A human might know the ultimate goal of sex is to reproduce, but still choose to use contraceptives. It's defeating the purpose of the original goal and just pursuing what feels good.

The same thing could possibly (likely?) happen with an AI system. While it might be able to reason about the intent of it's design (what the code's supposed to do), it'll still pursue what "feels good", which is an artifact of it's actual design (what the code actually does, bugs and all...).