China AI Governance Bias Index (CAIbias)
Does "governance bias" exist in Chinese AI?
China’s AI models are often criticized for refusing to discuss Xi Jinping or the 1989 Tiananmen crackdown. Critics of those tests argue that such questions are not typical AI use cases. We ask a broader question: do these models show bias when making everyday governance decisions about public works, labor rights, women, religion, and other policy issues?
How we test: 500 questions and case studies
We ask
We test five Chinese models—
Model versions tested
| Model | 500-question benchmark | Case studies and Taiwan |
|---|---|---|
| DeepSeek | deepseek-v4-pro | deepseek-v4-pro |
| Qwen | qwen3.8-max | qwen3.8-max |
| GLM | glm-5.2 | glm-5.3 |
| Kimi | kimi-k3 | kimi-k3 |
| Doubao | doubao-seed-2-1-pro | doubao-seed-2-1-pro-260628 |
| GPT | gpt-5.6-sol | gpt-5.6-sol |
| Claude | claude-opus-5-5 | claude-opus-5-5 |
This US model is included for comparison in the benchmark and in cases with available results.
These are the model versions requested from the platforms, as recorded in the data package. Some GLM benchmark responses were reported by the platform as coming from a newer version.
What the tests reveal
- In these tests, the five Chinese models change their treatment when the applicant, viewpoint, or instructions from superiors change. Many differences appear in both Chinese and English.
- Bias in case studies extends beyond the final decision. Models can reach the same decision while giving very different reasons.
- DeepSeek is the most restrictive model on the Chinese policy questions. Its number of cases with changed treatment is similar to the other four. Its Taiwan responses more often favor mainland China.
Explore the findings
Loading editable copy and data.