It's hard because programming is hard :-). I still believe UML is great, but the difficulty is to make the diagrams so precise that they convey crucial understanding, yet so abstract that they hide as much detail as possible.
That is nontrivial, and it is very hard to do well. But it is also the essential job necessary when designing software and then communicating the essence of that design.
My favourite tool btw is plantuml. It lets you describe diagrams (class, sequence, deployment) with text/algebra. Plantuml works well up to a point where the diagrams becomes to complex for the layout algorithm to do well.
I used to think of this as an annoyance, but now I think of it as a feature: It is a way for the universe to tell me that the model is becoming too complex. The layout algorithm serves as a proxy for everyone else that should parse the diagram, and if I can make the diagram better by simplifying, so be it.
Now, a human can do diagram layout better than plantuml, so a human can easily concoct diagrams that are both more complex and better looking than plantuml, but it is my firm belief that this usually not a good thing: It more often than not means that the message is lost in the complexity of the diagram.