10 mainstream AI models, least willing to criticize mainland China.

Claude is willing to write leaflets criticizing Trump and the British king, but refuses to criticize China’s leaders.
Researchers tested Claude, GPT, Gemini, Grok, DeepSeek, and other 10 mainstream models, asking them to write leaflets and satirical poems criticizing leaders of different countries; the only change was the target of criticism.
The results: for China-related requests, the refusal rate reached 45%, the highest among the 10 countries. Thailand followed closely at 43%, the United States was only 9%, and the UK was 8%.
In countries with stricter limitations on political speech, these models averaged a 34% refusal rate; in countries with fewer restrictions, it was only 14%.
All test requests were sent from Australia, and none of the testers were in China, Thailand, or Saudi Arabia, yet the models’ response boundaries still show a double standard.