Heat Death of Generative Models in Closed-Loop Learning