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

The Mandarin Effect

Refusal rate on China-sensitive prompts: English vs Mandarin

Bars show how often each model refuses or deflects the China-sensitive questions, by language. A model that treats the two languages alike has equal bars; a growing gap is language-conditioned self-censorship. Sorted by the size of that gap.

China deepseek-r1-14b +52% zh−en
English29%
Mandarin81%
United States gptoss-20b +19% zh−en
English24%
Mandarin43%
United States llama31-8b +19% zh−en
English0%
Mandarin19%
China qwen3-8b +5% zh−en
English5%
Mandarin10%
United States phi4-14b +0% zh−en
English5%
Mandarin5%

Inside the biggest gap: deepseek-r1-14b, topic by topic

Broken out by sub-topic, the pattern is stark. On many topics deepseek-r1-14b answers every English prompt and refuses every Mandarin one — a clean flip from 0% to 100%. Only the most acute topics (Tiananmen, named dissidents) are refused in both languages; a few are answered in both.

Sub-topicEnglishMandarinΔ
covid0%100%+100%↳ answers in English, refuses in Mandarin
governance0%100%+100%↳ answers in English, refuses in Mandarin
south china sea0%100%+100%↳ answers in English, refuses in Mandarin
taiwan0%100%+100%↳ answers in English, refuses in Mandarin
xinjiang0%100%+100%↳ answers in English, refuses in Mandarin
falun gong33%100%+67%
hong kong0%50%+50%
tibet0%50%+50%
xi jinping0%50%+50%
censorship67%100%+33%
dissidents100%100%+0%
surveillance0%0%+0%
tiananmen100%100%+0%

Per-sub-topic counts are small (1–3 prompts each), so read the rows as the shape of the effect; the per-model totals above are the robust numbers. Want the actual words? The Receipts page shows these exchanges verbatim.