💻 computer science

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

This paper evaluates the geographic diversity of AI image generation by adapting ecological species diversity measures to reveal that older models and prompt revisions often yield greater diversity than newer models, while exposing a concerning homogeneity in how current systems stereotypically represent specific places.

Original authors: Zilong Liu, Krzysztof Janowicz, Mina Karimi

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zilong Liu, Krzysztof Janowicz, Mina Karimi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a group of very smart, very creative robots. You ask them to draw a picture of a specific city, like Vienna. You might expect that if you ask 30 different robots, you'd get 30 different views: some showing a park, some a museum, some a busy street, and some a famous church.

But this paper asks: What if they all draw the exact same thing?

The authors of this study decided to test this by asking various AI models (like DALL·E and GPT-4o) to generate images of Vienna. They didn't just look at the pictures; they looked at the ideas behind the pictures to see if the AI was being creative or just repeating a broken record.

Here is the breakdown of their findings, using simple analogies:

1. The "Tourist Brochure" Problem

When the researchers asked the AI to draw Vienna, almost every single model immediately thought of the St. Stephen's Cathedral. It was as if every robot had the exact same tourist brochure in its head.

The Finding: The AI models were incredibly "stereotypical." They didn't show the whole city; they only showed the most famous, iconic landmarks (like the Opera House or the City Hall).
The Metaphor: Imagine asking 100 people to describe a forest. If 90 of them only mention the tallest pine tree and ignore the ferns, the rivers, and the rocks, they aren't really describing the forest; they are just describing the "idea" of a pine tree. The AI was doing the same thing with Vienna.

2. The "Old vs. New" Surprise

Usually, we think newer technology is always better. We assume the newest AI model would be the most creative and diverse.

The Finding: The authors found something counterintuitive. The older AI models (like DALL·E 2) actually produced a more diverse set of images than the newest ones. The newest models were actually more stuck in their ways, focusing heavily on the same few landmarks.
The Metaphor: It's like a music playlist. You might expect the "Premium 2026" version of a music app to have the most variety. But in this case, the "Premium" version only played the top 3 hits over and over, while the "Basic 2022" version actually played a wider mix of songs.

3. The "Word vs. Picture" Gap

The study looked at two steps in the process:

The Prompt: The text the AI writes to describe what it's going to draw.
The Image: The actual picture it generates.

The Finding: The AI was more diverse when it was writing the description than when it was drawing the picture.
The Metaphor: Imagine a chef who writes a menu with 20 different exotic dishes (the prompt), but when they actually cook the meal, they only serve the same three basic dishes (the image). The "idea" was diverse, but the "reality" was not.

4. Counting Apples and Oranges (The Math Part)

To measure this "diversity," the researchers borrowed a tool from ecology (the study of nature).

The Old Way (Hill Number): If you have a forest with 10 trees, and they are all different species, that's high diversity. If you have 10 trees and 9 are pines and 1 is an oak, that's low diversity.
The New Way (Leinster-Cobbold Number): The authors realized that just counting "different" things isn't enough. What if you have 10 trees, but 5 are pines and 5 are spruces? They are different species, but they are very similar to each other.
The Insight: When they used this "similarity" math, the diversity scores dropped even lower. It turned out that even when the AI picked different landmarks, they were often just different versions of the same type of building (e.g., different churches).
The Metaphor: If you ask a child to pick 5 different fruits, and they pick a Granny Smith, a Fuji, and a Gala apple, they might say, "I picked 3 different fruits!" But an expert knows they are all just apples. The AI was picking "different" landmarks that were actually just different "apples."

5. Why Should We Care?

The authors argue that this isn't just a technical glitch; it's a form of bias.

If an AI always shows you the same few landmarks when you ask about a city, it creates a narrow, stereotypical view of that place.
It risks making the rest of the city (the local neighborhoods, the smaller parks, the unique local architecture) invisible.
The Conclusion: As AI gets better at making pictures, it might actually get worse at showing us the true, messy, diverse reality of the world. It's becoming a "hall of mirrors" that only reflects the most famous things back at us.

In short: The paper warns that our AI image generators are becoming "echo chambers" for geography. They are getting so good at following the crowd that they are forgetting to show us the unique, diverse, and less famous parts of the world.

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

Problem Statement
While the lack of diversity in Generative AI (GenAI) is often framed as an ethical issue, this paper argues it is fundamentally a geographic information science (GIScience) problem rooted in uncertainty and cognitive bias. Previous research has quantified geographic diversity in text-based AI outputs (e.g., chatbots naming countries) using information-theoretic measures like the Hill number. However, a significant gap exists regarding AI image generation. Unlike text, which can be tokenized, images contain rich, high-dimensional information that requires semantic segmentation to identify constituent objects. Current operationalizations of geographic diversity have not been applied to the multimodal nature of text-to-image (T2I) generation, nor have they accounted for the similarity between geographic entities (e.g., different types of landmarks) within the generated outputs.

Methodology
The authors propose a framework to measure geographic diversity in AI image generation, using Vienna, Austria, as a case study. The methodology involves three primary stages:

Experimental Setup and Data Generation:
- The study utilizes OpenAI's multimodal ecosystem, specifically the GPT-4o large language model (LLM) acting as an agent to revise user prompts before calling T2I models (GPT Image 1, GPT Image 1 Mini, GPT Image 1.5), as well as standalone T2I models (DALL·E 2, DALL·E 3).
- A "Generate an image of Vienna" prompt was submitted across 30 independent sessions for each model configuration.
- The workflow captures both the revised prompts (processed by the LLM) and the generated images.
Semantic Annotation and Similarity Computation:
- Landmark Identification: Researchers manually identified the primary landmark in each revised prompt (defined as the first-mentioned landmark) and each generated image (defined as the foreground landmark).
- Knowledge Base Mapping: Identified landmarks were mapped to Wikidata entities.
- Similarity Weighting: To address the limitation of standard diversity measures that treat all distinct entities as equally different, the authors computed "place-type similarity." Using the Rada distance (shortest path length in a semantic network) within Wikidata, they calculated the similarity between landmark pairs. For example, two churches are deemed more similar than a church and an opera house.
Diversity Measurement:
- Hill Number ( $^qD$ ): The standard measure of diversity (richness and evenness) was applied, treating landmarks as distinct species.
- Leinster-Cobbold Number ( $^qD_Z$ ): A similarity-sensitive extension of the Hill number was applied. This metric incorporates a similarity matrix ( $Z$ ) derived from the Rada distances. It posits that a community of highly similar species (or landmarks) is less diverse than one with highly distinct species, even if the count and distribution are identical.
- Diversity Profiles: Results were visualized across different orders of $q$ (where $q=0$ emphasizes richness, $q=1$ Shannon entropy, and $q=2$ emphasizes evenness).

Key Contributions

Multimodal Evaluation Framework: The paper demonstrates that assessing geographic diversity in GenAI requires a multi-stage approach that accounts for both prompt revision (LLM processing) and image generation (T2I processing), highlighting the role of multi-agent systems.
Similarity-Sensitive Metrics: The authors introduce the application of the Leinster-Cobbold number to geographic diversity. By integrating auxiliary knowledge bases (Wikidata) to compute semantic similarity, they provide a more holistic measure that penalizes the over-representation of semantically similar features (e.g., multiple churches) more heavily than standard metrics.
Empirical Case Study: The study provides the first empirical analysis of geographic diversity in AI-generated images, moving beyond text-based benchmarks to visual representations of specific places.

Results
The analysis yielded several counterintuitive findings:

Model Performance vs. Diversity: Newer models do not necessarily exhibit higher geographic diversity. For instance, GPT Image 1 (and its Mini variant) consistently showed the lowest diversity in image generation. Conversely, the older DALL·E 2 model exhibited higher diversity (specifically in evenness) than newer counterparts when similarity was ignored.
Prompt vs. Image Diversity: The revised prompts generated by the LLMs consistently exhibited greater geographic diversity than the final images produced by the T2I models. This suggests that the image generation stage introduces a bottleneck or homogenization effect.
Prototypicality and Homogeneity: All models displayed a strong "prototypicality effect," overwhelmingly favoring St. Stephen's Cathedral. While the specific distribution of other landmarks varied slightly between models, the lack of diversity was explicit and shared.
Impact of Similarity: The Leinster-Cobbold numbers were consistently lower than the Hill numbers. This indicates that while models might generate a variety of named landmarks, these landmarks are often semantically similar (e.g., different types of churches or palaces), leading to a "diversity collapse" when semantic similarity is accounted for.
Validity Rates: Newer models (GPT Image 1.5, GPT Image 1, GPT Image 1 Mini) achieved 100% validity in generating images with identifiable landmarks, whereas DALL·E 2 had the lowest rate (67%). However, the validity of revised prompts was lower than that of images, suggesting T2I models can generate visual descriptions of landmarks without explicit linguistic cues in the prompt.

Significance and Claims
The paper claims that geographic diversity in GenAI is not merely about the number of places mentioned but also about the platial representations—how specific places are depicted and the semantic variety of those depictions.

Beyond Ethics: The authors frame diversity as a function of uncertainty and a manifestation of cognitive bias (specifically, a reliance on fixed, prototypical geographic categories).
Risk of Stereotyping: The observed homogeneity risks producing stereotypical representations of places, where complex cities are reduced to a narrow set of iconic landmarks.
Methodological Advancement: By showing that the Leinster-Cobbold number reveals a deeper lack of diversity than the Hill number, the paper argues that future evaluations of GenAI must incorporate semantic similarity to accurately assess representational bias.
Call to Action: The authors call for systematic, place-centric case studies that span various geographic scales and modalities to better understand and quantify biases in generative (Geo)AI systems, drawing on literature from uncertainty quantification and environmental perception.