AI Reliability Audit of AI-Generated Building Footprints — Abuja, Nigeria (2026)

Measuring when a deployed AI system can be trusted in high-stakes decisions. An independent, instance-level accuracy audit of a commercial GeoAI building-extraction tool against OpenStreetMap reference data over a defined 4.28 km² study area.

Core finding: the system generated 7.3× more building polygons than the baseline while failing to detect 43.6% of independently known buildings — aggregate coverage and per-instance reliability are distinct properties. Methods include greedy one-to-one IoU matching, threshold-sensitivity analysis, positional-offset statistics, and street-level ground-truth verification; findings are positioned against NIST AI RMF and EU AI Act Annex III.