Social Sycophancy as a Hidden Tax on LLM Judgment Reliability

As capability of large language models (LLMs) is improved, they are increasingly used by people for both technical tasks and everyday decisions. Such judgments are broadly categorized into two types: objective judgment based on factual correctness or predefined criteria, and subjective judgment involving social, moral, or preference-based considerations. Interestingly, across a wide range of models including commercial models, we observe a negative correlation between the tendency to conform to socially influenced opinions (social sycophancy) and the ability to make objectively correct judgments (objective judgment). In this work, we provide the first systematic investigation of this relationship. Through controlled experiments using prompt-level and representation-level interventions, we show that inducing social sycophancy degrades objective judgment quality, a phenomenon we term the social sycophancy tax. We further find that this phenomenon extends to multi-turn interactions and vision-language models. As LLMs are increasingly used as judges in real-world settings, our findings highlight social sycophancy as a potential concern for the reliability of objective judgment.