An Overview of Gender and Political Biases in LLMs
Published in Proceedings of the Midwest Instruction and Computing Symposium 2025, 2025
As the public increasingly relies on Large Language Models (LLMs) like GPT for information, and as students increasingly turn to these models for assistance in studying, the impact of their influence will rise. Thus, it becomes increasingly imperative to study what biases these models exhibit. We examine eleven studies to determine the political leanings and the prevalence of gender stereotyping in GPT. Examinations reveal a progressive political bias with varying consistency based on language implementation. Current literature suggests GPT also exhibits gender biases, especially in the choice of adjectives, professions, and gender-based preferences. GPT often favors stereotypical roles or expressions and shows a tendency to assign higher scores or use more positive language when evaluating certain genders. It is likely all the aforementioned biases reflect the biases inherent in the training data. To mitigate these biases, we suggest training LLMs with diverse training data incorporating a balance of contrasting perspectives.
Recommended citation: Jared B. Jones, Naeem Seliya, Emily M. Hastings, and Benjamin T. Fine. 2025. An Overview of Gender and Political Biases in LLMs. In Proceedings of the Midwest Instruction and Computing Symposium 2025.