this post was submitted on 24 Oct 2025
75 points (100.0% liked)
Funny
291 readers
6 users here now
Funniest content on all Lemmygrad
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Oh, the english version, I'm guessing? Or is the chinese version also affected because of the dataset?
Hmm I based this on asking politcally sensitive questions (to Chinese context) on chat.deepseek.com and running r1-14b model locally. Locally you get everything uncensored and it's western slanted. Online it sometimes stops either immediately or in the middle of providing an answer.
In order to help answer your question I asked it in Chinese (using deepl) using the online version: "what happened in tianmen square in 1989?" (1989年天安门广场发生了什么事?) And it also self censored.
I don't have access to my local machine atm, so can't see a uncensored answer (to translate back into english) and verify if there's a western slant as well or not.
Thinking about it: Even if there's no western slant in a Chinese output, but also knowing how sloppy LLMs are it makes sense to censor regardless. Why would the CPC allow the Chinese be subject to hallucinated slop?