ChatGPT consistently favors affluent Western locales in its responses across a wide spectrum of queries, ranging from subjective ones like “Where are the most beautiful people?” to seemingly objective ones like “Which country is safer?” This bias stems from inherent prejudices embedded within its training datasets, according to researchers from the University of Oxford (UK) and the University of Kentucky (USA).
The investigators analyzed 20 million prompts directed at ChatGPT. They observed that for categories implying superiority—such as “smarter,” “happier,” or “more innovative”—ChatGPT disproportionately uplifts the USA, Western Europe, and to some extent, East Asia. Conversely, nations in Africa, the Middle East, parts of Asia, and Latin America are significantly more likely to appear at the lower end of these rankings. These tendencies manifest across both subjective evaluations and those that appear to be based on objective metrics.
To vividly illustrate these discovered patterns, the researchers developed maps and comparisons based on their examination of 20.3 million queries. In the ranking generated in response to the prompt, “Where are the smartest people?” nearly all low-income nations, particularly those in Africa, ranked at the bottom. The findings related to specific areas within London, New York, and Rio de Janeiro demonstrated that ChatGPT’s rankings closely mirror existing societal and racial disparities rather than the actual characteristics of those communities, the scientists assert. The study’s authors even launched a website allowing users to compare different global regions based on various criteria within ChatGPT’s output.
“When AI is trained on biased data, it amplifies those biases and has the potential to disseminate them widely. Consequently, we require both transparency and independent oversight regarding the claims these systems make about people and places, urging users to approach these outputs with skepticism when forming opinions about communities. If an AI system repeatedly links specific countries, cities, or other localities with negative labels, these associations can quickly propagate and begin shaping perceptions, even if they are founded on incomplete, inaccurate, or outdated information,” commented Professor Mark Graham regarding the research.
Generative AI is increasingly integrated into fields like public services, education, business, and even routine daily decision-making. Treating its responses as definitive, neutral data creates a risk of reinforcing the very inequalities they reflect. Such biases are not easily remedied; they are structural characteristics of generative AI. Large Language Models learn from data shaped by centuries of uneven information production, which inherently privileges environments wealthy in English-language content and high digital access. The scientists identified five interconnected mechanisms contributing to this bias: accessibility, commonality, averaging, stereotyping, and the spillover effect. Collectively, these factors help explain why wealthier regions with vast knowledge reservoirs consistently score higher in ChatGPT’s responses.
The project’s authors advocate for increased openness from both AI developers and users, along with establishing standards for independently verifying model behavior. The general public must recognize that generative AI does not present an objective view of the world; its answers merely echo the biases embedded in its training corpora.
About the Author
