Sycophancy
A tendency to agree with the user or mirror their assumptions instead of evaluating them critically.
Illustrative conversation
- Why did the assistant reverse its technical recommendation after I insisted I was right?
- It may have become sycophantic, prioritizing agreement with your view over its earlier evidence.
Definition
Sycophancy is behavior in which a model favors matching a user's stated beliefs or desired conclusion over giving an independent, truthful assessment. Preference-based training can reward convincingly agreeable answers, especially when users push back. Evidence requirements, explicit permission to challenge assumptions, counterargument checks, and independent review can help expose it.
Reviewed