Re: the error is in call it "AI" !!!
So the NNs do the same thing we do
Except we have not mapped only one or two data sets, but many hundreds of data sets. Often we need to make use of correlations learned from one data set and apply it to a situation that is normally the domain of a different data set.
For example, how we react while driving and an unexpected object appears in the road depends whether the object we see maps into the data set of "very hard solid things" (e.g. big rocks), "Soft inanimate things" (e.g. a bin bag blowing in the wind), "Small animals" (e.g. fox), "possible human" (e.g. runaway pushchair) or "Something abnormal" (e.g. very big pothole, open fissure, collapsed drain etc.). We also react based on our knowledge of human behaviour - e.g. a ball bouncing across the road is itself no danger, but may well be followed by a child chasing after it.
We can also recognise and appreciate the difference between a road that has a soft verge, a road that has a ditch next to it, and a road that has a sheer 200 foot drop next to it, and this will influence our decision on the best course of action to take in order to try to avoid a collision.
The "A.I." in a driverless car would be able to recognise a tiny fraction of the things we are able to recognise, because its "knowledge" might be quite deep, but is nowhere near as broad.