Toward Causal Reasoning in GeoAI : A Three‐Tier Framework for Spatial Decision Support

ABSTRACT As AI has increasingly shifted toward a generative paradigm with large‐scale foundation models, GeoAI has likewise advanced rapidly in spatial prediction, representation learning, and workflow automation across spatial tasks. However, the identification of causal mechanisms in spatial systems remains comparatively underdeveloped. This limitation matters because spatial decision support often requires answers to intervention questions: what would happen if an exposure, policy, infrastructure investment, or emergency response were changed, and under what assumptions that answer is credible. This paper synthesizes recent developments at the intersection of GeoAI and causal inference and proposes a three‐tier framework for causal GeoAI organized by the depth at which geographic information is integrated into causal reasoning: Geo as Tool, Geo as Knowledge, and Geo as Representation. In this framework, “Geo” denotes the role of geographic information in causal reasoning, rather than GeoAI itself. The framework separates this integration depth from the technical pathways through which each tier can be implemented. We review how each tier supports causal discovery and causal effect estimation under three recurring spatial challenges: interference, spatial confounding, and heterogeneity across place and scale. We further identify new manifestations of classical spatial causal challenges under foundation‐model and agentic GeoAI settings, including automated identification errors, latent spatial confounding in embeddings, cross‐modal spatial support mismatch, infeasible generated counterfactuals, and transportability drift. The proposed framework clarifies the promise and limits of causal GeoAI: larger models and richer representations can support causal analysis, but they do not replace explicit identification assumptions, diagnostics, sensitivity analysis, and accountable decision‐making.