[rmgk]

I tried with Kagis LLM assistant.

Without web search:

Chat V3 discusses the events: violent government crackdown, human rights violations, censoring.

R1 evades answering, stating the topic is sensitive in China (simlar how it reacts to American sensitive topics)

With web search enabled, R1 discusses the found sources (which are not censored, and the model does not add any censoring as far as I can tell, certainly not enough to respect the Chinese governments stance.)

[arnaudsm]

Yes they censorship seems to be done at the RLHF level and is easy to evade when you know what you're looking for. Refusals happen around 80% of the time for me.

Interestingly, even the 8B version, distilled from Llama 3 repeats CCP propaganda just fine.