LLMs over-idealize disability, study finds, erasing real struggles

Researchers Marco Bombieri, Simone Paolo Ponzetto, and Marco Rospocher investigated how large language models (LLMs) represent disability by generating simulated social media posts from the perspective of individuals with disabilities. They compared these AI-generated posts with real posts written by people with disabilities, analyzing emotional tone, sentiment, and thematic content. The study, accepted for publication in ACM Transactions on Intelligent Systems and Technology, reveals two critical findings.

First, LLMs tend to idealize disability, producing overly positive stereotypes that fail to capture the nuanced challenges and lived realities of disabled individuals. This overcompensation, likely a result of debiasing efforts, creates a sanitized narrative that can be misleading. Second, the models exhibit a negative bias: topics such as career and entertainment are disproportionately linked to non-disabled individuals, reinforcing exclusionary narratives. The authors argue these findings align with broader concerns that LLMs struggle to reflect the diverse realities of marginalized communities, emphasizing the need for critical scrutiny in AI deployment.