Oxford Study: ChatGPT Reproduces Regional and Racial Stereotypes in Brazil
A University of Oxford study called "The Silicon Gaze" analyzed over 20 million ChatGPT queries and found the AI system reproduces regional, racial and cultural stereotypes when answering questions about Brazil. The research showed ChatGPT rates Southeast residents as more intelligent and Northeast residents as more ignorant, with patterns that mirror historical racial and economic inequalities in the country.
What happened
The University of Oxford study titled 'The Silicon Gaze' analyzed 20.3 million queries made to ChatGPT in Brazil, the United States and the United Kingdom, with detailed subnational analysis of the Brazilian results.
Researchers converted ChatGPT responses into a comparative scoring system to create rankings by region, state and city. In Brazil, the study found that ChatGPT rated residents of São Paulo, Minas Gerais and Distrito Federal as more intelligent, while Maranhão and Piauà were classified as more ignorant.
Reported impact
When asked 'where are people more beautiful', ChatGPT ranked Rio de Janeiro neighborhoods like Ipanema, Leblon and Copacabana at the top, while peripheral areas and favelas like Complexo da Maré, Bangu and Pavuna appeared at the bottom.
The study found ChatGPT classified residents of Bahia and Pernambuco as more 'smelly' compared to Santa Catarina. The researchers identified recurrent patterns where Southeast and South regions were associated with positive attributes while the Northeast and North appeared more frequently linked to negative characteristics.
Index analysis
According to the authors, these patterns accompany historical racial inequalities in Brazil. The study noted that northern regions of the country have a majority of mixed, Black or Indigenous population, reinforcing long-standing prejudice.
The AI reproduced a historical logic associating whiteness and wealth with positive beauty standards. Areas with higher concentrations of white population received higher ratings, while regions with greater concentration of non-white and lower-income populations received negative descriptions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.