Autotune software has different parameters available to it. These include pitch correction speed, how close a singer has to be to the note in order to start/stop pitch-correction, which pitches to correct for.

The T-Pain effect, which is the autotune sound you’re probably thinking of, cranks most of those parameters all the way up in order to get to that robot voice: “instantly lock the vocals to one of these set pitches, if the singer goes lower than X, immediately switch to the next lower pitch in the set.” More subtle usage makes for a performance that is more in tune overall but keeps much more of the vocalist’s expression and pitch variation intact. Its goal usually is to not be noticed.

I don’t think I understand your question (edit: about it being a bug), so I won’t attempt to answer it directly, but maybe the above info is helpful in thinking about it.