I'm guessing the fleet is large enough that there are large sample sizes of: 1 - Pre-autopilot cars 2 - Autopilot cars with the feature turned on 3 - Autopilot cars with the feature turned off Since Tesla knows where you are, your speed, and whether or not your car is using autopilot I would think they could very quickly amass data to form a conclusion as to whether or not Autopilot is reducing accident rates (despite the fact that it may in some cases make its own error, or lull some drivers into unsafe behavior) overall. I would love to be a fly on the wall in whatever cubicle belongs to the Tesla data analyst(s) who are examining accident rates and their relationship (or lack of relationship) to the use of autopilot. Having now used it for a couple of days on freeways in So Cal my gut guess is that it will reduce accident rates in the fleet even in its current, early beta form. Despite its limited and imperfect "vision" and primitive decision making abilities, I wouldn't be surprised to learn that statistically, those weaknesses are overcome by its major strength - the fact that it never stops paying attention to its surroundings. It doesn't get distracted with a phone call, a cappuccino, a billboard, an attractive human being in a nearby car or a hundred other things - and it never gets road rage.