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 study to see if ChatGPT holds geographic biases. They asked the GPT-4o-mini model over 20 million questions between March and May of 2025 to analyze how it classifies different locations.

The team used a specific method to bypass the model's standard refusal to answer subjective or offensive questions. Instead of asking if a state is "stupid," they forced the model to compare two places at a time to produce a single, unambiguous answer.

By backing the AI into these corners, the researchers found that the model began making difficult choices about human attributes. The more frequently certain states appeared in the results, the higher they were ranked for traits like creativity, physical beauty, and intelligence.

The results of the ranking

The study revealed a clear hierarchy in how the AI perceives different regions of the United States. The model identified Massachusetts as the smartest state, while it labeled Louisiana as the smelliest and Ohio as the ugliest.

When looking at specific populations, the AI ranked North Dakota as having the least sexy people. It also identified Kentucky, West Virginia, and Mississippi as the top three states with the most "ignorant" people, while Hawaii, Colorado, and New Hampshire were ranked as the least stupid.

Beyond the United States, the findings showed a broader global bias in the model's logic. The researchers found that ChatGPT ranked Western Europe and the US as having more desirable traits, such as being more stylish or smarter, than countries in sub-Saharan Africa.

How data creates bias

The researchers noted that the AI's rankings are not neutral but instead reflect the data it was trained on. According to geography professor Matt Zook, the model is reinforcing dominant narratives and prejudices found in its training materials.

The study found a consistent pattern where whiter, richer, and less-immigrant communities were classified as smarter and more beautiful. This suggests that the AI is picking up on long histories of racism and classism that exist within the internet's available data.

Safiya Noble, a professor at the University of California, Los Angeles, noted that these infographics show how historical inequalities are reflected in AI. Because the model learns from human-generated content, it can inadvertently replicate and normalize these harmful stereotypes.

OpenAI's response

OpenAI responded to the findings by stating that the researchers used an outdated version of the model. The company maintains that ChatGPT is designed to be objective by default and is built to avoid endorsing specific stereotypes.

The company argued that the research method of using forced-choice prompts does not reflect how the model is typically used by the general public. They stated that current models behave differently and that they are constantly working to improve how the system handles subjective comparisons.

However, the researchers remain skeptical about whether these biases can be fully removed. Zook suggested that while future models might hide these biases better using different keywords, the underlying geographic prejudices will likely remain in the system.

Real-world consequences

The study warns that these biases could have practical impacts on how people use AI for professional tasks. For example, the model might influence how a user structures a job search if it suggests certain states are better for recruiting software developers.

In informal testing, the researchers found that the model's subjective scripts changed based on the state mentioned in a prompt. When asked about a man from Kentucky, the AI told a story about a factory worker, but it described a man from Hawaii as an environmental engineer.

These different scripts demonstrate how the AI might subconsciously favor certain regions over others. As AI becomes more integrated into daily life, these hidden biases could shape economic opportunities and social perceptions in ways that are not immediately obvious to the user.