Modelling the Recommender Alignment Problem