Research Reveals Geographic Biases in ChatGPT's Responses

A recent study from the University of Oxford and the University of Kentucky indicates that ChatGPT exhibits notable geographic biases, as it categorizes various U.S. states based on perceived intelligence, desirability, and other traits. The researchers engaged OpenAI’s GPT-4o-mini model with over 20 million questions between March and May of last year, seeking to identify consistent classifications among locations.

The findings suggest that when pressured to make comparative judgments, the model ranked Massachusetts as the smartest state, Louisiana as the smelliest, and Ohio as the ugliest, among other assessments. Certain states, like Kentucky, West Virginia, and Mississippi, were identified as having the least intelligent populations, while Hawaii, Colorado, and New Hampshire were considered to have the most educated residents.

Matt Zook, a geography professor and co-author of the study, emphasized concerns over the normalization of negative stereotypes, stating, "We’re most concerned about how certain ideas get normalized, like the idea that people in Kentucky are stupider than anyplace else." The model's responses were found to reflect broader societal biases, with indications that wealthier and predominantly white communities were viewed as more favorable than others. This perpetuates narratives that align with existing prejudices in data sources.

The research also highlighted a disparity in how areas are perceived internationally, with Western countries classified as more desirable compared to sub-Saharan African nations. Safiya Noble, an academic at UCLA, criticized the inherent biases, noting that historical racism and classism are mirrored in the AI's training data. In response to the study, OpenAI contended that it is instrumental in minimizing subjective comparisons, stating that the research used an outdated model and asserting that ChatGPT aims to avoid reinforcing stereotypes. Nonetheless, Zook suggests that inherent biases may still influence the model's outputs, impacting real-world perceptions and decisions.