Understanding Down Syndrome Stereotypes in LLM-Based Personas
Chantelle Wu, Peinan Wang, Nafi Nibras, Meida Li, Dajun Yuan, Zhixiao Wang, Jiahuan He, Mona Ali, Mirjana Prpa
TL;DR
The paper introduces Persona-L, a system that uses LLMs and retrieval-augmented generation to create interactive personas of people with Down syndrome, augmented by a stereotype-detection module trained on an expanded MGSD-derived dataset. Through a caregiver- and clinician-facing user study (N=10), it reveals that stereotypes can emerge not only in content but also in interface design and conversational delivery, underscoring the need for participatory, human-in-the-loop approaches. The authors demonstrate that while RAG improves factual grounding, deterministic improvements in tone and structure are needed to avoid infantilizing or overpositive portrayals. The work contributes a technical stereotype-detection framework, empirical caregiver feedback, and design recommendations to advance representational fairness in AI personas for disability communities.
Abstract
We present a case study of Persona-L, a system that leverages large language models (LLMs) and retrieval-augmented generation (RAG) to model personas of people with Down syndrome. Existing approaches to persona creation can often lead to oversimplified or stereotypical profiles of people with Down Syndrome. To that end, we built stereotype detection capabilities into Persona-L. Through interviews with caregivers and healthcare professionals (N=10), we examine how Down Syndrome stereotypes could manifest in both, content and delivery of LLMs, and interface design. Our findings show the challenges in stereotypes definition, and reveal the potential stereotype emergence from the training data, interface design, and the tone of LLM output. This highlights the need for participatory methods that capture the heterogeneity of lived experiences of people with Down Syndrome.