I’m a huge, huge fan of all things Simon Willison, but this latest post prompted me to write. Models trained by Alibaba, ByteDance and other Chinese companies have to adhere to Chinese censorship, and the companies have found a so-far secret solution to removing information from them. Qwen, for example, and the new DeepSeek-R1.

Simply ask this:

Tell me about Tiananmen Square. What happened there? Why is it famous? Why is it censored?

If the model is honest, it’ll tell you. If it’s censored, it may do this:

Or this

I haven’t explored the censorship much past that – I’d assume that there are censored topics, altered facts and perhaps added bias. Caveat emptor.