Assessing the Geographic Diversity of AI's Platial Representations in Image Generation

Researchers evaluated geographic diversity in AI image generation models (GPT and DALL-E), finding that these systems produce stereotypical representations of places due to underlying model homogeneity. The study reveals counterintuitive results: older models sometimes show greater geographic diversity despite lower image quality, and the systems consistently depict identical prototypical features for specific locations.

This research addresses a critical but underexamined problem in generative AI: the systematic bias toward particular geographic representations embedded in widely-used image generation models. While AI ethics typically focuses on demographic diversity, this study reframes geographic bias as both an uncertainty and cognitive bias issue through the lens of geographic information science, applying ecological diversity measurement principles to evaluate AI outputs.

The findings expose a troubling pattern where state-of-the-art models like DALL-E and GPT exhibit explicit homogeneity in their geographic representations. Rather than producing varied depictions of places, these systems gravitate toward stereotypical prototypes—the same iconic features or representations repeated across queries. This homogeneity becomes particularly concerning as generative AI becomes increasingly multimodal and integrated into daily information consumption, from education to tourism to urban planning applications.

The counterintuitive discovery that older models sometimes generate more geographically diverse outputs than newer, higher-fidelity versions suggests that optimization for image quality may inadvertently narrow geographic representation. Additionally, the finding that prompt revision generates greater diversity than image generation itself indicates that model architecture fundamentally constrains the range of possible geographic outputs.

For developers and organizations deploying these models, this research signals the need for deliberate diversification strategies in training data and evaluation metrics. The risk extends beyond representation—stereotypical geographic depictions can reinforce harmful place-based narratives, influence public perception of regions, and create feedback loops where AI-generated imagery shapes human understanding of global geography. Future model development should explicitly incorporate geographic diversity measurement as a performance criterion alongside traditional quality metrics.

• →State-of-the-art AI image generators exhibit homogeneous geographic representations, consistently depicting identical prototypical features for specific locations.
• →Older AI models sometimes demonstrate greater geographic diversity than newer models despite producing lower-quality images, suggesting a trade-off between fidelity and representation breadth.
• →Prompt revision yields greater geographic diversity than direct image generation, indicating fundamental architectural constraints in current models.
• →AI systems risk producing and reinforcing stereotypical geographic narratives that shape public perception of places and regions.
• →Geographic diversity should become an explicit evaluation metric in AI model development, alongside traditional quality and safety assessments.