New study finds older AI models more geographically diverse than newer ones

A new research paper accepted at AGILE 2026, titled 'Assessing the Geographic Diversity of AI's Platial Representations in Image Generation' by Zilong Liu, Krzysztof Janowicz, and Mina Karimi, examines how AI image generators like GPT and DALL-E represent different places around the world. The authors argue that geographic diversity is not just an ethical issue but also a matter of uncertainty and cognitive bias from a GIScience perspective. They developed information-theoretic diversity measures inspired by ecological species diversity, incorporating similarity weighting to evaluate how varied the generated images are across geographic contexts.

The study reveals several counterintuitive findings. Older models, despite producing lower-quality images, often exhibit greater geographic diversity than their newer counterparts. Additionally, prompt revision—adjusting text prompts before generating images—yields more geographic diversity than the image generation step itself. The researchers observed explicit model homogeneity, where the systems consistently depict the same prototypical features for specific places, risking stereotypical representations. This work highlights a pressing need to measure and improve geographic diversity across modalities as AI becomes increasingly multimodal in daily use.