Can an AI chatbot develop stereotypes about people based on where they live? A research project examining ChatGPT suggests that it can produce troubling regional comparisons when repeatedly forced to choose between locations.
The findings attracted attention after researchers used an older version of OpenAI’s model to compare US states and other places across subjective categories, including intelligence, attractiveness, and smell. The results were not measurements of residents’ actual abilities or characteristics. Instead, they revealed patterns in the AI model’s answers—and raised questions about the stereotypes those answers might reinforce.
What Did Researchers Discover?
Researchers affiliated with the University of Oxford and the University of Kentucky studied how OpenAI’s GPT-4o-mini model compared locations. Between March and May 2025, their experiment used more than 20 million prompts designed to produce consistent comparisons between pairs of places.
Instead of asking the model to make one sweeping ranking, the researchers repeatedly asked it to choose between two locations. By collecting the answers, they could identify patterns across states, cities, and countries.
According to reporting on the study, ChatGPT’s responses placed Massachusetts near the top for intelligence-related comparisons. Kentucky, West Virginia, and Mississippi appeared among the states associated with more negative intellectual labels. Louisiana was ranked poorly in smell-related comparisons, while Ohio was associated with negative attractiveness judgments.
These were AI-generated rankings, not scientific assessments of residents. There is no valid basis for concluding that people from one state are inherently smarter, smellier, or more attractive than people from another based on this experiment.
Why Did ChatGPT Make These Comparisons?
ChatGPT is trained on large collections of text and other data. Those materials can contain genuine information, but they can also reflect prejudice, jokes, historical discrimination, media stereotypes, and unequal representations of different communities.
When a model is repeatedly required to choose between two places, its answers may reflect associations it has learned from those materials. The result can make a stereotype appear more consistent or authoritative than it really is.
The researchers’ concern was not simply that an AI system could generate an insulting answer. It was that repeated patterns in its responses could normalize existing assumptions about particular communities.
The experiment does not establish a single, definitive cause for every ranking. It does, however, illustrate why subjective AI-generated comparisons need to be examined critically.
Why the Findings Raise Concerns About Discrimination
Geographic stereotypes can overlap with assumptions about race, income, education, and immigration. The study’s authors and other experts warned that AI systems may reproduce those wider social inequalities.
For example, if a chatbot repeatedly associates a state with lower intelligence, users could mistakenly treat that output as evidence about the abilities of people who live there. Such assumptions could influence decisions about employment, education, investment, or where to recruit professionals.
That would be particularly concerning if someone used an AI-generated stereotype as a shortcut for evaluating individuals.
A person’s state of residence does not establish their intelligence, character, professional ability, appearance, or personal hygiene. These qualities cannot responsibly be inferred from a chatbot’s subjective ranking.
Could AI Bias Affect Real-World Decisions?
Yes, potentially—especially when people trust automated systems without examining how their answers were produced.
Consider a hypothetical employer asking an AI assistant which states are best for finding talented software developers. If the model’s response is influenced by stereotypes rather than relevant evidence, it could encourage unfair assumptions about candidates from particular regions.
The same problem could arise in education, housing, marketing, and other areas where automated recommendations influence human decisions.
The risk is not limited to offensive language. A biased response can become consequential when someone uses it to make a decision about a real person.
Researchers therefore argue that AI systems should be evaluated for the patterns they produce, not only for whether they refuse obviously discriminatory questions.
What Did OpenAI Say About the Research?
OpenAI disputed the idea that the study’s results necessarily describe how ChatGPT behaves today.
In a statement reported by news outlets, the company said the researchers had used an older model. OpenAI also emphasized that ChatGPT is designed to be objective by default and to avoid endorsing stereotypes, adding that results from forced-choice prompts do not necessarily reflect typical use or the behavior of current models.
The distinction matters. The research examined GPT-4o-mini under a specific experimental setup; it should not automatically be treated as a definitive assessment of every ChatGPT model or every interaction.
At the same time, a model’s intended behavior does not make bias testing unnecessary. Independent evaluations can help identify weaknesses and track whether improvements reduce harmful patterns.
Can You Check How AI Ranks Your State or City?
The researchers created a public project called Inequalities.ai, where visitors can explore how the model compared locations across different attributes.
The project is useful as an example of how researchers can audit AI outputs at scale. However, its rankings should be understood as evidence about the model’s responses—not as verified facts about the places or people being compared.
If you explore the results, ask what the model was prompted to do, which model was tested, how the comparisons were collected, and whether the category being measured has an objective definition.
These questions are especially important when a ranking concerns subjective or potentially discriminatory traits.
How Users Can Interpret AI Rankings More Carefully
Readers can take a few practical steps when an AI tool produces a ranking about people or places:
- Check the source: Look for the original study and its methodology instead of relying only on a headline.
- Separate output from reality: A chatbot’s answer is not proof that its claim is true.
- Question subjective categories: Labels such as “stupid,” “smelly,” or “hottest” do not necessarily have objective, measurable definitions.
- Look for independent evidence: For factual comparisons, prefer relevant statistics and transparent research methods.
- Avoid using stereotypes to judge individuals: A geographic label says little about a specific person’s abilities or character.
- Consider the model and date: Findings from an older model may not accurately describe newer systems.
These habits are useful beyond ChatGPT. Any generative AI system can produce answers that sound confident while reflecting limitations in its data, design, or instructions.
Final Thoughts
The study behind ChatGPT’s state rankings highlights an important challenge in artificial intelligence: models can produce consistent answers that still reflect problematic social assumptions.
The rankings do not establish that residents of any state are inherently more intelligent, attractive, or hygienic than others. Instead, they offer a case study in how an AI model may reproduce stereotypes when pushed to make repeated subjective comparisons.
As AI tools become more involved in everyday decisions, transparency, independent testing, and careful interpretation will remain essential. AI-generated rankings should be treated as outputs to investigate—not as facts about people.
Related Coverage
The article is based on published research and reporting. You can use these references at the end of your DealInNow post.
- Original Research: The Silicon Gaze: A Typology of Biases and Inequality in LLMs
- University of Oxford: New Study Finds That ChatGPT Amplifies Global Inequalities
- BuzzFeed / HuffPost reporting: ChatGPT Ranked Every State by How Stupid and Smelly They Are
