The Limits of AI: Geospatial Foundation Models Do Not Understand Geography

Geospatial foundation models can now learn powerful spatial representations from remote sensing imagery, maps, trajectories, and multimodal Earth observation data. But spatial representation is not the same as geographic understanding.

This interactive Field Note asks a stricter question: can a model remain reliable across regions, scales, and distribution shifts while reasoning about spatial dependence, regional heterogeneity, and mechanism? The central claim is simple: space can be represented; geography must be reasoned about.

Open the interactive essay full screen ↗

Core claim: spatial representations can be learned; geographic understanding must be tested across scales, regions, and mechanisms.