Is Spatial Bias a Form of Epistemic Injustice? Representation, Explainability, and Human Agency in Geospatial Artificial Intelligence

Geospatial artificial intelligence (GeoAI) increasingly mediates decisions about urban investment, environmental protection, disaster response, public health, agriculture, mobility, and the allocation of public resources. Its outputs are often treated as technically refined descriptions of spatial reality. Yet the data from which GeoAI learns are geographically uneven, institutionally selected, and shaped by unequal capacities to observe, record, classify, and contest. This work asks when spatial bias should be understood not merely as a technical defect but as a form of epistemic injustice. The argument proceeds in four stages. First, it characterizes GeoAI systems as epistemic infrastructures that participate in producing, rather than simply representing, geographical knowledge. Second, it distinguishes ordinary predictive error from geospatial epistemic injustice. Third, it proposes four jointly relevant conditions under which spatial bias acquires an epistemically unjust character: systematic representational asymmetry, displacement or devaluation of situated knowledge, decision-bearing authority, and a deficit of contestability. Fourth, it argues that conventional explainable-AI techniques are insufficient unless they are embedded in practices of geographical reason-giving, participatory validation, and institutional redress. The central conclusion is that fairness in GeoAI cannot be reduced to equalized model performance across regions. It requires protecting the agency of affected communities as contributors to, interpreters of, and challengers of spatial knowledge. The work develops a situated framework for responsible GeoAI based on representational adequacy, epistemic plurality, answerability, contestability, and proportionate human oversight.