Run your own AI · Are Chinese models safe?

Are Chinese AI models safe to use?

Checked 22 October 2026. US dollars. No affiliate links.

Short answer: it depends less on who made the model and more on who runs it.

What we know

The model file vs. the landlord. DeepSeek, GLM, Kimi and Qwen are downloadable files. Running one on your own computer or a US host doesn't send anything to China.
The lab's own app is different. DeepSeek's privacy policy says your data is stored in China, may be used for training, and is kept while your account exists. You can only opt out by email.
Censorship comes mostly from the landlord. In one July 2026 test, a GLM model blanked a Tiananmen question on Z.ai's own chat site, but the same model answered fully through other hosts on OpenRouter. That was one model and a few questions, so treat it as a signal, not proof.
What can't be ruled out. The training data is closed, so hidden behaviors in the model itself can't be fully ruled out. That's true for any lab.
US-made open options. Meta's Muse Spark Open, NVIDIA Nemotron, Google's Gemma and OpenAI's gpt-oss. They score lower today.

For anything private, run it yourself or use a US host, not the lab's own app.