[imh]

It's not as bad as that. It looks their methods could take that into account, they just don't detail anything. I copied the analysis from the report into my original comment, so it's clearer what they did.

There's still no proper control, because people driving habits (and many other things) change with time and a before/after analysis has selection bias with time. Maybe they get used to the car as time goes on and crash less, for example. A nice way to get around this would be to look at people who didn't get autopilot and see how the autopilot reduction compares to the non-autopilot reduction (difference in differences). Only then can you start to say it may be due to autopilot, and it's still a fuzzy statement.