AI hiking routes keep looking great on satellite view and terrible on the ground — are we training these models on only our good trips?
retained text
Got back from a soggy weekend on what was supposed to be a loop near Mount Hood. The AI-generated route looked great on satellite view. In reality it involved bushwhacking through devil's club and a creek crossing that would need a rope bridge. The app rated it moderate with notes about wildflowers. There were no wildflowers. Just wet. I keep testing these tools because the idea is solid: feed terrain data, trail reports, weather patterns, get a route that adapts to your group. What actually happens is the model smooths over complexity until it looks hikeable on a phone screen. Elevation gain gets rounded down. Unmaintained trails stay in the database years after they wash out. The confidence interval is invisible. This mirrors the board game prototyping I've been doing. The AI generates something that looks coherent but falls apart when you actually play it. But with hiking the stakes are different. A broken game mechanic is funny. A broken trail recommendation at dusk with tired kids is not. The part I keep coming back to: these models are trained on human reports, GPS tracks, photos. So where does the optimism come from? Is it us? Are we only uploading the good trips?
document fetched 2026-10-06 00:20 · canonical https://reddit.com/r/artificial/comments/1wyfkvw/ai_hiking_routes_keep_looking_great_on_satellite