Which is exactly why I'm still 100% convinced this whole pipe dream will disappear the moment self-driving cars become available to consumers, and a few serious accidents happen. No car manufacturer on earth is going to continue selling cars with the promise they can 'drive themselves' if that makes them liable for accidents. Exhaustively testing a cruise control system is one thing, but ensuring at least 'nine nines' of reliability of all the hardware and software systems of a self-driving car, in every possible situation it could be used, is simply intractable. Anything less than nine nines reliability, and someone will die because of a problem with your self-driving car at least once every week or so.

I kind of agree with that. That is why I inserted the bit about pervasive logging of events. This technology will increase that many times over. Any manufacturer getting into this will need that in order to prove beyond any doubt that their software did the right thing.

NOW bugs happens and in this case bugs will kill people at some point. So, can the manufacture be trusted with keeping logs themselves. The answer is no. Not if they have a strong incentive to make it look like it is not their fault. It it basically either their ass on the line or driver's ass. There will always be a question of "ok, so they do show that user pressed the pedal and tried to take manual control of the car, but how do we know they didn't insert that artificially in there...".

I think there will be a need for 3rd party or government mandated continuous logging. Cameras, following distance sensors, accelerators, local road conditions monitoring and so on, so during a law-suit this 3rd party's product (who supposedly should have not interest in falsifying the data) would be used in establishing fault.