Last November, Robin described a study where subjects were less overconfident if asked to predict their performance on tasks they will actually be expected to complete. He ended by noting that "It is almost as if we at some level realize that our overconfidence is unrealistic."

I think there's a less perplexing answer: that at some level we realize that our performance is not 100% reliable, and we should shift down our estimate by an intuitive standard deviation of sorts. That way, we can under-perform in this specific case, and won't have to deal with the group dynamics of someone else's horrible disappointment because they were counting on you doing your part as well as you said you could.