Ask an AI image model to picture "USA" and you'll mostly get cities.
Researchers tested two diffusion models — FLUX 1-schnell and Stable Diffusion 3.5-Large — by generating 150 street-view images each for every US state, every state capital, and a plain "USA" prompt. They then used a computer vision model called DINO-v2 to embed the images and measured visual distances with a metric called Fréchet Inception Distance. The pairwise clustering showed that neighboring states and capitals tended to group together, suggesting the models do carry some genuine geographic structure under the hood.
The problem shows up at the national scale. When the prompt was simply "USA," that regional diversity collapsed into what the authors call a metropolitan stereotype — frontier landscapes, deserts, tropical environments, rural towns, and small cities were statistically distant from what the models produced. In short: the models know the difference between Wyoming and Louisiana, but default to neither when asked to picture America.
That matters because these models are increasingly used for urban planning and scenario generation, not just art. A planning tool that treats every American city as a variation on the same dense downtown is baking a geographic blind spot into real decisions. It's a familiar pattern — earlier bias audits of image generators found similar skews around race and gender — and it suggests the problem of representational collapse is structural, not accidental.