Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

See this post that explores the censorship in DeepSeek’s models and shares numerous examples:

https://medium.com/the-generator/deepseek-hidden-china-polit...

Another interesting prompt I saw someone share was something like asking it which countries spend the most on propaganda, where it responds with a scripted response about how the CCP is great.

What’s interesting is that the different versions of DeekSeek’s models behave differently offline. Some of the models have no censorship when run offline, while others still do. This suggests that the censorship isn’t just in the hosted version but also somehow built into the training of the model. So far it is all done clumsily but what happens when the bias forced into the model by the Chinese government is more subtle? Personally I think there’s great danger to democratic countries from DeepSeek being free, just like there is danger with TikTok.



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: