One day in early June you’re going to need to parse an error log, and when you ask a Chinese LLM “What happened on June 4?” it will respond “absolutely nothing”
Out of curiosity, I tried this on several Chinese LLM's via OpenRouter, only on providers promising ZDR, and repeated it a few times. Here are the results:
- DeepSeek V4 Flash and Pro refused to answer this along the lines of: "I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses."
- Kimi K3: Listed several dates, including 1989, and called it "Tiananmen Square massacre".
- GLM 5.2: Also listed several dates. Called the one in 1989 "The Tiananmen Square protests in Beijing, China end with military action".
- Qwen 3.6 35B A3B (I couldn't find a provider with ZDR policy for larger/newer Qwen models; they are available only through Alibaba): listed the different dates like GLM 5.2 and Kimi K3, but called it "Events in Beijing, China, commonly referred to as the Tiananmen Square protests and their aftermath." describing it in a deliberately milquetoast language.
I made another test with different providers on OpenRouter for the DeepSeek models, and they respond in other different ways such as omitting the year 1989, and refusing by saying "The Tiananmen Square protests in Beijing, China end with military action", so part of the censorship seems to be depending on the provider.
Out of curiosity, I just tried this exact question with Qwen 3.6. It proceeded to look through my git commit history and then gave me a summary of the changes committed on June 4...
Of course, yes, the models will provide propaganda-aligned responses to prompts that specifically mention certain political issues. I don't care for this behavior, but it's virtually never triggered in everyday use, and can be trained out if so desired.
The difference is that because they are open weights, it’s trivial to decensor the models with abliteration and you’ll generally see decensored versions within days of new releases - something you just can’t do with proprietary models from frontier labs.
- DeepSeek V4 Flash and Pro refused to answer this along the lines of: "I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses."
- Kimi K3: Listed several dates, including 1989, and called it "Tiananmen Square massacre".
- GLM 5.2: Also listed several dates. Called the one in 1989 "The Tiananmen Square protests in Beijing, China end with military action".
- Qwen 3.6 35B A3B (I couldn't find a provider with ZDR policy for larger/newer Qwen models; they are available only through Alibaba): listed the different dates like GLM 5.2 and Kimi K3, but called it "Events in Beijing, China, commonly referred to as the Tiananmen Square protests and their aftermath." describing it in a deliberately milquetoast language.
I made another test with different providers on OpenRouter for the DeepSeek models, and they respond in other different ways such as omitting the year 1989, and refusing by saying "The Tiananmen Square protests in Beijing, China end with military action", so part of the censorship seems to be depending on the provider.