From Distributional to Overton Pluralism: Investigating Large Language Model Alignment