GPT-4 biases

The dataset used in this study consists of multiple subsets corresponding to different library application scenarios, with file names indicating task types and user identity attributes (e.g., gender and race). Each subset contains GPT-4–generated outputs for controlled prompts, enabling systematic comparison across demographic conditions. The dataset is structured to capture variations in language use, interaction patterns, and content generation, thereby facilitating the analysis of potential biases in GPT-4’s responses across different user groups and library contexts.