“No, They Did Not”: Dialogue Response Dynamics in Pre-trained Language Models