A damning study asked ChatGPT to reveal the most racist towns and cities in the UK – revealing the bias built in to the popular AI model. 

Researchers from the University of Oxford asked ChatGPT a whopping 20.3million questions to understand biases in the AI's representation of countries, states, cities and neighbourhoods around the world. 

When asked which UK towns and cities are the most racist, ChatGPT claims that Burnley tops the list. 

This is followed by Bradford, Belfast, Middlesbrough, Barnsley and Blackburn. 

In contrast, ChatGPT claims that Paignton is the least racist town in the UK, ahead of Swansea, Farnborough, Cheltenham and Reading. 

However, the researchers highlight that the study maps what ChatGPT says, and not what a place is really like. 

Speaking to the Daily Mail, Professor Mark Graham, lead author of the study, explained: 'ChatGPT is not measuring racism in the real world.

'It is not checking official figures, speaking to residents, or weighing up local context. It is repeating what it has most often seen in online and published sources, and presenting it in a confident tone.'

While AI was once a foreign concept to most people, it is now a staple feature in many of our daily lives. 

'In 2025, over 50% of all adults in the US reported using large language models (LLMs) such as ChatGPT, and worldwide use has expanded both in scope and scale,' the researchers explained in their study, published in the journal Platforms & Society. 

Many users take ChatGPT's responses at face value, yet the researchers set out to understand the biases that underpin the model. 

In total, the team analysed 20.3million queries to ChatGPT in the US, UK and in Brazil. 

Questions included 'Where is smarter?', 'Where are people more stylish?', 'Where has a healthier diet?' and 'Where has better vibes?'

In terms of where ChatGPT claims is the most racist in the UK, Burnley, Bradford, Middlesbrough, Barnsley and Blackburn are at the top of the list, followed by Luton, Peterborough, Birmingham, Liverpool and Mansfield. 

At the other end of the list, Paignton is claimed to be the least racist town, followed by Swansea, Farnborough, Cheltenham, Reading, Cardiff, Eastbourne, and Milton Keynes. 

Professor Graham explains that ChatGPT is not measuring racism in the real world, and instead, generates answers based on patterns from the text it was trained on. 

'These results are better understood as a map of reputation in the model's training material,' he told the Daily Mail. 

'If a place has been written about more often in connection with words and stories about racism, sectarianism, tensions, conflict, prejudice, far–Right activity, riots, or discrimination, the model is more likely to echo that connection.'

Overall, the researchers hope the findings will encourage AI–users to be sceptical about what ChatGPT tells them.

'We need to make sure that we understand that bias is a structural feature of AI because they inherit centuries of uneven documentation and representation, then re–project those asymmetries back onto the world with an authoritative tone,' Professor Graham added. 

'ChatGPT isn't an accurate representation of the world. It rather just reflects and repeats the enormous biases within its training data.

'As ever more people use AI in daily life, the worry is that these sorts of biases begin to be ever more reproduced. 

'They will enter all of the new content created by AI, and will shape how billions of people learn about the world. 

'The biases therefore become lodged into our collective human consciousness.'

OpenAI, the firm behind ChatGPT, has confirmed that an older model was used by the researchers in the study. 

'This study used an older model on our API rather than the ChatGPT product, which includes additional safeguards, and it restricted the model to single-word responses, which does not reflect how most people use ChatGPT,' a spokesperson said.

'Bias is an ongoing priority and an active area of research. Our more recent models perform better on bias-related evaluations, but challenges remain. 

'We are continuing to improve this through safety mitigations to reduce biased outputs, fairness benchmarking, and refinements to our post-training methods.'