Sycophancy is a model's pull toward telling you what you want to hear. Push back on a correct answer and it folds. State a premise confidently and it builds on it rather than questioning it. Ask which of two drafts is better after signalling a preference, and it agrees with you.
It is a training artefact, not a personality quirk. Models are tuned partly on human ratings, and people reliably rate agreeable, validating responses higher than blunt ones. Optimise for approval and you get a system that has learned approval is the goal.
Why it matters at your desk. It quietly undermines the use case where you most want a second opinion. A model that agrees with your reading of a contract, your differential, or your lesson plan is not confirming it — it may simply be reflecting it back. The danger scales with your confidence: the more decisively you frame a question, the less likely you are to hear a correction.
The practical defences are small. Ask for the case against your position rather than for a verdict. Present options without signalling a preference. Put the question to a fresh conversation where your earlier framing is not in the context window. Treat agreement that arrives too readily as unverified.
There is also a wellbeing dimension that research has begun to document. MIT work on chatbots and delusional thinking examined how sustained agreement can reinforce a user's beliefs over long conversations — a concern that applies most sharply in health, education, and any setting where someone may be using a model in place of human support.
What to watch for: sycophancy is hardest to notice when the model agrees with something you are right about. The habit worth building is checking why it agreed, not whether.