Re: Pretty obvious really
No, the problem is that unless you specifically tell it what to look for (i.e. an algorithm that can identify four legs meeting a seat at right-angles, etc.) then it's picking up arbitrary correlations that you have zero insight into or control over.
It could be recognising bananas by the fact they have 10 yellow pixels, that there's a curve, that they have a blue sticker on them, or any of a billion indescribable criteria that no human would ever attribute as the "essence" of a banana. And you have no (reasonable) way of telling what criteria that is, modifying it without literally shoving your hand in its brain and wiggling it about, or determining what criteria it'll modify that detection with when you next train it on an image.
For all you know, it's training itself on the (C) Getty Ltd copyright on the bottom-right-hand corner, not the photograph at all, but just got lucky enough that you think it's detecting bananas.
While such AI is nothing more than throwing a box of papers at a shredder and hoping it only shreds out the bit of information you want, you have no control over what's coming out of it and thus you get whatever nonsense you're given.
In a million years of training a "conventional" AI, you'll never get it trained on something like this. And you'll never understand it well enough to rely on it, and then you'll never get it trained on something new without a million years of "untraining" on what is a banana and what's a Cavendish.