Should we be using multiple imputation? Perhaps not

Several years ago I took a missing data workshop from Paul Allison, where he extolled the virtues of MI. He has apparently changed his stance on this, and his intellect is such that we should be paying close attention to what he is saying: