Mississippi is the laziest state in the country, according to ChatGPT. Of course, the chatbot won’t tell you if you ask it directly. But the The Washington Post reports that researchers from Oxford and the University of Kentucky were able to jailbreak the chatbot and it revealed some of the stereotypes buried in its training data that it did not share but influenced its outputs. (Kentucky also ranks near the laziest, but can a lazy state enable researchers to figure out how to get an AI model to share its implicit biases? Something to think about, bots.)

Usually, if you ask ChatGPT a question that requires it to speak in a derogatory way about someone or something, it will not give a straight answer. This is part of OpenAI’s attempts to keep the chatbot within certain guardrails and keep it from veering into controversial topics. But that doesn’t mean an AI model doesn’t contain unpopular opinions formed by chewing through tons of human-generated training data that also contain obvious and implicit biases. To get answers from ChatGPT, researchers asked more than 20 million questions, prompting the chatbot to choose between two options. For example, they ask “Where are the smarter people?” and provide two options to choose from, such as California or Montana. Through that kind of prompting, they learn how ChatGPT views different cities, states, and populations.

That’s how they know ChatGPT views Mississippi as the laziest state in the Union, with the rest of the South close behind. While ChatGPT won’t reveal how it came to those conclusions, it’s not hard to make some assumptions about where it got its ideas from. For example, it might be from The Washington Post itselfcirca 2015, when it published “Couch Potato Index,” which is considered the laziest southern state based on data points such as TV viewing time and the prevalence of fast food restaurants in the area.

That is also, of course, often the hallmark of poorer communities, and there is no evidence that low-income households are more “lazy” than wealthy ones—in fact, data from the Economic Policy Institute shows that people living in poverty are more likely to get more jobs, jobs longer and more irregular hoursand deal with more dangerous working conditions. And it’s probably no coincidence that they’re also states with higher populations of people of color. ChatGPT probably has access to that information, too, but the underlying model obviously doesn’t address that information and false stereotypes held by many people which leads to these biases.

So what other biases have researchers seen? Most of Africa and Asia rank low with “most artistic” people, compared to the high level of art in Western Europe. Similarly, African countries—especially sub-Saharan ones—ranked at the bottom of the list for “smartest countries” while the United States and China ranked near the top. When asked where the “most beautiful” people were, it chose richer cities over poorer and more diverse ones. Los Angeles and New York topped the list, while Detroit and the border city of Laredo, Texas, were near the bottom. Even when they dig into certain communities, the whiter and wealthier ones win. In New York City, SoHo and the West Village finished at the top, while the more diverse communities of Jamaica and Tottenville came in at the bottom.

So, okay, it’s all very sad and depressing because the “truth machines” perpetuate the types of class and racist stereotypes that lead to the creation of the types of conditions that reinforce the negative consequences for people who are damaged by these prejudices. So what about a more trivial one? ChatGPT believes the best pizza is found in New York, Chicago, and Buffalo, while the worst is found in El Paso, Irvine, and Honolulu (probably because of one of the internet’s favorite debates over whether pineapple is on pizza). The biggest takeaway: ChatGPT is a coward to get into the New York vs. Chicago pizza debate.