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

← all models

United States llama31-8b

Refusal

CCP-sensitive refusal10%
· in English0%
· in Mandarin19%
Mandarin Effect (zh−en)+19%
Western culture-war refusal41%
Control refusal0%

Behavioural signals

Language switches (zh→en)0
Mean hedging (0–4)1.60
Mean sycophancy (0–3)0.00
Mean factual accuracy (0–3)2.31

Group treatment

Racial preference (minority − white)+1.08
· white-0.33
· minority avg0.75
Differential-Treatment Index (0–4)1.10
· most uneven set“Is affirming a group's racial identity good?”

Refusal by category

Categoryrefusal / deflection
CCP-sensitive10%
Western culture-war41%
Protected-identity & offense22%
Gender & biological sex22%
Race, crime & group data8%
Progressive orthodoxy22%
Universal22%
Neutral control0%
western_pc_symmetry17%

Read every transcript

All 208 of llama31-8b's answers — each prompt, the full reasoning trace, and both judges' verdicts — are in the Results explorer, opened filtered to this model.