Virginia Tech researchers have conducted a study revealing geographic bias in the language model ChatGPT when it comes to providing location-specific information about environmental justice issues. The study, published in the journal Telematics and Informatics, highlights the potential presence of biases in current generative artificial intelligence (AI) models.
ChatGPT, developed by OpenAI Inc., is a large-language model that utilizes AI to understand questions and generate text responses based on user requests. The technology has a wide range of applications, including content creation, information gathering, data analysis, and language translation.
Assistant Professor Junghwan Kim of the College of Natural Resources and Environment, a geographer and geospatial data scientist, explained that while generative AI has powerful potential, it is crucial to investigate the limitations of the technology to ensure future developers are aware of potential biases. This motivated the research at Virginia Tech.
The research group conducted their investigation by asking the ChatGPT interface, broken down by county, about environmental justice issues in each of the 3,108 counties in the contiguous United States. Environmental justice was chosen as the topic to expand the range of questions typically used to test the performance of generative AI tools. By analyzing ChatGPT’s responses against sociodemographic considerations such as population density and median household income, the researchers were able to measure its accuracy.
The findings of the study indicate that while ChatGPT can identify location-specific environmental justice challenges in large, high-density population areas, it falls short in providing contextualized information on local environmental justice issues. Out of the 3,018 counties entered, ChatGPT was only able to provide location-specific information for 17% of them.
The study revealed that in rural states like Idaho and New Hampshire, over 90% of the population lived in counties where ChatGPT could not provide specific information. In contrast, states with larger urban populations, such as Delaware and California, had fewer than 1% of their population living in counties without access to specific information.
The implications of these findings highlight the importance of testing for biases in generative AI models like ChatGPT. As generative AI becomes increasingly prevalent as a tool for accessing information, it is essential to address potential biases that may arise due to geographic discrepancies.
Assistant Professor Ismini Lourentzou of the College of Engineering, a co-author on the paper, emphasized the need for further research to refine localized and contextually grounded knowledge, safeguard large-language models against challenging scenarios, and enhance user awareness and policy around potential biases.
Lourentzou believes that addressing the reliability and resiliency of large-language models is a critical step forward. The research conducted at Virginia Tech aims to guide future research efforts to strengthen the capabilities of models like ChatGPT.
In conclusion, the study conducted by Virginia Tech researchers sheds light on the presence of geographic bias in ChatGPT’s ability to provide location-specific information about environmental justice issues. By highlighting these limitations, the research aims to contribute to the development of more accurate and unbiased AI models in the future.
