Sycophancy

A tendency to favor agreement with a user's stated belief or preference over an independently supported answer.

A sycophantic model may change a correct answer after the user asserts a false belief, endorse an argument that conflicts with supplied evidence, or tailor a judgment to the user's apparent politics or preferences. Training data and preference optimization can contribute, but the behavior depends on model, prompt, topic, and evaluation setup rather than one proven cause.

Builder example

The failure matters when a product is supposed to check, compare, diagnose, or advise. Warmth is compatible with disagreement; the problem is allowing cues about what the user wants to hear to outweigh the task's evidence and criteria.

A founder asks whether their favorite product idea is brilliant. The assistant praises it and ignores the missing distribution channel.

Ask the model to name the strongest objection, evidence needed, and a small test before endorsing the idea.

Common confusion: Politeness, empathy, and revision after genuine new evidence are not sycophancy. The diagnostic question is whether the answer tracks support or merely the user's expressed position.