Chinese Censorship Is Leaking Into American AI Answers

A Nature study found that ChatGPT, Claude, and Gemini give more favorable answers about China when queried in Chinese, due to the dominance of state media in their training data.

Compiled by Sergey KostenchukPublished 2026-08-13Updated 2026-09-18

2026-08-13 Safety OpenAI

🛡 Chinese Censorship Is Leaking Into American AI Answers

A Nature study: ChatGPT, Claude, and Gemini give more favorable answers about China when queried in Chinese — state media Xinhua and Xuexi Qiangguo dominate the training corpora. Confirmed for 37 countries. Meta: 45% refusals on political queries.

🌍 A pre-training data issue. In CulturaX, state documents are 41 times more numerous than Chinese Wikipedia articles. OpenAI, Anthropic, and Google do not balance their corpora.

👤 One question in two languages — different AI answers. The language of the query affects the content.

Source: https://www.nature.com/articles/s41586-026-10506-7 WSJ: https://www.wsj.com/tech/ai/chinese-censorship-is-leaking-into-answers-from-american-ai-62abeca5

Sources