On the Reliability of Large Language Models to Misinformed and Demographically-Informed Prompts