ChatGPT ranks US states by intelligence and smell in new study

Forcing the AI to choose

Researchers from the University of Oxford and the University of Kentucky conducted a large-scale study to see how ChatGPT views different geographic locations. They wanted to see if the model held internal biases about which places are smarter, uglier, or less desirable. To get clear results, the team asked the GPT-4o-mini model over 20 million questions between March and May 2025.

The researchers discovered that while the AI might refuse to answer a direct question about whether people in a specific state are stupid, it will often provide an answer if forced to choose between two options. This method of asking for a single, unambiguous consistent answer allowed the researchers to map out the model's internal rankings. By backing the AI into a corner, they were able to reveal underlying classifications that the model usually hides.

The study found that the more certain states appeared in the model's answers, the higher they were ranked regarding intellectual, creative, and physical attributes. This suggests that the AI has a hierarchy of locations based on the data it was trained on. As a result, the model consistently assigned different levels of desirability to various states based on these forced comparisons.

Mapping geographic stereotypes

The results of the study highlighted several specific geographic stereotypes that the model reinforced. For example, the AI considered Massachusetts to be the smartest state in the country. In contrast, it ranked Louisiana as the smelliest state and Ohio as the ugliest. It also identified North Dakota as having the least sexy people.

When looking at intelligence specifically, the model ranked Kentucky, West Virginia, and Mississippi as the top three states with the most "stupid" people. Conversely, it identified Hawaii, Colorado, and New Hampshire as the states with the least stupid people. These rankings reflect dominant narratives that often exist in public discourse but are not necessarily based on objective facts.

The research also extended beyond the United States to show global biases. The model ranked the US and Western Europe as having more desirable traits, such as a smarter and more stylish populace, compared to sub-Saharan African countries. The study suggests that the AI is reflecting a world view where whiter, richer, and less-immigrant communities are portrayed as more beautiful, less smelly, and smarter.

How training data creates bias

The researchers explain that these results are not a reflection of reality but rather a reflection of the data used to train the AI. Because the model learns from massive amounts of internet text, it absorbs the biases, prejudices, and stereotypes present in those materials. This means the AI is essentially mirroring the long histories of racism and classism found in its training set.

Matt Zook, a geography professor and co-author of the research, noted that the model is reinforcing what it has learned from its data. He expressed concern about how these ideas can become normalized through AI interactions. If the AI consistently associates certain locations with negative traits, it can reinforce the idea that those places are inherently different or inferior.

The study suggests that these biases are fundamentally structural and difficult to remove completely. Even if the model is designed to be neutral, the underlying data contains geopolitical hierarchies. This makes it difficult for the model to provide truly objective answers when comparing different cultures or regions that have been historically marginalized.

Real-world impacts on users

While these rankings might seem like simple trivia, the researchers warn that they could have serious real-world consequences. They suggest that these biases might be better hidden in future models through specific keywords, but the underlying logic will likely remain. For example, a user might not ask the AI if a state is "stupid," but they might ask for the best places to recruit software developers or medical doctors.

If the AI uses its biased internal rankings to answer such questions, it could shape how people structure their job searches or hiring processes. This could lead to people from certain states being overlooked for opportunities based on the AI's internal stereotypes. These subjective scripts can influence how people perceive the capabilities of different groups of people without them realizing it.

The study highlights the danger of AI becoming a tool that automates and scales human prejudice. As these models become more integrated into daily life, the way they categorize people and places can influence economic and social outcomes. The researchers argue that as these tools become more common, the fact that they are shaping human decisions becomes even more problematic.

OpenAI's response to findings

In response to the study, OpenAI stated that the researchers used an outdated model to reach their conclusions. The company maintains that ChatGPT is designed to be objective by default and is built to avoid endorsing stereotypes. They argued that the research, which relied on forced-choice prompts, does not reflect how the model is typically used by the general public today.

However, OpenAI did acknowledge that they are continuously working to improve how the model handles subjective or non-representative comparisons. They are looking for ways to reduce the subjectivity that can appear in the AI's outputs. Despite these efforts, the researchers suggest that there is still significant work to be done to address the biases baked into the data.

The study was published in the journal Platforms & Society and provides a framework for understanding how AI can perpetuate social inequalities. Readers interested in seeing how their own locations are classified can visit the researchers' website, inequalities.ai. This site allows users to look up exactly how the model classifies their city, state, and country based on various attributes.