I don't agree to the conclusion that AP is a net loss of safety based on three accidents as these three incidents represent a very very minute fraction of overall miles that AP has worked correctly and safely. Incorrectly applied seat belts or child seats that are incorrectly used also kill but that doesn't mean seat belts or child seats are net loss of safety.
For collisions, we don't have enough public data. However, we can judge safety as listed by
fatalities: one fatality per two billion miles when Autopilot is turned off, and one fatality per 130 million miles when Autopilot is turned on. This is fairly strong evidence that
Autopilot does not prevent fatalities.
Returning to collisions on "restricted roads": Elon has claimed the accident rate on Autopilot is halved. However, the normal collision rate for cars is roughly 400 per hundred million miles travelled (
https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812032). Half that (200) times the number of Autopilot miles driven (130 million) yields 260 collisions. Thus, even if this claim were true, we should have 260 collisions as data, but we've heard of only a handful. The other 255+ were not picked up by the press, for whatever reason (no police report, in a foreign country, driver doesn't blame Autopilot, Tesla non-disclosure agreements). Similarly, Autopilot could have prevented a huge number of collisions that we haven't heard about.
So, we have no way to judge, via statistics, whether Autopilot on all "restricted" roads prevents more collisions than non-Autopilot.
However, Tesla does have these statistics, and we do know that Tesla restricts Autopilot speeds on non-divided highways ("restricted" roads). Why? For safety, obviously. This is very, very strong evidence that Tesla believes Autopilot to be riskier on non-divided highways than on divided highways. If they had statistics that proved that Autopilot was as safe on restricted roads, they would have left AP speed up to the discretion of the driver, as they do on highways.
And we have corroborating evidence: we have seen two drivers on AP driving on restricted roads that did not prevent accidents which any normal driver could have prevented. The only AP fatality seen so far was on a highway in which vehicles are permitted to cross traffic. We have evidence that, in these accidents, Autopilot drove terribly. We have warnings like "There may be situations in which Traffic-Aware Cruise Control does not detect a vehicle, bicycle, or pedestrian" listed in the manual. And lastly, we have the owner's manual stating outright that "Autopilot is intended for use on divided highways."
Finally, we have not heard much about Autopilot accident prevention on non-divided highways. At best, we have one instance of AEBS activating, which is always on and distinct from Autopilot.