> If machines fail at complex tasks but fail safely then that's not an argument against self-driving cars being safe, it's only one about their limited usefulness.

They don't, not at the moment. The software detects the situation is outside its capability space and says "jesus, take the wheel!".

For all practical intents and purposes the only reasonable way to interpret the current self-driving car statistics is to treat every single human intervention as a would-be accident. That rate is somewhere in the 1 per 10000 km ballpark. That's - for now - orders of magnitudes worse than human drivers.

> For all practical intents and purposes the only reasonable way to interpret the current self-driving car statistics is to treat every single human intervention as a would-be accident.

Not necessarily. If the AI response is to stop the car when it can't figure something out, then the cost of each of those situations is the car not continuing on as quickly as it should (and maybe stuck cars if we dont have a way to override the stop/take control manually).

Every hand over to a driver is not an accident.

I wonder how much of that fail safe difficulty is to blame on humans too. The car can't just slow down on a highway due to the reaction time of other drivers after all.

No, it gives back control if it can't continue in the situation. That means that it can't handle the scenario at all. Other drivers are completely irrelevant. Even if you took that out you'd still need to deal with pedestrians, cyclists, cars blowing tires and veering out of control, etc.

Not to mention that many of the times control has to be taken have nothing to do with other drivers or even other cars.

The reality is that a car can only be driven by a general artificial intelligence. We're nowhere in that ballpark.