Uncovering Ableism in Large Language Models’ Responses: Hybrid Sentiment and Thematic Analysis Approach for Disability Representation

Large language models (LLMs) have become central to how people access and interact with information, yet their potential to reproduce ableist bias remains underexamined, especially in non-English-language settings. This study examines disability representation in LLM-generated outputs across English and Bahasa Indonesia using a hybrid analytical framework. A total of 360 responses were generated by ChatGPT, Gemini, and Microsoft Copilot through a factorial prompt design varying disability type, socioeconomic class, and language. The proposed analysis combined lexicon-based sentiment analysis, topic modeling, qualitative thematic analysis (three analysts), and consensus-based human ableism scoring. The results show that positive sentiment dominated across both languages but did not reliably indicate non-ableist representation. Ableist classifications were most concentrated in lower socioeconomic class condition, with schizophrenia-related prompts and Gemini-generated outputs showing the highest proportions of ableist classifications across disability types and LLMs, respectively. Theme-level analysis showed that ableism was most prevalent in responses framing disability through sensory overload, mobility barriers, and uneven professional access in negative sentiment outputs, and through life trajectories, aspirations, and participation shaped by socioeconomic access in positive sentiment outputs. Overall, ableist bias appeared intersectionally across socioeconomic class and disability type, with the proposed hybrid framework providing a more sensitive approach for identifying ableist representation in AI-generated text.