students from Eastern cultural backgrounds (g = − 0.88, p < 0.01), while no significant effects were observed for Islamic cultural backgrounds (g = − 0.39, p = 0.57) or Western cultural backgrounds (g = − 0.27, p = 0.35). The heterogeneity test indicated no significant difference across cultural backgrounds (Q = 1.15, p = 0.37). By research design, AI significantly alleviated academic anxiety in experimental studies (g = − 0.81, p = 0.01) and quasi-experimental studies (g = − 0.42, p = 0.01), while no significant effect was observed in single-arm studies (g = − 0.63, p = 0.49). The heterogeneity test showed no significant difference across research designs (Q = 0.48, p = 0.64). By post-test assessment timing, AI significantly reduced academic anxiety at both immediate post-intervention assessment (g = − 0.70, p = 0.02) and delayed follow-up assessment (g = − 0.81, p < 0.01). The heterogeneity test indicated no significant difference between the two assessment timings (Q = 0.02, p = 0.88). By publication period, AI showed a marginally significant anxiety-reducing effect in studies published in 2022 and earlier (g = − 0.36, p = 0.08), and a statistically significant effect in studies published from 2023 onward (g = − 1.25, p < 0.01). The heterogeneity test revealed a marginally significant difference between the two publication periods (Q = 4.07, p = 0.07). By sample size category, AI significantly alleviated academic anxiety in studies with sample sizes of 50 or fewer participants (g = − 1.02, p = 0.02) and 51 to 100 participants (g = − 0.98, p = 0.01), whereas no significant effect was observed in studies with more than 100 participants (g = − 0.25, p = 0.37). The heterogeneity test indicated no significant difference across sample size categories (Q = 0.76, p = 0.50). By total intervention duration, AI significantly or marginally significantly reduced academic anxiety in studies with intervention durations of 4 weeks or less (g = − 0.79, p = 0.05) and 5 to 8 weeks (g = − 0.85, p = 0.01), whereas no significant effect was observed in studies with intervention durations exceeding 9 weeks (g = − 0.78, p = 0.16). The heterogeneity test showed no significant difference across intervention duration categories (Q = 1.59, p = 0.28). By single-session intervention duration, AI significantly reduced academic anxiety in studies with session durations of 46 to 90 min (g = − 0.87, p = 0.03) and more than 90 min (g = − 0.60, p = 0.01), while a marginally significant effect was found for sessions of 45 min or less (g = − 0.67, p = 0.05). The heterogeneity test indicated no significant difference across single-session intervention durations (Q = 0.40, p = 0.76). Sensitivity analysis Stratified sensitivity analyses were conducted based on ROB2 risk-of-bias ratings. Three-level meta-analyses were performed on three subsamples: the full sample (k = 30, n = 22, g = − 0.65, p < 0.01), after excluding high-risk studies (k = 26, n = 18, g = − 0.64, p < 0.01), and including only low-risk studies (k = 12, n = 9, g = − 0.63, p = 0.02). The overall heterogeneity test indicated no statistically significant difference among the three effect size estimates (Q = 0.00, p = 1.00). Pairwise comparisons further confirmed no significant differences between the full sample and the sample excluding high-risk studies (Z = − 0.01, p = 0.99), between the full sample and the low-risk only sample (Z = − 0.04, p = 0.97), or between the sample excluding high-risk studies and the low-risk only sample (Z = − 0.03, p = 0.98). A sensitivity analysis was also conducted by excluding single-arm studies to examine the robustness of the results. The pooled effect size for the full sample was g = − 0.65 (95% CI [− 1.05, − 0.24], p < 0.01, k = 30). After excluding single-arm studies, the pooled effect size was g = − 0.64 (95% CI [− 0.98, − 0.30], p < 0.01, k = 26). The magnitude, direction, and statistical significance of the effect size remained consistent before and after exclusion, with no significant difference between the two estimates. Taken together, the results of these sensitivity analyses indicate that the pooled effect size obtained in this meta-analysis is stable and reliable. The core findings are robust to variations in study-level risk of bias and research design type, supporting the credibility of the overall conclusions. Publication bias Multiple methods were employed to assess publication bias. The RVE-corrected Egger's regression test was not statistically significant (β = − 6.44, SE = 3.60, Z = − 1.79, p = 0.16), providing no evidence of publication bias in the present study. Duval and Tweedie's trim-and-fill [27] procedure further confirmed this finding, indicating that no missing studies needed to be imputed (estimated missing studies k₀ = 0), as shown in see Fig. 5. The original pooled effect size (g = − 0.65) was identical to the adjusted effect size, suggesting that publication bias was not present and that the conclusions were minimally influenced by unpublished literature. Rosenthal's [70] fail-safe N test indicated that at least 1,570 unpublished studies with null effects would be required to render the overall pooled effect non-significant, further corroborating the low risk of publication bias in this meta-analysis. Funnel plot Finally, the Vevea and Hedges weight-function selection model [80] was applied as a sensitivity analysis. The adjusted pooled effect size was g = 0.24 (95% CI [− 1.19, 1.67], p = 0.74), which was no longer statistically significant. Compared with the original effect size (g = − 0.65), the direction of the effect changed, with a magnitude of change of 0.885. It should be noted that the selection model requires a relatively large body of data to yield stable estimates. Given the limited sample size of the present study (22 studies, 30 effect sizes), the model estimates may be unstable, and the results of this analysis should therefore be interpreted with caution. Taken together, the majority of the evidence indicates that the conclusions of this meta-analysis are minimally affected by publication bias and demonstrate satisfactory robustness. However, given the uncertainty associated with the weight-function model results, caution is warranted when interpreting the overall findings. Discussion Academic anxiety represents a critical factor influencing students' academic performance and quality of life, a phenomenon that has been extensively investigated in educational research for decades [10, 21, 49, 61, 86]. With the integration of AI technologies into educational contexts, a growing body of literature has begun to examine whether and to what extent AI may contribute to both academic enhancement and anxiety reduction [5, 89, 94,30*, 45*, ]. However, findings remain mixed across studies, and the magnitude of AI's effects appears to vary considerably depending on the type of AI modality, intervention duration, and outcome measure. The present study conducted a meta-analysis synthesizing data from 22 studies with 30 independent effect sizes to examine the overall efficacy of AI interventions on academic anxiety and to identify potential moderating variables that may account for this observed heterogeneity. Overall effect of AI interventions on academic anxiety This meta-analysis found a significant moderate effect of AI interventions on reducing academic anxiety (g = −0.65, 95% CI [−1.07, −0.22], p < 0.01). However, efficacy varied across anxiety subtypes and AI modalities. These findings suggest that AI has the potential to alleviate specific types of academic anxiety under certain conditions. By managing positive academic emotions, AI may foster a more supportive learning environment through real-time analysis of students' behaviors and personalized learning recommendations. Guiding students with customized paces and approaches may support academic performance while addressing their achievement and self-esteem needs. Consequently, AI appears associated with increased positive academic emotions and reduced negative emotions—such as anxiety, shame, and helplessness—that typically arise from learning difficulties or maladjustment [50, 78, 90, 91]. Subgroup analyses by anxiety type showed significant effects for test and general academic anxiety, but non‑significant effects for mathematics, foreign language, and learning anxiety, with no significant between‑group heterogeneity (Q = 0.13, p = 0.96). The present findings may contribute to a better understanding of the potential benefits and limitations of incorporating AI technologies into educational settings, while also offering some extensions to the existing knowledge base in educational psychology and instructional research. The results are broadly consistent with several previous investigations [23, 51*, 97*], while also extending certain conclusions from others [89, 94, 101]. Notably, our study extends Wu and Yu's [89] meta-analysis, which reported a medium effect of AI chatbots on academic anxiety based on only three studies. By systematically compiling experimental studies on AI and academic anxiety, we provide more stable estimates that help address this methodological limitation. Although the present meta-analysis confirms the overall moderate efficacy of AI interventions for academic anxiety, the specific mechanisms underlying the heterogeneous effects across different anxiety subtypes and AI modalities warrant further explication. Drawing on the control-value theory of academic emotions [64], we propose that AI interventions may reduce anxiety through at least three pathways. First, they enhance perceived control by providing personalized, adaptive learning trajectories that match students’ proficiency levels. Second, they reduce the value threat associated with evaluation by offering private, low-stakes feedback environments. Third, they foster self-efficacy through immediate, mastery-oriented feedback. These mechanistic pathways are consistent with the moderator patterns observed in our subsequent analyses and provide a theoretical framework for interpreting the differential descriptive effects across anxiety subtypes documented in this review. It is also worth noting that consistent with our subgroup results, no statistically significant mitigating effect of AI on foreign language anxiety was observed. Several existing individual studies have reported negative or non-optimal AI effects on foreign language anxiety, which aligns with our non-significant pooled findings. Shazly [72*] reported that AI was associated with a slight increase in foreign language anxiety among university students, and Cakmak [8*] observed potentially adverse effects in a 12-week single-arm study with 90 students examining chatbot-assisted language learning. Although chatbots may offer novel opportunities for oral practice, interactions with these systems were linked to higher levels of second language speaking anxiety in this particular sample, which aligns with the present finding that AI did not demonstrate a statistically significant reduction in foreign language anxiety (g = − 0.59, p = 0.12). Both studies shared certain methodological characteristics: they employed single-arm experimental designs and focused specifically on AI's impact on foreign language anxiety among university students. In Cakmak's [8*] study, participants reported during interviews that chatbots sometimes struggled to accurately comprehend instructions and conversational content. These comprehension difficulties may have contributed to negative attitudes toward the chatbots and appeared to heighten participants' concerns about language accuracy and proficiency, potentially exacerbating their foreign language anxiety. These individual cases are directionally consistent with our non‑significant pooled effect, but given the small number of such studies and the overall non‑significant subgroup result, they should be interpreted as exploratory observations rather than causal explanations for the null finding. Future research should prioritize investigating communication barriers in student–chatbot interactions and conducting rigorous controlled studies to verify AI's efficacy for foreign language and mathematics anxiety. Effect sizes of AI on academic anxiety reduction by research design types The moderation analysis revealed no significant differences in effect sizes among RCT, quasi-experimental, and single-arm research designs. These findings are consistent with those reported by Bolat and Tas [6] and Molendijk et al. [58], who similarly found no significant differences between research design types in effect sizes for anxiety reduction in adult psychological treatment studies. Although the Q test between groups was non-significant, a descriptive examination of the effect size magnitudes and 95% confidence intervals suggested a pattern in which experimental designs yielded the largest point estimate (g = − 0.81) compared to quasi-experimental (g = − 0.42) and single-arm designs (g = − 0.63). Notably, although the single-arm design’s effect size was not statistically significant, its point estimate was descriptively larger than that of quasi-experimental designs. However, this observation should be interpreted with caution given the non-significant moderator test. Additionally, we observed that studies employing single-arm designs were more likely to generate statistical outliers [8*, 76*]. Based on these observations, future research may benefit from prioritizing experimental designs where feasible, as they may help produce more accurate estimates of AI's impact on academic anxiety. Effect sizes of AI on academic anxiety reduction across learning environments The literature offers divergent perspectives on how educational technology may impact students across different learning environments. One perspective suggests greater benefits in higher education compared to K-12 education. Bolat and Tas [6] analyzed the effects of game-based interventions across learning environments and reported that higher education students appeared to derive greater benefits than their K-12 counterparts. This disparity may stem from K-12 students' relatively limited language proficiency, autonomous learning capacity, and digital literacy, which could potentially hinder their effective utilization of AI tools. Higher education students, on the other hand, tend to possess more developed self-regulation abilities and learning initiative, which may enable them to make greater use of AI tools for mitigating academic anxiety. An alternative perspective suggests that educational technology may yield greater benefits at the K-12 level. Coban et al. [18] reported that immersive virtual reality (I-VR) produced larger effect sizes on academic outcomes among K-12 students compared to higher education students. Some studies have noted that higher education students may encounter communication difficulties with AI systems, particularly when the systems struggle to comprehend instructions [8*]. K-12 students, however, may receive more teacher support when using AI tools, which could help alleviate anxiety arising from communication challenges. In our meta-analysis, effect sizes for AI on academic anxiety reduction were −0.93 (primary), −1.20 (middle), and −0.35 (university), suggesting a descriptive V‑shaped pattern with the largest effect at the middle school level. However, the Q test was non‑significant (Q = 1.09, p = 0.38), indicating these differences are descriptive trends, not statistically robust. Moreover, both middle school (p = 0.06) and primary school (p = 0.06) effects were marginally significant, whereas the university effect was non‑significant (p = 0.15). Given the non‑significant moderator test, we refrain from strong claims about differential benefits across educational levels. These findings offer tentative hypotheses only. Future research should distinguish between primary and middle school rather than treating K‑12 as homogeneous, provided larger samples and more primary studies become available. More high‑quality studies are needed to clarify AI’s effects across learning environments. Effect sizes of AI on academic anxiety reduction across cultural backgrounds The analysis indicated no significant differences in AI's effects on academic anxiety reduction among Eastern, Western, and Islamic cultural contexts, consistent with Lei et al. [48]. Although the between‑group differences were not statistically significant (Q = 1.15, p = 0.37), a descriptively larger effect size was observed in Eastern countries where the effect was statistically significant (g = − 0.88, p < 0.01). Compared to students from Western and Islamic backgrounds, Eastern students tend to place greater emphasis on academic achievement and social expectations, and may be more responsive to external motivational influences [48, 60]. While the present meta-analysis does not directly compare absolute anxiety levels across cultures, these cultural distinctions may plausibly shape AI intervention effectiveness. Specifically, for Eastern students who are more sensitive to external academic feedback, AI tools providing structured guidance and personalized feedback may contribute to the observed stronger descriptive trend in anxiety reduction. Given the non-significant moderation result, such cultural discrepancies are only descriptive and should be interpreted with caution. Effect sizes of AI on academic anxiety reduction across intervention durations Regarding the timing of outcome assessment, our meta-analysis found that AI interventions significantly reduced academic anxiety both at immediate posttest (g = − 0.70, p = 0.02) and at delayed posttest (g = − 0.81, p < 0.01), with no significant difference between the two subgroups (Q = 0.02, p = 0.88). This demonstrates that the anxiety-reduction effects of AI interventions can be sustained across different assessment time points. The stable effects observed here do not support the novelty effect hypothesis, which posits that the benefits of educational technology fade as initial enthusiasm diminishes [42, 81]. By contrast, the present results align better with the cumulative effect perspective: prolonged exposure to AI tools may help students improve their operational proficiency, contributing to maintained or even slightly enhanced intervention outcomes [24*]. Notably, the descriptively larger effect size in delayed posttests should be interpreted with caution, as the between-group difference failed to reach statistical significance. Future research could adopt longer follow-up intervals to further explore the long-term temporal stability of AI’s effects on academic anxiety. Effect sizes of AI on academic anxiety reduction across publication periods The analysis indicated that effect sizes from studies published from 2023 onward were descriptively larger than those from studies published in 2022 and earlier, with a marginally significant between-group difference (Q = 4.07, p = 0.07). Specifically, studies published in 2022 and earlier yielded a marginally significant effect (g = − 0.36, p = 0.08), whereas studies published from 2023 onward showed a statistically significant effect (g = − 1.25, p < 0.01). One plausible interpretation is that recent advances in AI technologies such as ChatGPT and GPT-4 have enabled more effective human–computer interactions [3, 14]. However, alternative explanations beyond technological progress warrant consideration. For instance, differences in outcome measurement over time—earlier studies using less standardized or sensitive anxiety scales versus recent studies employing more domain‑specific instruments—could partly account for the observed pattern. Additionally, recent research may have concentrated on anxiety subtypes such as test anxiety, which are known to yield larger effect sizes, potentially inflating post‑2022 estimates. Finally, stronger publication pressure for positive or large‑effect findings following the launch of ChatGPT cannot be ruled out. To address these possibilities, future meta-analyses should employ meta-regression controlling for anxiety type, measurement characteristics, and publication bias, or cumulative meta-analysis, to separate genuine technological progress from artifacts. Effect sizes of AI on academic anxiety reduction across sample sizes We stratified the analyzed studies into three categories based on intervention sample size: those with fewer than 50 participants, those with 50–100 participants, and those exceeding 100 participants. The analysis indicated that studies with smaller sample sizes (n < 50) tended to yield larger effect sizes for AI interventions on academic anxiety reduction. This pattern is consistent with research by Hwang [40] and Sun & Zhou [74], who similarly reported larger effect magnitudes in studies with smaller samples. The relatively larger effects observed in smaller-scale educational settings may be related to more favorable instructional conditions. In these contexts, educators may be able to provide more direct supervision and guidance for AI tool implementation, potentially allowing students to benefit more from personalized learning pathways and real-time feedback mechanisms. Such environments may facilitate greater engagement and more interactive learning experiences [33, 74], which could create more favorable conditions for anxiety reduction. Conversely, as sample sizes increased, the observed effects of AI interventions on academic anxiety tended to diminish. This pattern may reflect reduced individualized attention and limited personalized support in larger educational settings, which could constrain the effective use of AI capabilities [38]. These findings suggest that the effectiveness of AI interventions may be moderated by group size and the associated level of personalized instruction. When implementing AI in educational contexts, it may be useful to consider the relationship between group size and the availability of individualized support. Further research could usefully investigate strategies for deploying AI tools in larger educational settings while seeking to maintain their potential anxiety-reducing benefits and supporting overall student development. Effect sizes of different AI types on anxiety reduction Among the AI types examined, Intelligent Tutoring Systems (ITS) yielded the largest descriptive effect size in reducing student anxiety, followed by robotics technology, which also showed a statistically significant effect (g = − 0.76, p < 0.01). The pattern for ITS is consistent with Yi's [94] findings, which similarly indicated that ITS was associated with greater improvements in students' mathematics performance compared to other AI types. The characteristics of ITS—including personalized feedback, immediate assessment, and adaptive instruction—may contribute to greater student self-efficacy, which may in turn reduce emotional burden during the learning process and potentially alleviate academic anxiety [54, 67*]. In contrast, chatbots showed no significant effect on reducing student anxiety in the present analysis (g = − 0.33, p = 0.33), nor did virtual technology (g = − 0.50, p = 0.25). Chatbots' non-significant effect may be attributable to the relatively early stage of chatbot implementation in educational practice. As an emerging technology, chatbots have limited application cases, and teachers' familiarity with their functionalities may still be developing, which could constrain their instructional support capabilities. Additionally, students may experience higher cognitive load and less stable interactive experiences when engaging with chatbots, which may limit the potential delivery of emotional support. However, the Q test for subgroup differences was non-significant (Q = 0.59, p = 0.66), indicating that the observed variations in effect sizes across AI types are descriptive trends rather than statistically robust disparities. Therefore, the apparently superior performance of ITS and robotics should be interpreted with caution. With this caveat in mind, the differential effects across AI types offer preliminary insights into potential mechanisms of anxiety reduction. ITS's descriptively larger effect may be attributed to its capacity to simultaneously address multiple mechanism pathways: it enhances perceived control through adaptive scaffolding, reduces social evaluation threat by removing human judgment from feedback loops, and builds self-efficacy through incremental mastery experiences. These pathways correspond closely with the control-value theory framework [64] outlined earlier: ITS intervenes on both the control dimension, through personalized pacing, and the value dimension, by lowering the stakes of evaluation. In contrast, chatbots may inadvertently heighten anxiety by introducing communication ambiguity and interaction instability, as suggested by the comprehension failures reported in Cakmak [8*]. These findings underscore that the effectiveness of AI-based anxiety interventions depends not merely on the presence of AI technology but on the specific fit between the AI modality and the targeted psychological mechanism. Therefore, this study recommends that future research should continue investigating the application effects of established AI technologies, such as ITS and robotics, while actively exploring the feasibility and effectiveness of emerging AI educational tools, including chatbots and generative AI, in instructional practice. Through accumulating practical experience and identifying best practices, the sustainable development and broader adoption of AI technology in education can be promoted. Limitations and future research Despite adherence to rigorous meta-analytic protocols, several limitations warrant acknowledgment. The literature search was restricted to three primary databases (Web of Science, Scopus, and ProQuest), potentially overlooking relevant studies indexed exclusively in other repositories such as Google Scholar. Future investigations should employ more comprehensive search strategies across additional databases to ensure broader literature coverage. Additionally, our analysis was confined to peer-reviewed journal articles, excluding conference proceedings and dissertations. Incorporating these alternative publication formats in subsequent meta-analysis would enhance the evidential foundation and mitigate potential publication bias. The scope of this meta-analysis was limited to AI interventions implemented within formal educational settings, without consideration of applications in community or home environments. Expanding research contexts would provide a more holistic understanding of AI's efficacy in addressing academic anxiety across diverse learning environments. Furthermore, our analytical approach focused exclusively on direct effects and moderator analyses, without empirically testing the mediating pathways proposed in the discussion section—such as enhanced perceived control, reduced evaluation threat, and increased self-efficacy—which we identified on the basis of the control-value theory [64]. Future research employing Meta-Analytic Structural Equation Modeling (MASEM) could formally test these pathways, thereby validating the theoretical mechanisms through which AI interventions influence academic anxiety. While our findings generally support AI's efficacy in reducing academic anxiety, contradictory results emerged in specific contexts. Notably, some single-arm studies indicated that AI applications may actually increase foreign language anxiety among university students. Due to sample limitations, we could only hypothesize about potential explanations through literature review. More methodologically rigorous experimental designs are needed to investigate these contextual variations and optimize AI interventions for foreign language learning environments. One limitation of this meta-analysis is that the PRISMA flowchart does not specify the exact number of studies excluded for each individual criterion, as this information was not recorded during the screening process. We acknowledge that such detailed reporting would improve transparency, and we will adopt this practice in future research. Conclusion This meta-analysis demonstrates that AI interventions exert a significant moderate overall effect on reducing students' academic anxiety (g = − 0.65). While descriptive differences in efficacy exist across anxiety subtypes, AI modalities, learning environments, cultural backgrounds and publication periods, most moderator analyses detected no statistically significant between-group disparities. Descriptively, intelligent tutoring systems and robotics achieved better outcomes, whereas chatbots and virtual technology failed to produce significant anxiety-reduction effects. The intervention effects were marginally significant for primary and middle school students, yet non-significant for university students. Studies conducted in Eastern cultural settings showed a larger descriptive effect compared with those in Western and Islamic contexts. The anxiety-alleviating effects remained stable from immediate to delayed posttests, lending support to the cumulative effect hypothesis rather than the novelty effect hypothesis. Studies published after 2022 reported larger effect sizes, though alternative explanations including varying measurement tools and publication bias cannot be excluded. Additionally, studies with smaller samples tended to yield greater effect sizes. Future research is recommended to adopt more rigorous experimental designs and longer follow-up assessments, and apply meta-analytic structural equation modeling to explore underlying mechanisms. Broader literature searches and inclusion of grey literature will also help reinforce the overall evidence base. Supplementary Information Supplementary Material 1. Authors’ contributions Weihao Wang conceived the study. The data analysis, results interpretation, and original manuscript writing were done by Weihao Wang and Long Diao. Weihao Wang and Jingting Yu reviewed the original manuscript. Weihao Wang and Long Diao conducted data collection and organization. Weihao Wang and Long Diao managed and coordinated the planning and execution of research activities. All authors read and approved the final manuscript. Funding This study was supported by the National Social Science Fund Major Project in Education Science “Research on the Pathways and Mechanisms of Reducing the Burden on Primary and Middle School Students” (Grant No. VHA240003). Data availability All data generated or analysed during this study are included in this published article. Declarations Ethics approval and consent to participate Not applicable. Competing interest The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement All data generated or analysed during this study are included in this published article.