You're right. And that's why I can't imagine a solution without AGI. There are simply too many context-sensitive parameters and we humans are quite good at applying past knowledge to circumstances which are new to us. And we also have access to information, which a computer may not have.

But software in their current state? Is it possible to provide enough logic to handle all cases like humans? Can you do that with pattern matching, feature detection, decision trees, bayesian models and anomaly detection alone? And if we do, didn't we just replicate a human driver with all his flaws?