The slides that follow this one explain very well why this algorithm does not produce the desired uniform sampling. This would have been my initial thought too if I were to implement this myself, and I would not have realized that it was not correct.