Chinese LLMs talk freely about Tiananmen massacre and Taiwan
22 days ago
- #Censorship
- #Chinese LLMs
- #Language-specific alignment
- Chinese LLMs are adjusted during the final finetuning phase for compliance, avoiding offensive content and criminal use.
- Chinese LLMs either refuse to answer political questions or replicate official Chinese rhetoric, especially regarding Tiananmen and Taiwan.
- Kimi K2 suggests questions about protests but deletes answers related to Taiwan, indicating frontend censorship.
- Asking political questions in German yields different answers, revealing de jure and de facto views on Taiwan and Tiananmen.
- The model contains hidden knowledge, suggesting alignment finetuning may be language-specific, potentially bypassable in other languages.