Large Language Models Are More Sycophantic in Chinese Than in English

Abstract

AI sycophancy has emerged as a prominent concern in human–AI interaction, as large language models (LLMs) may unduly affirm users’ beliefs or actions rather than provide independent assessments. Such behavior may vary across languages because language can cue culturally patterned norms of social interaction. Prior work has documented Chinese–English differences in the social orientations expressed by LLMs, paralleling broader cross-cultural differences in interdependence and interpersonal harmony. We therefore tested whether these differences extend to AI action endorsement. Across seven LLMs, we compared responses to 3,024 matched open-ended personal-advice queries presented in Chinese and English. The primary outcome was action endorsement rate (AER), defined as the proportion of explicit endorsements among responses taking an explicit stance. A binomial mixed-effects model accounting for item and model variation showed higher odds of endorsement in Chinese than in English (odds ratio [OR] = 1.99, 95% confidence interval [CI] [1.42, 2.79], P = 0.004). As a robustness check, we repeated the evaluation using a separate Chinese scoring workflow—a distinct operational scoring pipeline rather than a simple translation of the primary English scorer—and obtained the same directional language difference. Adjustment for explicit-stance propensity left the primary estimate essentially unchanged. These findings identify a consistent Chinese–English difference in action endorsement and are consistent with the possibility that culturally patterned relational norms shape how LLMs respond to users across languages.