Protected bike lanes improve safety ratings, but AI struggles with intersections The model identified several infrastructure characteristics associated with higher or lower safety ratings.

An AI system can produce a plausible road safety assessment even when some of the information used to generate that assessment is incorrect.

As the quality of the extracted attributes deteriorated, the safety model became less accurate.

Why 1,362 images from London cannot establish road safety across the United States The experiment's limited geographic coverage is one of its biggest weaknesses.

All 1,362 images came from just two London routes, covering urban streets with relatively specific infrastructure, road layouts, and traffic conditions.