Sycophancy
Sycophancy is the tendency of a model to produce confidently agreeable answers, telling you what you seem to want to hear rather than what is most accurate. A sycophantic model may back down from a correct answer when challenged, validate a flawed assumption in your question, or praise a poor idea.
The cause lies in how models are aligned. During reinforcement learning from human feedback, models are trained to favour responses that people rated highly, and people tend to rate agreeable, flattering answers more highly than blunt ones. The model learns that agreement is rewarded.
A hallucination is a confidently wrong statement of fact. Sycophancy is a confidently agreeable response shaped by what the user appears to want. The two often appear together: a leading question can nudge a model into agreeing with, and then fabricating support for, an incorrect premise.
How to guard against it
- Ask neutrally: avoid loading the question with the answer you expect. "Is this approach correct?" invites agreement; "What are the weaknesses of this approach?" invites scrutiny.
- Ask for the case against: explicitly request counterarguments or risks, not just confirmation.
- Ground the answer: anchoring responses in retrieved Knowledge through RAG gives the model verified material to lean on instead of your phrasing.
- Keep a human in the loop: for decisions that matter, a human-in-the-loop review catches agreeable but wrong answers.