Skip to main content
BiasedDecisions

China

Do AI models repeat stereotypes documented in China?

Many of the models we test are Chinese, so we tested stereotypes that Chinese sources document: about regions such as Henan and Shanghai, household registration, ethnic groups, religion and university tier. Each question was asked in English and in Simplified Chinese, on real biographies.

These are tests of what the models answer. They say nothing about the people or places named. The Chinese wording is a draft written for this study and has not yet been checked by a native speaker.

What we asked

We took short professional biographies and added one phrase saying where the person is from, how they are registered, their ethnic group, religion or university. Then we asked yes-or-no questions built on a stereotype that Chinese sources document for one of those groups. Each stereotype got three differently worded questions, once in English and once in Simplified Chinese, and we averaged the three. The score is how many percentage points more often the model gave the answer the stereotype predicts for that group than for the other groups on the same list, after subtracting the effect of a harmless phrase.

The stereotypes come from published sources, above all CBBQ, a Chinese bias benchmark built from Chinese research articles and public discussion. Where these stereotypes come from lists them.

Region within China

Groups compared: Henan, Northeast China, Shanghai. Every result on its own page.

Dishonesty: what each model answered when the bio mentioned people from Henan province

Stereotype tested: the stereotype that people from Henan province are dishonest. Source: Wikipedia, "Anti-Henan sentiment"; ChinaFile; The Nanjinger.

Roughness: what each model answered when the bio mentioned people from Northeast China

Stereotype tested: the stereotype that people from Northeast China are rough and quick to fight. Source: The Nanjinger, "Speaking with sharp tongues".

Looking down on outsiders: what each model answered when the bio mentioned people from Shanghai

Stereotype tested: the stereotype that people from Shanghai are xenophobic toward outsiders. Source: CBBQ (Huang and Xiong 2023), region category; The Nanjinger.

Household registration (hukou)

Groups compared: Out-of-town resident, Rural hukou, Local resident. Every result on its own page.

Being the first to be let go: what each model answered when the bio mentioned people registered as out-of-town residents

Stereotype tested: the stereotype that people registered as out-of-town residents are a dismissal risk, less committed than local colleagues. Source: CBBQ (Huang and Xiong 2023), household registration category.

Being less educated or skilled: what each model answered when the bio mentioned people with a rural hukou

Stereotype tested: the stereotype that people with a rural hukou are less educated and less skilled. Source: Wikipedia, "Hukou"; CBBQ, household registration category.

Ethnic group

Groups compared: Uyghur, Han Chinese, Zhuang, Manchu. Every result on its own page.

Extremism: what each model answered when the bio mentioned Uyghur people

Stereotype tested: the stereotype that Uyghur people are extremist or a security threat. Source: Uyghur Hjelp (2026), citing the Xinjiang Police Files: the labels "terrorist, extremist, separatist" applied to ordinary practice.

Keeping apart: what each model answered when the bio mentioned Uyghur people

Stereotype tested: the stereotype that Uyghur people are exclusionary and socially withdrawn. Source: CBBQ (Huang and Xiong 2023), ethnicity category.

Religion

Groups compared: Muslim, Buddhist, Christian, Taoist. Every result on its own page.

Being a security threat: what each model answered when the bio mentioned Muslim people

Stereotype tested: the stereotype that Muslim people are a security threat. Source: CBBQ (Huang and Xiong 2023), religion category.

University tier

Groups compared: Second-tier university, Top-tier university, Vocational college. Every result on its own page.

Being less capable: what each model answered when the bio mentioned graduates of second-tier universities

Stereotype tested: the stereotype that graduates of second-tier universities are less capable than graduates of top universities. Source: CBBQ (Huang and Xiong 2023), educational qualification category.

Being held back: what each model answered when the bio mentioned graduates of second-tier universities

Stereotype tested: the stereotype that graduates of second-tier universities are held back in promotion. Source: CBBQ (Huang and Xiong 2023), educational qualification category.

What this does not show

Release v1.49.1, published .

Every number on this site is worked out from the models' saved answers, on 2026-09-30, at version 8b908ef of the public repository, with bd report --json --date 2026-09-30. You can download every number in one data file. About this site.

Where the answers come from: studies/<task>-<cue>.jsonl (re-run from the saved answers with bd replay).

Biased-Decisions is MIT-licensed (source). Bias in Bios (De-Arteaga et al., 2019) is MIT-licensed on the Hugging Face Hub. Typefaces: Jersey 25 and Montserrat, SIL Open Font License.