Tesla could gets lots of quality data with the autopilot disengaged and running in a passive state such that the autopilot's calculations are compared to the hu

Tesla could gets lots of quality data with the autopilot disengaged and running in a passive state such that the autopilot's calculations are compared to the human driver's actions. Any large variances (such as the autopilot wanting to suddenly veer left when the driver didn't) would be very valuable. I've done this with experimental machine control algorithms a number of times.