Street View AI Often Just Repeats What Cities Already Know
Your city may already know more than the AI staring at street photos.
A new study asks a blunt question: when AI looks at a street photo, does it learn anything your city hall doesn't already know? Researcher Kaizhen Tan compared predictions from three AI vision-language models — systems that can describe and reason about pictures — against seven urban traits, using five public data sources. The traits ranged from road damage and curb ramps to house prices, population, building type and building height.
For several of these, existing records matched or beat the photos. Official data was just as good at spotting road damage and curb ramps, and nearly as accurate on population. House prices were better predicted from existing records too. Photos genuinely added value for three things: what a building is used for (office, shop, home), what type of building it is, and how many floors low-rise buildings have.
Distance and framing mattered a lot. The photo advantage grew by 5.7 percentage points for every doubling of distance to the nearest labelled building — meaning images help most in places with thin records. It shrank for tall buildings, whose rooflines often fall outside the camera frame.
There's a catch worth flagging. When a photo and an official record disagreed, the models often just went with the record — so "AI accuracy" partly reflects copying, not seeing. The study also found one popular building-height dataset was itself built from OpenStreetMap floor values, making those comparisons somewhat circular. The practical takeaway: street-view AI is a gap-filler, not a replacement. Cities, insurers and mapmakers should check what data they already own before paying to analyze millions of photos. And anyone using AI-generated maps should ask where the "truth" came from — if the labels and the official records share an origin, the impressive score may just be an echo.
- For road damage, curb ramps, house prices and population, existing city records were as good as or better than AI reading street photos.
- Photos won on building type, building use and low-rise floor counts — and got 5.7 percentage points more useful for every doubling of distance from labelled buildings.
- When a photo and an official record disagreed, the AI often just repeated the official record, so accuracy scores can be misleading.
Why It Matters
It shows cities and companies when street-photo AI is worth paying for — and when it just echoes old records.