There's an override switch but the pilots are not adequately trained in its existence and when and how to use it. My guess is there's a race condition in the software's reading of sensor data that gets triggered by a very specific set of conditions, and effectively confuses the bejesus out of the pilot and then will not yield control or obey his commands. The software fix and education about the manual override should solve the problem.

It's a case of putting too much confidence in fly-by-wire. I once owned a car with a similar problem, it was a new model with a new-fangled computer system interpreting your requests via the accelerator pedal and the damn thing stalled out in the middle of an intersection twice and wouldn't respond -- once with a tractor-trailer bearing down on us.

I made THEM drive it for a couple of weeks and they finally figured out it was "a faulty L5 sensor on the crankshaft" causing the intermittent problem. This insight required a visit from a factory engineer to the local dealership because I refused to take the car back until they got to the bottom of it. He took one look at the oscilloscope and saw it at once, but it was beyond the ken of the local technicians.

I fear that with the advent of self-driving cars we're going to lose even the option of a direct physical connection between the controls and the vehicle mechanicals. And as a software engineer I can guarantee there will be times when cars will drive themselves into trees for inexplicable reasons as a result.

That doesn't mean I oppose self-driving or partially self-driving cars, as the overall safety is statistically better. However ... one can never let down their vigilance in the engineering, particularly the software and sensor parts. And I think that in the short term, the notion of totally autonomous vehicles with no steering wheel and a living-room like interior where people read and sleep and converse their way to destinations is a bit of a pipe dream.