Exploring the limitations in how ChatGPT introduces environmental justice issues in the United States: A case study of 3,108 counties

Abstract

The potential of Generative AI, such as ChatGPT, has sparked discussions among researchers and the public. This study empirically explores the capabilities and limitations of ChatGPT, specifically its portrayal of environmental justice issues. Using OpenAI’s ChatGPT API, we asked ChatGPT (GPT-4) to answer questions about environmental justice issues in 3,108 counties in the contiguous United States. Our findings suggest that ChatGPT provides a general overview of environmental justice issues. Consistent with research, ChatGPT appears to acknowledge the disproportionate distribution of environmental pollutants and toxic materials in low-income communities and those inhabited by people of color. However, our results also highlighted ChatGPT’s shortcomings in detailing specific local environmental justice issues, particularly in disadvantaged (e.g., rural and low-income) counties. For instance, ChatGPT could not provide information on local-specific environmental justice issues for 2,593 of 3,108 counties (83%). The results of the binary logistic regression model revealed that counties with lower population densities, higher percentages of white population, and lower incomes are less likely to receive local-specific responses from the ChatGPT. This could indicate a potential regional disparity in the volume and quality of training data, hinting at geographical biases. Our findings offer insights and implications for educators, researchers, and AI developers.