1st International Workshop on
Geography According to Foundation Models

By Prof. Dr. Johannes Scholz

University of Salzburg

Foundation models are rapidly reshaping how geography is represented, reasoned about, and operationalized in GeoAI. Recent evaluations show that large language models can infer topological spatial relations from geometries with moderate accuracy, but they still struggle with geometric fidelity, contextual nuance, and region-specific knowledge (Ji et al., 2025). At the same time, industry initiatives such as Google's Population Dynamics Foundation Model and trajectory-based mobility models demonstrate how multimodal geo-foundation models are becoming infrastructures for population analysis, mobility prediction, and climate-relevant decision support (Schottlander & Shekel, 2025). A growing body of systematic reviews highlights both the promise and the limitations of these models: while commercial LLMs can interpret geospatial concepts and generate functional code, they remain constrained by opaque training data, uneven global representation, and limited autonomy in complex spatial tasks (Dorobantu & Badea, 2026).

Against this backdrop, this talk examines how foundation models construct geographic knowledge — often over-privileging well-represented regions while being inaccurate or hallucinating about underrepresented places. We connect these representational biases to emerging benchmark results in geocoding, elevation estimation, and spatial reasoning, raising critical questions about whether and how LLMs should be treated as geographic knowledge generators. Complementing this, we foreground uncertainty as a cross-cutting challenge in GeoAI. Across domains such as wildfire risk, cattle movement, energy modeling, and supply-chain prediction, both aleatoric uncertainty (sensor noise, measurement error, stochastic processes) and epistemic uncertainty (model structure, data sparsity, domain shift) impact model outputs and their downstream use. Strategies such as probabilistic modeling, ensemble learning, and spatially explicit uncertainty mapping offer promising mitigation pathways — but they must be adapted to the new realities of foundation-model pipelines, where uncertainty is often hidden behind fluent text or single deterministic predictions. By bringing together conceptual, empirical, and applied perspectives, this talk aims to articulate a shared research agenda for equitable and trustworthy geo-foundation models.