High Utility, Low Trust: A Systematic Review and Meta-Analysis of the "Language Pivot" and the Perception–Performance Paradox in Generative AI

This repository contains the supplementary data and methodological appendices supporting the systematic review and meta-analysis titled "High Utility, Low Trust: A Systematic Review and Meta-Analysis of the ‘Language Pivot’ and the Perception–Performance Paradox in Generative AI." The study synthesizes the "Post-ChatGPT" evidence base (2023–2025) regarding Generative AI in education (n = 50 studies), investigating the divergence between measured learning gains and user trust (the "Perception–Performance Paradox") and testing the hypothesis that language domains exhibit stronger efficacy than STEM domains (the "Language Pivot"). Repository Contents: This dataset includes three files containing statistical outputs, risk-of-bias appraisals, and bibliometric visualizations: Online Resource 1 (Meta-Analysis Data & Diagnostics): · Table S1.1: Study-level effect sizes, sample sizes, and metadata for the 22 comparisons included in the quantitative synthesis. · Table S1.2: Random-effects model summaries, subgroup analyses (Domain, Outcome Type), and sensitivity analysis results (e.g., RCT-only). · Table S1.3: Risk-of-bias distribution for the quantifiable subset (k = 22). · Table S1.4: Egger’s regression test results for publication bias. · Figures S1.1–S1.4: Funnel plots and Forest plots stratified by outcome type and academic domain. Online Resource 2 (Risk of Bias Assessment): · Table S2.1: Risk-of-bias appraisal approach and decision rules (RoB 2, ROBINS-I, JBI). · Table S2.2: Item-level risk-of-bias appraisal and primary concerns for all included studies (N = 50). Online Resource 3 (Methodological Appendices): · Table S3.1: Detailed eligibility criteria (PICOS) and operational definitions. · Table S3.2: Complete database search strategy (Web of Science, Scopus, IEEE, ACM). · Figure S3.1: Evidence map of study distribution by publication year and domain. · Figure S3.2: Geographic distribution of included evidence.