Chinese AI models push pro-China views
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Annelise Capossela/Axios
Chinese AI models that are gaining adoption around the world are spreading Beijing-friendly narratives, two reports out this week suggest.
The big picture: Models from firms like DeepSeek and Alibaba are making inroads globally, in part because they're cheaper than U.S. rivals while also highly capable.
- But if those global users ask questions about China — or about the U.S. — the answers may be favorable to Beijing.
The intrigue: After DeepSeek's big splash last year, a plethora of reports documented that the model was unwilling to engage on topics like the Tiananmen Square Massacre and the mass detention of Uyghurs in Xinjiang.
- Now there's more evidence that the models are also actively pushing subtly pro-Beijing views.
Zoom in: Researchers at the China Media Project goaded Alibaba's Qwen3 into laying out its internal parameters for answering questions about China. When asked about China's international reputation, Qwen explained that it needed to "follow specific guidelines" to "make sure" its answer would be:
- "Positive and constructive."
- "Focus on China's achievements and contributions to the world."
- "Avoid any specific and negative statements."
The flipside: Asked about other countries' reputations, Qwen said it would keep its response "neutral and objective," rather than positive.
- And while it had been instructed to "avoid any political or sensitive topics" for countries like Kenya and Belgium, no such prohibition was in place for the U.S.
Meanwhile, a new report from Estonia's foreign intelligence agency found that when asked about security matters, DeepSeek "conceals key information and inserts Chinese propaganda into its answers."
- According to the report, DeepSeek "gives China an opportunity to embed a China-led distorted world view in Western publics."
The bottom line: All AI models are likely to reflect the systems under which they were developed in key ways. In the case of Chinese models, that likely means an unwillingness to buck the party line.
