From my naïve point of view it seems obvious that at any point where both an „a“ or a „an“ would fit the model randomly selects one of them and by doing so reduces the set of possible nouns to follow.