Publication

Abstract : Large Language Models (LLMs) hold increasing potential for supporting autistic learners through personalised instruction, adaptive content generation, communication scaffolding, and structured learning support. However, most existing AI educational systems are developed without grounding in the lived realities of parents, tutors, and educators who provide educational support to autistic learners, particularly within low-resource contexts such as India. Current systems also frequently reflect neurotypical communication assumptions that create accessibility barriers for neurodivergent users. This study employed a qualitative requirements analysis to investigate how LLM-based educational tools may be more effectively designed for neurodiversity-informed educational contexts. The study combined semi-structured interviews with eight Indian parents, teachers, and tutors, alongside a netnographic analysis of 14 global online sources. Findings revealed that participants already engage in extensive adaptive educational practices involving interest-based learning, multimodal instruction, visual scaffolding, and flexible pacing. The analysis identified key system-level requirements for neurodiversity-aligned LLM tools, such as increased interest-based anchoring, goal-oriented interactions, privacy settings, multilingual accessibility, and simplified interfaces for parents. The study also identifies the "neurotypical default problem" and proposes mitigation through neurodiversity-affirming communication styles. The paper contributes empirically grounded design requirements and proposes a human-in-the-loop architectural framework positioning caregivers as mediators of AI-assisted learning.