Virginia Tech researchers have identified geographic biases in the performance of ChatGPT, a large-language model developed by OpenAI Inc. The researchers found limitations in the tool’s ability to provide location-specific information about environmental justice issues. Published in the journal Telematics and Informatics, their findings shed light on potential biases present in current generative artificial intelligence (AI) models.
ChatGPT is an AI model designed to understand questions and generate text-based responses. Its applications range from content creation and information gathering to data analysis and language translation.
Assistant Professor Junghwan Kim of the College of Natural Resources and Environment, who specializes in geography and geospatial data science, expressed the need to investigate the limitations of generative AI to ensure the recognition of potential biases by future developers. This motivated the research conducted by the team.
To measure ChatGPT’s performance, the researchers elicited information on environmental justice issues from the tool by asking questions related to each of the 3,108 counties in the contiguous United States. By assessing ChatGPT’s responses against sociodemographic factors such as population density and median household income, the researchers could evaluate its accuracy. Environmental justice was chosen as a topic to expand the range of questions typically used to test the capabilities of generative AI tools.
The research group discovered that while ChatGPT was effective in identifying location-specific environmental justice challenges in densely populated areas such as Los Angeles County, California, it struggled to provide contextualized information and identify local environmental justice issues.
Given the growing prominence of generative AI as a tool for information gathering, it is essential to test for potential biases in modeling outputs. By exposing geographic biases in ChatGPT, the researchers hope to contribute to the improvement of similar programs.
Assistant Professor Ismini Lourentzou, a co-author of the study from the College of Engineering, outlined three areas of further research for large-language models like ChatGPT. Lourentzou stressed the importance of addressing issues concerning reliability, resiliency, and capabilities to enhance the performance of these models.
The researchers’ investigation into ChatGPT’s limitations adds to Assistant Professor Kim’s previous work exploring the tool’s understanding of transportation issues in the U.S. and Canada. Kim’s research group, Smart Cities for Good, focuses on using geospatial data science methods and technology to address urban social and environmental challenges.
As the reliance on large-language models like ChatGPT grows, it becomes crucial to evaluate their capabilities and identify potential biases. Kim emphasized the need to study the disparity of information between big and small cities, as well as between urban and rural environments.
By shedding light on the existing geographic biases in ChatGPT, this research serves as a starting point to anticipate and mitigate disparities in information provided by AI models. The findings pave the way for future research and development to enhance the capabilities and reliability of generative AI models.
By recognizing and addressing biases, developers and programmers can ensure that AI tools like ChatGPT provide more accurate and inclusive information, benefiting users across various locations and populations.
