Neurotypical Assumptions in Large Language Models Across Individual Support and Interpersonal Communication

Abstract

Large language models are increasingly part of how people seek support and communicate with others, while also reproducing normative assumptions about how people think, interact, and express themselves. This dissertation investigates how these assumptions surface for neurodivergent people across individual and interpersonal contexts. My preliminary work examined how neurodivergent adults use LLMs for communication and emotional support, finding that responses often missed users’ thought processes, communication styles, and support needs. Building on this work, I am developing a dataset of 150 dyadic conversations, with utterance-level annotations and written rationales from both autistic and neurotypical participants. I use these annotations to examine how understanding, misunderstanding, and repair unfold across dyads, and to evaluate whether LLM interpretations of the same moments align more closely with autistic or neurotypical participants’ perspectives. This dissertation contributes empirical evidence and participant-grounded methods for examining the normative assumptions embedded in AI systems that shape interpersonal communication.