AutismCarebot: Emotion-First, Source-Aware Conversational Support for Autistic Users

Abstract

Autistic people often seek support for stress regulation, communication challenges, and navigating services, yet general-purpose chatbots may miss neurodiversity-affirming language and can provide ungrounded advice. We present AutismCarebot, a web-based conversational agent that combines (1) an emotion-first response structure (validation, normalization, and a clear next step) with (2) retrieval-augmented generation (RAG) to provide concise guidance grounded in curated autism resources with in-text citations. We designed the interaction to reduce cognitive load through quick prompts, progressive disclosure of detail, and a built-in “Take a Break” module offering guided breathing and grounding. We evaluated AutismCarebot through an automatic comparison with an unmodified large language model baseline and a survey study (n=18) including autistic adults, caregivers, educators, and students. Results indicate AutismCarebot produced more empathetic and better-grounded responses than the baseline, with no unsafe outputs observed in the audit. The survey feedback highlighted AutismCarebot's soothing tone, clarity, and practical advice, while emphasizing the necessity to control the level of links and provide multimodal support.