Addressing Methodological Considerations in the Study of Generative AI: A Case Study on Bias

Exploring the dynamics of generative AI systems requires flexible methodological frameworks to adapt to their continuous changes and complexity. This presentation discusses methodological considerations for effectively studying generative AI systems based on a case example. We draw on one of our previous studies, in which we analyzed responses from two popular AI models to homophobic statements. The prompts varied by including or omitting contextual user information. Our interest lay in how such information might alter the responses. The subtleties in the responses necessitated qualitative in-depth analysis besides quantitative statistics to effectively identify nuanced differences between categories. We argue that although quantitative methods are valuable for broad analyses, in-depth qualitative analysis should be considered to fully capture the complexities and subtle biases of newer generative AI systems. Our study demonstrates the application of mixed-methods research in AI bias studies, providing a framework for when and how to integrate these methodologies.

Additionally, our methodological decisions included strategies for eliciting responses from the models (opting for bias attack instructions by formulating negative statements), determining the formulation of these negative statements compared to standard survey questions, managing variability in responses under identical conditions, and addressing potential differences in responses based on the user's location. We also evaluated whether to utilize generative AI tools in the analysis process, such as for category creation. This presentation outlines the solutions we adopted to address these considerations.