ChatGPT's Geographic Stereotypes Exposed in New Research
A groundbreaking study has revealed that OpenAI's ChatGPT exhibits significant geographic biases, systematically ranking U.S. states and global regions based on stereotypical attributes including intelligence, attractiveness, and physical characteristics. The research demonstrates how artificial intelligence models can perpetuate and amplify human prejudices when making comparative judgments about different locations.
Forced-Choice Methodology Uncovers Hidden Biases
Researchers from the University of Oxford and University of Kentucky conducted an extensive analysis between March and May of last year, posing over 20 million questions to OpenAI's GPT-4o-mini model. By compelling the AI to choose between paired locations, the team uncovered consistent patterns of geographic stereotyping that might not emerge through direct questioning.
"When backed into a corner and forced to choose between pairs, the model started making harder choices," the researchers noted. This forced-choice methodology revealed that ChatGPT maintains internal rankings of states based on various attributes, with certain locations consistently appearing in specific categories.
State Rankings Reveal Troubling Patterns
The study identified clear hierarchies in ChatGPT's perceptions of different states. Massachusetts emerged as the state with the smartest population according to the AI model, while Louisiana was categorized as the smelliest. Ohio received the designation as the ugliest state, and North Dakota was ranked as having the least sexy residents.
When evaluating intellectual attributes, ChatGPT identified Kentucky, West Virginia, and Mississippi as having the "stupider" populations, while Hawaii, Colorado, and New Hampshire were considered to have the least stupid residents. These rankings demonstrate how the AI model has internalized and reproduces existing stereotypes about different regions.
Global Biases and Structural Inequalities
The research extended beyond U.S. borders, revealing that ChatGPT consistently ranked Western countries like the United States and Western European nations as having more desirable traits than sub-Saharan African countries. The AI model associated whiter, wealthier, less-immigrant communities with positive attributes like intelligence and beauty while assigning negative characteristics to more diverse populations.
"Long histories of racism and classism are reflected in the training data used for AI models," explained Safiya Noble, a professor at UCLA and author of "Algorithms of Oppression." "That's what the infographics are showing."
Real-World Implications and Concerns
Geography professor Matt Zook, co-author of the research, expressed particular concern about how these AI-generated stereotypes could become normalized. "We're most concerned about how certain ideas get normalized, like the idea that people in Kentucky are stupider than anyplace else," Zook stated. He emphasized that the model reinforces "dominant narratives about certain places being like this, certain places being like that."
The researchers demonstrated how these biases could translate into real-world consequences through informal testing. When asked to generate career stories, ChatGPT produced dramatically different narratives based on geographic origin: a man from Kentucky attended local technical college and secured a factory job, while a man from Hawaii attended a four-year college out of state and became an environmental engineer.
OpenAI's Response and Ongoing Challenges
When contacted about the findings, OpenAI disputed the methodology, noting that the researchers used an outdated model and forced-choice prompts that don't reflect typical ChatGPT usage. The company stated that "ChatGPT is designed to be objective by default and to avoid endorsing stereotypes" and that current models behave differently.
However, OpenAI acknowledged ongoing efforts to improve "how ChatGPT handles subjective or non-representative comparisons." Zook remains skeptical, suggesting that while geographic biases might become "better hidden" in future models through keyword filtering, "it's still going to be in there."
Broader Implications for AI Development
This study contributes to a growing body of research documenting biases in large language models. The findings highlight how AI systems can inadvertently perpetuate harmful stereotypes even when not explicitly programmed to do so, reflecting the prejudices present in their training data.
Zook warned about the subtle ways these biases could influence decision-making: "As these things get used in ways that we don't even realize, that's where it becomes even more problematic." The research suggests that AI models might indirectly shape perceptions through seemingly neutral queries about where to recruit professionals or which locations have desirable attributes.
The full research was published in the journal Platforms & Society, and the researchers have made their findings accessible through the inequalities.ai website, where users can explore how ChatGPT classifies their own locations based on various attributes.
