TLDR: The paper introduces GeoBS, an information-theoretic framework to quantify geographic bias (geo-bias) in AI models. Unlike previous model-specific or spatially implicit methods, GeoBS offers a model-agnostic and spatially explicit way to evaluate geo-bias by comparing observed spatial performance patterns against unbiased reference patterns. It reinterprets existing metrics and proposes three novel scores (Scale-Grid SRE, Distance-Lag SRE, Direction-Sector SRE) to explicitly capture multi-scalability, distance decay, and anisotropy in bias. Experiments show both task-specific and foundation models exhibit geo-bias, highlighting the need for spatial fairness evaluation.
The rapid integration of Artificial Intelligence (AI) models, particularly powerful foundation models, into various aspects of our lives has brought immense benefits. However, this widespread adoption also raises critical ethical questions, especially concerning bias. While much attention has been given to social biases related to gender, race, or religion, a less explored but equally significant issue is geographic bias, or geo-bias.
Geo-bias refers to systematic disparities in how AI models perform across different geographic regions. Imagine an AI model that works flawlessly in one city but struggles significantly in another, or performs well in certain directions but poorly in others. This uneven performance can lead to unfair outcomes, impacting resource allocation, exacerbating social inequalities, and creating vulnerabilities in critical systems.
Previous attempts to quantify geo-bias often fell short. Many existing measures were designed for specific models or tasks, making it difficult to compare the fairness of different AI systems. Others were ‘spatially implicit,’ meaning they could tell you that bias existed, but not what specific spatial factors—like distance, direction, or scale—were contributing to it. This left researchers without a clear, universal framework to understand and address the root causes of geo-bias.
Introducing GeoBS: A New Framework for Geographic Bias Evaluation
To bridge these gaps, researchers have developed an innovative information-theoretic framework called GeoBS (Geo-Bias Scores). GeoBS provides a model-agnostic, universally applicable, and spatially explicit way to evaluate geo-bias in AI models. At its core, GeoBS treats the performance of an AI model across different locations as ‘spatial point patterns’ on the Earth’s surface. The more these observed patterns deviate from a predefined ‘unbiased’ or homogeneous pattern, the more geographically biased the model is considered to be.
The GeoBS framework offers a clear way to categorize geo-bias metrics based on three key factors:
- The Map Used: Whether the metric analyzes the distribution of data points (unmarked) or the model’s performance at those points (marked).
- The Reference Pattern: The ideal ‘unbiased’ spatial pattern against which the observed pattern is compared (e.g., a random distribution).
- The Difference Measure: The mathematical tool used to quantify how much the observed pattern differs from the reference pattern.
This framework can even reinterpret existing geo-bias measures. For instance, the widely used Unmarked SSI and Marked SSI scores, which measure dataset and model performance geo-bias respectively, can be understood within GeoBS as quantifying bias related to spatial proximity using an information-theoretic difference measure.
Novel Geo-Bias Scores: Unpacking Spatial Factors
Building on this framework, GeoBS introduces three novel geo-bias scores, known as Spatial Relative-Entropy (SRE) scores, which explicitly consider intricate spatial factors:
- Scale-Grid SRE: This score assesses ‘multi-scalability’ by dividing a region into equal-sized squares. It reveals if low-performing points concentrate at specific spatial scales.
- Distance-Lag SRE: Focusing on ‘distance decay,’ this score partitions a region into concentric rings. It helps determine if model performance changes systematically with increasing distance from a central point.
- Direction-Sector SRE: Addressing ‘anisotropy,’ this score divides a region into equal-angle sectors. It quantifies whether the model performs differently depending on the direction.
These SRE scores use Kullback–Leibler (KL) divergence, a concept from information theory, to measure the ‘information gap’ between the performance distribution in a specific spatial patch and the overall performance distribution of the region. This allows researchers to pinpoint exactly what kind of spatial factors are driving the observed geo-bias.
Also Read:
- Uncovering Hidden Biases in AI: Introducing humancompatible.detect
- Robust Preference Optimization: Enhancing LLM Alignment by Tackling Noisy Human Feedback
Experimental Insights
The researchers conducted extensive experiments across three tasks (geo-aware image classification, geo-aware image regression, and remote sensing image classification), eight datasets, and eight models. The findings were illuminating:
- Geo-bias is an independent dimension of evaluation. A model might have high overall accuracy but still suffer from significant geo-bias, underscoring the importance of evaluating spatial fairness alongside general performance.
- For task-specific GeoAI models, geo-bias often appears to be more dependent on the dataset itself rather than the specific model. This suggests that biases in data sampling or spatial coverage can heavily influence model fairness.
- For general-purpose foundation models, while they might be less affected by data geo-bias due to their massive training datasets, their powerful learning capabilities can sometimes lead them to overfit implicit geographical information. This means their susceptibility to geo-bias might depend more on their internal architecture and how well they learn spatial features.
The development of GeoBS not only advances our technical understanding of geographic bias but also lays a crucial foundation for integrating spatial fairness into the entire lifecycle of AI systems—from design and deployment to evaluation. To facilitate its adoption, the researchers have also implemented a user-friendly Python package called GeoBS for efficiently computing these scores. You can learn more about this research by reading the full paper available at arXiv:2509.23482.
