Large Language Models Still Exhibit Bias in Long Text