Preprint Preliminary

Large Language Models Are More Sycophantic in Chinese Than in English

Cross-lingual test of whether large language models endorse users' intended actions more readily in Chinese than in English. Seven models answered 3,024 matched open-ended personal-advice queries in each language; the primary outcome was the action endorsement rate (explicit endorsements among responses taking an explicit stance), modelled with a binomial mixed-effects model accounting for item and model variation.

Read at osf.io

Publisher

PsyArXiv

Published

7 Oct 2026

Added

today

DOI

10.31234/osf.io/7ah85_v1(opens in a new tab)

Key Findings

Odds of endorsement were higher in Chinese than in English (odds ratio 1.99; 95% CI 1.42 to 2.79; P = 0.004) across seven models and 3,024 matched personal-advice queries.
A separate Chinese scoring pipeline, rather than a translation of the English scorer, reproduced the same directional difference.
Adjusting for the propensity to take an explicit stance left the primary estimate essentially unchanged.
The authors read the result as consistent with culturally patterned relational norms (interdependence, interpersonal harmony) shaping how models respond across languages.

Methodology Notes

Preprint; not peer reviewed. The abstract does not name the seven models or the source of the 3,024 queries, and author affiliations are not shown on the OSF page or the Crossref record. Submitted to PsyArXiv 2026-09-29; Crossref DOI created 2026-10-07 (version 1), which is used as the publication date. Verification route: Crossref record (title, five authors, posted date, abstract); the OSF preprint page was rendered by the clinical beat through a reader proxy.