Classroom Implementation of AI ‐Assisted Fraction Instruction for Students With Mild Autism

ABSTRACT Background Fractions challenge students with autism spectrum disorder (ASD) because they require coordinating part‐whole reasoning, symbol‐magnitude mapping, and proportional thinking. Objective To examine pre‐post change and classroom feasibility during participation in AI‐assisted fraction instruction among sixth graders with mild ASD (FSIQ ≥ 85). Methods A 12‐week, classroom‐embedded, one‐group pretest‐posttest implementation study ( n = 33) used AI Lobe lessons integrating virtual manipulatives, interactive number‐line tasks, and scaffolded feedback. AI Lobe functioned as a constrained, rule‐based adaptive learning environment in which student response data informed task sequencing, difficulty progression, and feedback selection; it was not an autonomous generative AI tutor. Pre/post assessments indexed conceptual understanding, and analyses emphasised distributional change and robust nonparametric inference. Results Post‐test scores exceeded pre‐test scores for most students (median gain = 14.5, IQR = 6.2–22.1; p < 0.001). Visual summaries indicated broadly distributed positive change. Implementation fidelity met the a priori threshold (≥ 80% of core elements across observed sessions), and most students completed 24–30 lessons. Positive pre‐post change was observed across equivalence, comparison, and number‐line subscales after Holm‐Bonferroni adjustment. Conclusions The teacher‐mediated, multi‐representation program appeared feasible in mainstream classrooms, and students showed substantial pre‐post change during the instructional period. Because the study had no comparison group and assessment form was confounded with measurement occasion, the findings constitute preliminary evidence of feasibility and temporal association, not causal effectiveness. Implications Controlled trials should use comparison groups, formally equated or counterbalanced assessment forms, maintenance and transfer measures, and adequately powered moderator analyses.