GeoAI Detective

How AI geolocates photos

AI image geolocation is the task of guessing where a photo was taken from the pixels alone — no GPS, no EXIF. Modern vision models are surprisingly good at it on easy cases and surprisingly bad on hard ones. Here is what they’re doing under the hood.

The cues a vision model reads

A vision model doesn’t “know” geography — it pattern-matches surface cues it learned from millions of captioned images: the side of the road people drive on, bollard and road-line styles, license-plate shapes, architectural vernacular, vegetation and climate, sun angle and shadows, written scripts and languages, even utility poles and antenna styles. It combines those into a best-guess region, much like a skilled human geo-guesser — but without a grounded world-model behind the inference.

Why it is sometimes confidently wrong

Its error is bimodal. On a famous landmark it can place the photo within a block from memory. But on a generic field, a back-alley, or a place built to look like somewhere else, the surface cues point the wrong way — and because the model has no calibrated sense of doubt, it commits to a confident answer that can land thousands of kilometres from the truth. A 2026 study (arXiv 2502.11163) also documented a systematic bias: accuracy fell for less-developed and sparsely-populated regions.

See it happen, one photo a day

GeoAI Detective turns those failure modes into a daily game: you and a pre-computed AI detective both guess where a real, human-curated photo was taken, and you watch the AI reason — and miss. Free, no signup, no AI-generated images. Play today’s case → Curious how often it misses big? See how good AI geolocation actually is.