Instead you need to count the number of times someone has actively taken control to avoid a collision, which is rare enough it’s hard to find examples.
Having an alert driver paying attention is a net benefit, but in practice people don’t continuously pay full attention for long stretches with autonomous driving. If you assume people are paying attention say 90% of the time in practice then their only catching less than 90% of the potential accidents they could avoid. Further, assuming identical rates of accidents to the average person the other car is at fault 1/2 the increase is further reduced.
As such, I think the regulation is a good idea for now as the net benefit to fully autonomous driving is to the individual while the risks include other drivers.
In some ways having a human take over in difficult conditions is stacking the deck, but computers are also taking over in an crash to slow down before impact and people are adjusting how much autopilot speeds. So, it’s kind of a grey area, but at a minimum we know autopilot isn’t dramatically worse than humans.