Encoding Inequity: Examining Demographic Bias in LLM-Driven Robot Caregiving
Raj Korpan
TL;DR
The paper investigates demographic bias in LLM-driven robot caregiving by prompting a state-of-the-art LLM with 138 demographic labels across four Categories and twenty Subcategories to generate caregiving narratives. It systematically analyzes word count, syntactic complexity, sentiment, and text similarity via TF-IDF and SBERT, complemented by TF-IDF and SBERT clustering and ANOVA-based significance testing. The findings reveal category-dependent disparities: age and disability labels tend to produce more verbose or simplified descriptions, while gender/sexuality and race/ethnicity narratives exhibit distinct, often less positive patterns, with notable lower sentiment for disability and LGBTQ+ identities. The study underscores ethical implications for DEI in HRI and advocates for bias mitigation, intersectional analyses, and principled guidelines to ensure fair, affirming, and inclusive robotic caregiving across diverse populations.
Abstract
As robots take on caregiving roles, ensuring equitable and unbiased interactions with diverse populations is critical. Although Large Language Models (LLMs) serve as key components in shaping robotic behavior, speech, and decision-making, these models may encode and propagate societal biases, leading to disparities in care based on demographic factors. This paper examines how LLM-generated responses shape robot caregiving characteristics and responsibilities when prompted with different demographic information related to sex, gender, sexuality, race, ethnicity, nationality, disability, and age. Findings show simplified descriptions for disability and age, lower sentiment for disability and LGBTQ+ identities, and distinct clustering patterns reinforcing stereotypes in caregiving narratives. These results emphasize the need for ethical and inclusive HRI design.