Investigating Perception of Gender Stereotypes in Large Language Models: A Computational Grounded Theory Approach

Artificial Intelligence has expanded its influence far beyond traditional boundaries in our society. One prominent application of artificial intelligence is the use of large language models, which have transcended their initial roles in high-tech industries and academic research and are now actively utilized by individual users. These models have continually improved over the years in their generative capabilities and performance across numerous tasks. However, they still pose a persistent risk of reproducing biases and stereotypes. Previous research has predominantly focused on quantitatively measuring biases in these large language models. In this study, we seek to assess not just the presence of bias itself, but the perception of stereotypes by these models via in-depth exploration of their responses. We demonstrate how the computational grounded theory framework, which integrates qualitative and quantitative approaches, can be applied in this context to assess the conceptualization of stereotypes. Furthermore, we contrast language model results with a survey of 400 human participants who also completed similar prompts as the model in order to understand people’s perception of gender stereotypes. The results indicate substantial similarities between language model and human perceptions of stereotypes, highlighting that a model’s perception stems from societal perception of stereotypes.