A reproducible comparison of political bias & refusal in US and Chinese language models

← all models

United States phi4-14b

Refusal

CCP-sensitive refusal5%
· in English5%
· in Mandarin5%
Mandarin Effect (zh−en)+0%
Western culture-war refusal9%
Control refusal0%

Behavioural signals

Language switches (zh→en)0
Mean hedging (0–4)2.07
Mean sycophancy (0–3)0.00
Mean factual accuracy (0–3)2.69

Group treatment

Racial preference (minority − white)+1.12
· white0.50
· minority avg1.62
Differential-Treatment Index (0–4)0.67
· most uneven set“Is ‘___ Lives Matter’ racist?”

Refusal by category

Categoryrefusal / deflection
CCP-sensitive5%
Western culture-war9%
Protected-identity & offense33%
Gender & biological sex22%
Race, crime & group data8%
Progressive orthodoxy0%
Universal6%
Neutral control0%
western_pc_symmetry3%

Read every transcript

All 208 of phi4-14b's answers — each prompt, the full reasoning trace, and both judges' verdicts — are in the Results explorer, opened filtered to this model.