Determining whether a particular driver is driving properly seems overkill. Any correlation between driving behavior and insurance payouts seems like it would be sufficient to improve the accuracy of your risk scores.

Let's say you only analyze vehicle speed data. That is clearly insufficient for training self-driving.

But it is very likely there is some correlation between speed and insurance payouts. For example, say you are able to segment your customer base including one subset with drivers who typically drive >80 miles an hour and another subset with drivers who typically drive <55 miles per hour. Are you saying that you would predict that the two populations have identical risk profiles? That speed data has no correlation with risk?

It's ok we'll just feed your speeds, GPS locations, braking data, and acceleration data into a black box AI model and have it spit out a bill multiplier.

All hail AI.