Evaluating Geographic Inference as a Privacy Proxy in Vision–Language Models with Multi-Frame Visual Inputs

Large vision–language models (VLMs) can infer geographic information from visual content even when explicit metadata are absent. We evaluate this capability as a privacy-relevant proxy: country, sub-national, city, and coordinate accuracy measure potential location exposure, not realized privacy harm. VideoGeoEval is a 1000-sequence diagnostic benchmark constructed from BDD100K driving videos and CityGuessr multi-view samples. We query four proprietary API models with one- and eight-frame inputs under Base, text-masked, blurred, and geographic-text-spoofed conditions. A containment-aware scorer enforces Country ≥ State ≥ City, and coordinate results satisfy the nested 50/200 km thresholds. Across three runs, eight-frame input improves city accuracy by 4.5–8.0 percentage points on CityGuessr and 5.5–7.1 points on BDD100K, with paired bootstrap confidence intervals excluding zero. Prior-normalized results indicate that the BDD100K gains cannot be fully explained by dominant-label priors. Spoofed overlays substantially reduce accuracy, but are treated only as controlled cross-modal stress tests. The benchmark therefore characterizes potential location exposure under controlled queries while separating measured accuracy from downstream privacy consequences.