Asian Journal of Education and Social Studies

Artificial intelligence (AI) now influences educational decisions through intelligent tutoring, learning analytics, automated feedback and generative conversational systems. Yet the prominence of these applications has outpaced agreement on what constitutes educational benefit, how risks should be governed and whether teachers are gaining or losing professional discretion. This critical narrative review integrates empirical studies, meta-analyses, conceptual scholarship and authoritative educational guidance published between 2011 and 28 August 2026, with earlier historical work considered where essential. The analysis differentiates conventional predictive and adaptive AI from generative systems, evaluates learning outcomes against the independence of demonstrated competence, and connects pedagogical effectiveness to implementation conditions, academic integrity and teacher roles. Established intelligent tutoring research provides comparatively mature evidence of improvements over some conventional instructional conditions, whereas recent generative AI evidence is more heterogeneous and sensitive to task design, assessment conditions and learner characteristics. Experimental findings indicate that better assisted performance cannot automatically be interpreted as improved unassisted learning. Educational implementation is further constrained by model reliability, curriculum fit, data protection, linguistic fairness, infrastructure, evaluation capacity and institutional accountability. Evidence on AI-generated text detection raises particular concerns about false accusations and unequal treatment of multilingual writers. Teacher expertise remains central to defining learning goals, interpreting algorithmic outputs, designing valid assessments and responding to learners' social and developmental needs. The review argues that the key unit of evaluation is not the AI tool alone but the instructional arrangement in which it operates. It proposes a conditional framework linking tool function, learning mechanism, verification, human oversight and equitable access. Research priorities include independent delayed assessments, multi-site implementation studies, subgroup-sensitive harm evaluation, longitudinal teacher-work studies and transparent audits of data governance. AI is best understood as a contingent educational resource, not an autonomous substitute for teaching or a uniform solution to educational inequality.