Born With a Silver Spoon? Investigating Socioeconomic Bias in Large Language Models
Abstract: Socioeconomic bias in society exacerbates disparities, influencing access to opportunities and resources based on individuals' economic and social backgrounds. This pervasive issue perpetuates systemic inequalities, hindering the pursuit of inclusive progress as a society. In this paper, we investigate the presence of socioeconomic bias in large language models. To this end, we introduce a novel dataset {\sc SilverSpoon}, consisting of 12000 samples that provide a multifaceted analysis of this complex issue. This dataset has three subsets. The first 3000 samples focus on normative judgement evaluation, consisting of hypothetical scenarios in which people of different socioeonomic class make difficult decisions. This subset of the dataset has a dual-labeling scheme and has been annotated by people belonging to both ends of the socioeconomic spectrum. The second subset of this dataset focuses on demographic driven profession prediction, and consists of 8000 samples that investigate socioeconomic bias across a plethora of combinations of gender, race and location. Finally, the third subset of the dataset focuses on contextual narrative bias analysis. This subset consists of 1000 LLM generated stories, which have been leveraged to detect the presence of subtle stereotypes against certain socioeconomic classes belonging to various demographic groups. Using {\sc SilverSpoon}, we evaluate the degree of socioeconomic bias expressed in state-of-the-art large language models. We also perform extensive quantitative and qualitative analysis to analyze the nature of this bias. Our analysis reveals that state-of-the-art large language models exhibit implicit and explicit socioeconomic bias, which is further augmented by stereotypes emanating from a combination of gender bias and racial bias. To foster further research in this domain, we make {\sc SilverSpoon} and our evaluation harness publicly available.
Paper Type: Long
Research Area: Ethics, Bias, and Fairness
Research Area Keywords: Socioeconomic Bias, Fairness in AI, LLM Bias
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 1196
Loading