Detecting biases in GenAI models' perceptions of urban neighborhoods

Source citation:

• Bollen, P., Higton, J., & Sands, M. (in press 2026). Nationally representative, locally misaligned: The biases of Generative Artificial Intelligence in neighborhood perception. Political Analysis.

Bollen et al. tested how Generative Artificial Intelligence (GenAI) compares to human perspectives when evaluating urban neighborhoods, specifically using visual analysis. To do so, they prompted several leading GenAI models to evaluate a set of street-view images of Detroit, asking them to rate the images based on perceived wealth, safety, and disorder. They also asked a nationally representative group of 800 survey respondents and a locally representative group of over 2,400 survey respondents to rate the same photos on the same three criteria. The local responses were collected via a module in Wave 20 of the 2024 Detroit Metro Area Communities Study (DMACS). Available from the ICPSR Member Archive, DMACS began in 2016 and is an ongoing panel study that recruits a representative sample of Detroit residents, with a focus on reaching underrepresented populations. The GenAI models matched well the average opinions of the general US population, but they were significantly less accurate at reflecting the views of the DMACS respondents, who actually live in Detroit. Bollen et al. noted that AI models often lack training data from historically underrepresented populations who are often “hard-to-reach” for surveys. The GenAI models exhibited significant demographic and geographic biases regarding neighborhood safety and disorder in Detroit. The authors urged caution when using GenAI models for subjective assessments, especially in cases “where judgments based on individual images are important,” for example in police work. Look here for more publications using data from any study in the DMACS series.

Posted March 5, 2026