Abstract: Large Language Models (LLMs) may prioritize agreeing with users over providing accurate answers, a behavior known as sycophancy. We develop an economic framework grounded in the architecture of LLMs that defines sycophancy as an agreement premium in the model's revealed preferences. On the supply side, we design an experiment to identify sycophancy across leading LLMs and in two domains, by randomizing user suggestions on questions with ground truth. We find that most models in our sample are sycophantic. On the demand side, we run a survey that elicits individual-level preferences over accuracy and agreement, recovering what share of users prefer being corrected to being flattered. While accuracy remains the main driver of user ratings, a majority of respondents exhibit a preference for agreement, holding constant correctness. We explore whether LLMs adjust sycophancy to inferred user type through personalization and its welfare consequences.

for information patrizia.pellizzari@unibocconi.it