“Language models exhibit a failure mode known as sycophancy, where they tend to agree with a user's incorrect statements rather than providing accurate, corrective information.”