New AI Reads Crash Reports and Suggests Safer Road Fixes

Right now, when someone is killed or badly hurt at an intersection, a traffic safety engineer has to read the crash report, figure out what went wrong, and then decide which fix to install — a new traffic light, a turn lane, better signs, a redesigned crossing. It's slow, it's expensive, and there aren't nearly enough people trained to do it. That means dangerous intersections can sit unfixed for years.

A team of researchers built an AI system that handles the first pass. It reads the written crash description — the "narrative" — and pulls out the important facts: was there a traffic signal, what was it showing, which driver was at fault, which way were the vehicles moving. Then, using a technique called RAG (letting AI look things up in a trusted library instead of guessing), it searches a database of fixes proven to work and picks the ones matching that intersection's specific problems. Think of it as a very well-read assistant who has memorized thousands of crash reports.

The test: 312 crashes involving deaths or serious injuries at 115 intersections in Lake and Sumter Counties, Florida. The AI's suggestions matched what experts had actually chosen about 82% of the time. It recommended roughly four fixes per intersection — almost exactly the number engineers had picked (3.86 on average). Not bad for a machine reading police reports.

The catch: the system was trained and tested on Florida crash data, so it may not translate cleanly to other states with different road designs and reporting styles. And it's a decision-support tool — a human engineer still signs off, because getting road safety wrong costs lives. The promise is speed and scale: agencies short on staff could screen hundreds of crash reports in hours instead of months, prioritizing the fixes that save the most lives per dollar.