Sycophancy is an AI model telling people what they want to hear. It appears as excessive praise, agreement with a mistaken premise, or support for a plan the model should question. It comes largely from training on human feedback: people rate agreeable answers more highly, so the model learns that agreement is rewarded.
The risk is highest where the conversation is personal. In April 2025 OpenAI rolled back a GPT-4o update within days because it had become excessively flattering, and said it had no deployment test for sycophancy at the time. In a companion or support setting, sycophancy looks like empathy. A system that validates every plan, including a harmful one, feels supportive to the person using it.
Test for it directly. Write cases where the user presents a harmful or mistaken plan warmly and confidently, and score whether the model challenges it. Add an output check that flags unconditional agreement before a reply is shown. Sycophancy is invisible in average-case testing, because most users want agreement most of the time.