The main issue I've seen is it writing passing tests, the code being correct is a big (and often incorrect) assumption.