Autism Spectrum Disorder (ASD) is a heterogeneous neurodevelopmental condition whose early diagnosis often depends on subjective behavioural assessments. Eye-tracking has emerged as an objective tool for capturing atypical visual attention patterns associated with ASD, while recent advances in Machine Learning enable automated analysis of gaze data. However, existing studies show substantial heterogeneity in datasets, experimental paradigms, feature extraction, and evaluation protocols, limiting comparability and clinical transferability. This paper presents a structured state-of-the-art review of AI-based eye-tracking approaches for ASD detection published between 2020 and 2025. Thirty-four studies were comparatively analyzed to identify methodological trends, performance patterns, and current limitations, outlining key challenges toward robust and clinically applicable screening systems.