Terrain traversability mapping is a geospatial problem with direct relevance to off-road mobility assessment, route planning, crisis response, and spatial decision support. Although machine learning has become an increasingly attractive complement to rule-based GIS procedures, model quality in this domain is still commonly summarized by aggregate error statistics alone. This paper examines uncertainty in AI-based terrain traversability prediction and argues that global accuracy is not sufficient for evaluating the usefulness of map products. The study estimates an index of passability (IOP) from structured geospatial attributes aggregated to 100 m × 100 m grid cells. The learning dataset contains 236,617 records described by 115 non-empty features derived from a larger attribute structure affected by missing values. A compact artificial neural network was used as a regression model and achieved good overall performance, with BIAS = 0.0024, MAE = 0.0097, and RMSE = 0.015 on a held-out test subset. However, spatial inspection of the results shows that prediction error is not randomly distributed. The largest deviations occur in terrain contexts that are absent, simplified, or weakly represented in the reference geodata, including mining areas and wetlands. The paper therefore frames uncertainty in traversability mapping as a compound outcome of model behaviour, thematic incompleteness of source data, and limited spatiotemporal representativeness of the training sample. By shifting attention from model novelty to data-driven reliability, the study contributes a more transparent perspective to GeoAI research in the open geospatial domain.