It’s hard to take the learning out of a machine learning system and make it understandable to a human.   It’s also often hard to port the system from one problem to another, even to a closely related problem.