AccessEval: Disability Bias in Large Language Models

AccessEval examines whether mentioning a disability changes the quality and tone of a language model’s response. The benchmark compares paired neutral and disability-aware queries across 21 models, six domains, and nine disability types.

Authors: Srikant Panda, Amit Agarwal, and Hitesh Laxmichand Patel. Venue: EMNLP 2025. Recognition: Best Social Impact Paper Award.

What the study measures

The evaluation considers factual accuracy, sentiment, and social perception. The paper reports more factual errors, negative tone, and stereotyping in disability-aware responses, with differences across domains and disability types.

Paper and citation

Read the original paper and publication record for the methodology, results, and citation details. DOI: 10.18653/v1/2025.emnlp-main.1653.

Explore my research interests and publications.