It's just a surprisingly difficult hard problem to solve with a computer. People usually get PhD's for solving problems like this one (after dedicating three years or so of their lives to it).
Humans are so very good at taking in visual information and effortlessly processing it that we take it for granted. Image recognition is as about a hard a computer science problem as exists.
False positives on cloud detection is better than false negatives I think (which was the problem previously). If the telescope doesn't observe for half an hour when it could, that is better than observing for half an hour when it shouldn't. It wastes the same amount of time at the end of the day (because half an hours work extra needs to be done another night), but the latter wastes a load of our users' time while they wait again for their picture to be taken, and is frustrating.
Anyway, I'm working on it (as much as my limited expertise allows). It is slowly getting better.