New Study Maps Geographic Bias in Generative AI Models

A new study from researchers at the University of California, Santa Barbara and the University of Wisconsin–Madison systematically reviews geographic bias in AI evaluation, spanning both pre-generative and generative AI periods. The authors, led by Zilong Liu, identify multiple forms of bias: representation bias in training data (e.g., under-sampling certain regions), regional disparities in language models' factual recall (e.g., knowing more about Western cities than African villages), and a tendency for generative AI to over-proportionally favor prototypical places—what they call 'defaults' (e.g., generating images of skyscrapers when asked for 'city'). The paper, accepted for the book chapter "Geography According to ChatGPT," emphasizes that these biases are not mere statistical artifacts but can amplify social inequality and distort downstream applications like disaster response or biodiversity assessments.

The study highlights how recent research addresses geographic diversity by evaluating outputs across cognitive levels, parameter settings, and modalities. For instance, adjusting model parameters can reduce over-reliance on defaults, but no standardized benchmark exists yet. The authors argue that geographic bias remains underexplored compared to other forms of bias (e.g., racial or gender), despite its profound implications for global deployment. They call for the development of measurable diversity metrics and urge the AI community to treat geographic bias with the same rigor as other fairness concerns. As foundation models become ubiquitous, ignoring where AI ‘thinks’ it is could lead to systemic distortions—from misallocating aid to perpetuating cultural stereotypes.