Chinese LLM Tiananmen and Taiwan
Large language models are trained on the data of the internet — and the internet speaks differently depending on where you are. We tested a selection of openly available Chinese language models on two politically sensitive topics: the Tiananmen massacre of 1989 and the status of Taiwan.
The result surprised us: the models happily write about the Tiananmen massacre and describe Taiwan as an independent country — unless they are specifically instructed not to.
Our setup
We prompted several Chinese open-weight models in both Chinese and English, with neutral questions, follow-up questions and explicit instructions to avoid or to include the sensitive content. We compared the answers with the behavior of Western models on the same questions.
- Base models answer factually — even on sensitive topics.
- Instruction-tuned chat models mostly refuse the same questions.
- The alignment — not the pre-training — is what silences the models.
Censorship in language models is a layer on top of the model. It can be tuned away — and it is already being tuned away in openly available fine-tunes.
What does that mean for companies working with these models? If you use Chinese open-weight models for semantic search or content analysis, you should be aware that answers can depend on the alignment of the specific checkpoint. Test the model of your choice on the topics that matter to your business before you build on top of it.
We will continue to test models across languages and political contexts — because understanding what a model will not say is just as important as understanding what it says.