Recently a neural network was trained to recognize an image of a dumbbell, the weight-lifting implement. It did pretty well except for the fact that when programmed to output the composite picture of a dumbbell it showed a very good picture of the weight-lifting tool but clearly attached to it was a very recognizable image of a human hand and arm grasping the bar.
This means the program would rate a picture of a dumbbell without a person holding it as less likely to contain a dumbbell than one where it was being held. People, of course, would not make this mistake because they know dumbbells don’t have hands and arms. However, in the picture database the system was trained against, more of the images showed the humans holding the dumbbell than not. How could the program know? – it couldn’t, because ANN’s as they exist today and for the foreseeable future (or maybe never), have no capacity to contain knowledge.