Key points are not available for this paper at this time.
Background The increased prevalence of ASD has generated a pressing demand for flexible therapeutic and educational tools. AI has been suggested as a potential bridge to this gap, but the translation from a model to a clinical application necessitates rigorous assessment. The purpose of the present review is to compile and examine the existing literature to demonstrate AI interventions that have advanced from a proposed model to being used with human participants. Methods We systematically searched five databases (Embase, PubMed, ScienceDirect, IEEE Xplore, Web of Science; Jan 2016–Dec 2025) for AI-based ASD interventions. Two reviewers independently assessed eligibility. Inclusion criteria were as followed: (1) participants with confirmed ASD diagnoses; (2) an intervention sample size of N ≥ 6; (3) AI as a central therapeutic, educational, or rehabilitative component; and (4) multi-session protocols with specified timeframes. Study types ranged from system development and feasibility trials to RCTs. Results 14 studies met inclusion criteria. AI (e.g. robotics, VR, and wearables) functioned as a social mediator, improving social-emotional outcomes (e.g., ADOS, SRS scores) by reducing cognitive load. Significant mechanisms included real-time task adaptation and precise behavioral monitoring via e.g. eye-tracking. However, significant heterogeneity was observed in intervention dosage (median 4–12 hours). Most studies were limited by small, male-dominated samples (N < 20) and a total absence of adult participants. Conclusion Based on the findings, AI seems highly promising for patient-specific ASD therapy via proactive, data-driven scaffolding. In order to move toward implementation of AI in ASD care, more RCTs and crucial augmentation of representation gaps concerning adult and female ASD phenotype studies are required.
Kuca et al. (2026) studied this question.