Investigating Social Bias in GPT-4: A Study on Race Representation

Large language models (LLMs) have become increasingly powerful, achieving impressive results across various tasks. However, a critical challenge lies in their potential to inherit and perpetuate social biases present in the data they are trained on. This study delves into the issue of social bias within the GPT-4 model, specifically focusing on the bias in how it represents race. We propose a novel method for quantifying bias in GPT’s outputs. This method will involve using binary responses of their API to analyze any preconceived notions. The resulting data will be visualized using graphs to illustrate any patterns or disparities in treatment. Further analysis will be conducted to explore the root causes of these biases and their potential impact. Finally, we will present and discuss potential solutions aimed at mitigating social bias in LLMs. By investigating and addressing social bias within GPT-4, this study aims to contribute to the development of fairer and more responsible LLMs.