Examining and Mitigating Ability-bias in LLMs via Self-Reflection

Abstract

Large language models (LLMs) (e.g., ChatGPT) are rapidly integrating into our daily lives, fundamentally shaping how we engage with, process information or make decisions. Despite their significant potential, LLMs can encode social biases (e.g., gender, culture) that amplify problematic and stereotypical representations of marginalized groups. Given the discriminatory impact that bias in LLMs can have on people with disabilities, in this work we examine ability bias in LLMs. We analyze LLM responses to a set of carefully crafted prompts across different abilities, and explore self-reflection through prompt chaining as a debiasing approach. Our findings surface linguistic associations encoded in LLMs with different disabilities. We note the types of justifications or rationalizations provided as explanations in LLM responses — which has implications on the trust associated with LLM responses. Our proposed approach of model self-reflection demonstrates improvement in LLM responses and thereby contributes to debiasing literature.