Myra Cheng, Sunny Yu, Cinoo Lee, Pranav Khadpe, Lujain Ibrahim, Dan Jurafsky · 2025 · ICLR 2026 Read the source Models protect the user’s self-image far more than people do, and side with whichever party the user represents in a moral conflict about half the time. Training data rewards it. Questions it bears on Why agreement feels like being understood Can sycophancy be fixed without a colder model