“Sycophancy, or the tendency to agree with a user's biases, is a significant issue for large language models.”