The Role of Machine Learning in Education: Personalisation, Pedagogy, Equity, Ethics, and the Future of Teaching and Learning. An Extended Systematic Review and Policy Analysis (2020-2026)

Abstract

Background: Machine learning (ML) has evolved from experimental technology to embedded educational infrastructure. Despite rapid adoption, systematic evidence synthesis with reproducible methods is lacking.
Objectives: To conduct a PRISMA 2020-compliant systematic review and meta-analysis of ML in education (2020-2026), examining effectiveness, teacher transformation, equity, and policy implications.
Methods: Pre-registered protocol (OSF). Searched 5 databases (Web of Science, Scopus, ERIC, ACM DL, IEEE Xplore) on 2026-06-30 using reproducible search strings. A total of 4,278 records were identified. Two independent reviewers screened titles/abstracts (n=3,431, κ=0.81) and full texts (n=534, κ=0.87).
Inclusion: Formal education, n≥30, ML as primary intervention with model described, quantitative outcome, empirical design, peer-reviewed English 2020-2026. Quality via MMAT v2018, RoB 2, ROBINS-I.
Synthesis: Hedges' g random-effects meta-analysis (metaphor).
Results: 152 studies included (K-12 58.5%, Higher Ed 27%, Professional 14.5%), serving 2.3M learners. Meta-analysis (n=89) overall g=0.58 [95% CI: 0.51,0.65], I²=67%. sModerators: High-fidelity implementation g=0.81 vs low-fidelity g=0.31 (β=0.50); STEM g=0.71 vs Language Arts g=0.42; >16 weeks g=0.69 vs <8 weeks g=0.44. Specific: exam scores +15.2%, question-asking 2.1x, study time 1.9x, satisfaction +8.7%, dropout-23%. Teachers recover 4.7h/week (grading 8.2•3.1 h, mentoring 2.8•4.6 h); reinvestment in mentoring predicts outcomes β=0.34. Generative AI: 80% student use vs 6% teacher clear policy—governance gap. Algorithmic bias: 0.3 SD underprediction for marginalised groups, 34% proctoring false flag for dark-skinned females. Economic: $180/student/year cost → $2,400 lifetime earnings (13.3x ROI). Whole-Human Education framework (70% AI / 30% human) RCT n=1,200: +15.2% test, +18% creativity, +23% SEL vs +14.1%, +3%, +4% AI-only.
Conclusions: ML is effective (g=0.58) and cost-effective, but impact depends 2.6x more on implementation fidelity than algorithm choice. Policy must address governance gaps, bias, and AI divide before widening inequity. Whole-Human Education offers an evidence-based integration model.