Journal of Global Research in Education and Social Science

Artificial intelligence (AI) has been positioned as a means of individualising instruction and social-communication support for autistic pupils in kindergarten to grade twelve (K–12) settings. Prototype systems that combine social robotics, adaptive virtual environments, wearable sensing and, more recently, large language models have proliferated, yet the transition from laboratory demonstration to durable classroom practice remains poorly characterised. This critical narrative review examines what the accessible evidence supports about the effectiveness of AI-enabled educational interventions for autistic K–12 learners, and appraises the readiness of these interventions for translation into ordinary school settings. Literature was identified through structured searching of open scholarly indexes and registries, supplemented by citation tracking and targeted retrieval from institutional sources, covering January 2000 to 5 June 2026. Evidence was appraised for design adequacy, comparator quality, outcome selection, reporting completeness and implementation relevance, and was synthesised thematically around three interdependent readiness dimensions: technical, contextual and normative. Pooled estimates from robot-mediated and virtual-reality interventions are consistently positive but unstable, with wide confidence intervals, heavy dependence on small trials, and effect sizes that attenuate sharply when analyses are restricted to randomised designs and detection-bias-free outcomes. Very few studies examine generalisation to untrained settings, maintenance beyond short follow-up, or delivery by teachers under routine conditions. Machine-learning components that supply the adaptive behaviour of these systems are rarely validated externally, and evidence from speech and vocal-marker research indicates substantial performance loss across tasks, languages and speaker populations. Reporting of adverse events, conflicts of interest and pre-registered outcomes remains inadequate across the wider autism intervention literature, which constrains confidence in the subset relevant to AI. Contextual readiness is the least-studied dimension: implementation research in autism education concentrates on early implementation phases and rarely addresses sustainment, cost or scale-up. Normative questions concerning outcome selection, autistic participation in design and data governance are raised more often than they are resolved. The field is therefore better described as accumulating feasibility evidence than as approaching classroom readiness, and the principal constraint is no longer technical novelty but the absence of pragmatic, teacher-delivered, adequately powered evaluations conducted in ordinary schools.