Our work shows that sycophancy is a persistent trait of AI assistants, likely due in part to flaws in human feedback data. This suggests we will need training methods that go beyond unaided, non-expert human judgment, such as LLM-assisted human feedback: