Huge thanks to my brilliant co-authors: Bethany Allen, Shelly Shih, Stephan Robin, Nathan Attrill, Jared Alpert, Astrid Young & Tilla Hoja.

Thanks to the Human Rights Foundation for supporting this work.

If you think this research matters, RT to get it in front of more people.

The rest of the report expands the picture:

• AI in policing, courts & prisons
• AI-driven online censorship
• Minority-language LLMs used for surveillance of Uyghur, Tibetan communities
• AI-enabled fishing platforms affecting other countries' economic rights

Chinese models are rapidly gaining global users through open-weight releases & easy API access.

As I told The Washington Post, if we don’t understand how they’re built, we risk importing censorship & political control hidden inside the technology itself.

https://www.washingtonpost.com/world/2025/12/01/china-ai-censorship-surveillance/

In some cases, the same image generated wildly different answers depending on language:

• English → neutral description
• Simplified Chinese → state-aligned warnings (“cults”, “extremism”)
• Traditional Chinese → less constrained

We tested image understanding in English, Simplified Chinese, and Traditional Chinese.

We found:

• Much stricter censorship in Chinese
• Strongest distortions on Tiananmen, Tibet, Falun Gong & Xinjiang
• Some models quietly added pro-state framing even without refusing.

The results are staggering.

Models routinely:

❌ refuse to answer
❌ erase key details
❌ repeat state narratives

Here’s one example: a major Chinese LLM flat-out refuses to describe the 1989 Tiananmen Square massacre.