Assessing the Reliability of LLMs Annotations in the Context of Demographic Bias and Model Explanation