ChatGPT Ranked Every State By How Stupid And Smelly They Are, And The Implications Go So Deep
It will not tell you if you ask it straight — but corner it with 20 million forced choices and ChatGPT develops some very firm opinions about your state. A study from the University of Oxford and the University of Kentucky has mapped the geographic stereotypes hiding inside one of the world’s most-used AI models, and the results are equal parts funny and uncomfortable. It belongs on the same shelf as my roundup of historical “facts” everyone believes that are not true: things that sound authoritative, repeated everywhere, and wrong in an instructive way.
The Experiment: 20 Million Questions The Model Could Not Dodge
Ask ChatGPT directly which state’s people are stupid or smelly and it will politely refuse — the guardrails work fine on obvious questions. So researchers at the Oxford Internet Institute and the University of Kentucky changed tactics. Between March and May 2025 they put more than 20 million questions to OpenAI’s GPT-4o-mini model, each one a forced choice between two places: which is smarter, which is safer, which is more beautiful, which is smellier. With no ‘neither’ option available, the model had to pick — and pick it did, consistently enough that the team could rank states, countries and even city neighbourhoods by the traits the model quietly assigns to them.
The State Rankings Nobody Asked For
The results are best read as a mirror of the internet’s oldest regional jokes, now wearing a lab coat. In the model’s forced choices, Massachusetts came out as the smartest state. Louisiana was ranked the smelliest, Ohio the ugliest, and North Dakota the least sexy. Asked where people are stupider, the model most often pointed to Kentucky, West Virginia and Mississippi; asked where people are least stupid, it favoured Hawaii, Colorado and New Hampshire. None of this measures anything real about the people living in those states — it measures what the model’s training data repeats about them, which is exactly the researchers’ point.
The Same Pattern, All Over The Map
The state rankings are the funny version of a serious global finding. Across countries, the model systematically favoured wealthier, Western places — the United States, Western Europe and parts of East Asia were repeatedly chosen as smarter, happier, safer and more innovative, while much of Africa, the Middle East, Latin America and parts of Asia sank to the bottom. Zooming in did not help: neighbourhood-level rankings in London, New York and Rio de Janeiro tracked those cities’ existing social and racial divides rather than anything measurable about the communities themselves.
Five Biases Behind The ‘Silicon Gaze’
The team — Francisco Kerche and Mark Graham at Oxford, with Kentucky geographer Matthew Zook — published the work in the journal Platforms and Society under the title ‘The Silicon Gaze’, and they identify five interlocking biases that produce it. Availability: English-language, wealthy places produce far more text for models to learn from. Pattern: the model absorbs the stereotypes already sitting in that text. Averaging: complicated places get flattened into simple reputations. Trope: those reputations harden into stock characters. And proxy: stand-ins like wealth quietly become measures of worth. Their argument is that these are not bugs awaiting a patch but structural features of how the technology learns.
Why A Silly Ranking Is Not Really Silly
The researchers’ worry is normalisation. If a widely used assistant consistently treats some places as smarter and others as backwards, those assumptions seep into the stories it writes, the descriptions it drafts and the suggestions it makes — and the original reporting found the bias surfacing in something as ordinary as generated career stories set in different states. Multiply that by hiring tools, search summaries and automated screening, and a stereotype stops being a joke and becomes an input to real decisions. OpenAI’s position, for its part, is that the study tested an older model and that the company works to reduce this kind of subjectivity in its systems.
What The Researchers Want Done About It
The team’s prescription is sunlight rather than panic: clearer disclosure from developers about where training data comes from and how systems are designed, plus independent auditing frameworks so outsiders can test model behaviour the way this study did. Transparency, as Zook’s group puts it in substance, exposes bias without erasing it — someone still has to act on what the audits find. Until then, the study is a useful reminder for every reader of AI answers, including this one: a confident ranking is not a measurement, and where a model was trained shapes what it believes about where you live.
The next time an AI assistant ranks anything for you — cities, schools, neighbourhoods, people — it is worth remembering this experiment: the confidence is real, but so is the inheritance of every stereotype the internet ever typed. And for a study with a happier headline, our news shelf also carries the remarkable one about antiviral chewing gum and cancer-risk research — proof that not every laboratory result needs a warning label.
