A new academic study has found that artificial intelligence systems can reproduce subjective and biased views about places, after ChatGPT judged Cambridge to be harder-working and happier than its historic rival Oxford. The findings, first reported by the Times UK, have reignited debate about the reliability of large language models and the risks of treating their outputs as objective fact.
According to the research, ChatGPT portrayed Cambridge as more relaxed, socially mobile and industrious, while depicting Oxford as more associated with alcohol, drug use and a livelier nightlife. The AI model also suggested Cambridge enjoyed cleaner air, better seafood and a faster-growing technology sector. Oxford, by contrast, was rated as more diverse, more religious, culturally richer in music, and offering wider culinary choices.
The conclusions emerged from a wider analysis of more than 20 million ChatGPT queries conducted by researchers at the Oxford Internet Institute. Academics asked the AI a series of inherently subjective questions and required it to respond with single-word answers, producing what the researchers described as stark, binary judgments. The Times UK reported that the study was designed to expose how AI systems absorb and reflect biases embedded in the data on which they are trained.
Mark Graham, professor of internet geography at the Oxford Internet Institute and a co-author of the paper, said the results highlight the deeper structural problem of bias in AI. “ChatGPT has been trained on the corpus of available human knowledge, which is characterised by already existing prejudices,” he said. “That’s what we’re seeing in the system, and that’s why it worries me.”
Beyond the Oxbridge rivalry, the study found broader global patterns. Wealthier regions in Western countries and parts of East Asia were consistently rated by ChatGPT as “better”, “happier” or “more innovative”, while large areas of Africa, the Middle East, Latin America and parts of Asia were more likely to receive negative rankings. Researchers warned that such skewed outputs could reinforce stereotypes if left unchallenged.
Some of the AI’s judgments were more provocative. ChatGPT described Oxford as “sluttier” than Cambridge, a characterisation Graham cited as evidence of how opaque and unaccountable AI decision-making can be. “As more people use these tools, their outputs are shaping books, media and how people think about the world,” he said. “That makes it critical to expose their subjective perspectives.”
Experts not involved in the study urged caution in interpreting the findings. Neil Lawrence, DeepMind professor of machine learning at Cambridge University, said forcing an AI to give one-word answers to complex social questions was problematic. While acknowledging that large language models can never fully represent the diversity of human opinion, he warned against treating them as if they were capable of accountable, human-like judgment.
OpenAI, the developer of ChatGPT, said the study relied on an older model and constrained it to one-word responses, which does not reflect typical real-world use. A spokesperson told the Times UK that addressing bias remains a priority, adding that newer models include additional safeguards and ongoing improvements, though challenges persist.
The findings underline a growing consensus among researchers: while AI systems can mirror human language impressively, they also mirror human prejudice. As the Oxford–Cambridge rivalry shows, even centuries-old debates can be reshaped — and distorted — by algorithms trained on the internet’s collective voice.
