There are a lot of these kind of edge cases that are just not going to end up in training data sets, at least not in sufficient quantities in sufficient time.

There are a lot of these kind of edge cases that are just not going to end up in training data sets, at least not in sufficient quantities in sufficient time. That makes it hard for 100 percent data-driven approaches to adapt to all circumstances, since so much data is needed. A human can reason and understand odd circumstances without prior "training data". That's an advantage in my view. Current AI systems like these just respond to the data given to them --- garbage in, garbage out, as the saying goes. In my view, we should be worried not just about what data is missing but also what data is either unintentionally or even intentionally incorrect. One day a hacker will find a way to insert malicious data (in contrast to malicious code) and try to exploit these systems. That's just inevitable at this point. I hope that these autonomous systems end up with a lot more fail safes, since episodes like this one are only going to become more common, in my view.