New Safety Check Tells Cities When to Trust Crash-Prediction AI

Cities and states increasingly rely on AI to guess how bad a car crash will be. Those guesses shape real decisions: which intersections get rebuilt, where ambulances wait, which roads get extra patrols. The problem is that nobody could say, in hard numbers, how much to trust any single prediction. The data is messy. Police injury ratings agree with hospital records only about half the time, and the same model used in a new city or a new year is essentially guessing in the dark.

A new paper by Amir Rafe and Subasish Das offers a fix: a certification layer that snaps onto any existing prediction model without changing it. Think of it like a nutrition label or a warranty. Instead of one guess, the model now says something like "this crash is at least a B on the injury scale" — and the label states how often that claim will actually hold. The math works even when the underlying data is unreliable, biased, or from a place the model has never seen. It also handles severity-weighted risk, meaning it can focus on the crashes that hurt people most.

The honest twist is what the researchers proved about limits. Every certified prediction comes with an unavoidable vagueness, and no better model can shrink it below a certain floor. Testing on 5.2 million Texas crash records across four decades, the tool pinned down that floor for vulnerable road users — pedestrians, cyclists, motorcyclists — and flagged which remaining uncertainty is simply unknowable from the data.

The goal isn't perfect prediction. It's knowing exactly how much to trust a prediction before you bet a budget, an ambulance, or a life on it. The framework is released as free, open-source software.