Measuring Political Bias in Large Language Models: What Is Said and How It Is Said