Examining large language models within autism-related contexts: A systematic review of bias and (MIS) representation

Abstract

Artificial Intelligence (AI) is quickly becoming an integral part of everyday life. AI is no longer limited to advanced applications; it is embedded in applications and tools we use daily to streamline tasks and solve problems. Large Language Models (LLMs), such as ChatGPT and Claude, are designed to help users make decisions, produce recommendations, and provide more personalized solutions to user inquiries. The immense reach these systems have had in such a short period has significantly impacted education, work, and interpersonal interactions. AI systems are only becoming more common, and it is imperative that developers work to address bias within them to promote fair use for all, not only the neurotypical. Research investigating the bias harbored by LLM platforms is particularly limited when it comes to addressing bias against individuals with autism. We applied inductive thematic analysis, followed by deductive thematic analysis, to analyze 34 scientific research papers gathered from two major academic databases: ACM Digital Library and IEEE Xplore. This method helped us learn more about the current empirical understanding of LLM bias as it relates to the context of autism. We found that bias within LLMs has the potential to harm individuals with autism who are looking for answers to problems for which LLMs have been postulated as a solution. LLMs also have the potential to reinforce negative stereotypes, leading users with autism who already face bias in daily life to continually struggle when faced with this new technological frontier.