alt Hacker News> I have serious concerns about how these models might reflect Chinese government perspectives (try asking them about Tiananmen Square).
And I have serious concerns about the American ones. Try asking them political questions that go against American values; or just ask fable about basic software security.
Replies
What are some example questions you would pose that are against American political values?
Getting opinionated replies about politics does feel less dangerous than the model shutting up completely when asked about past government atrocities. At least in the west we can freely discuss and criticize.
This seems like an easily testable assumption. Without the easily obtained data this comment is useless speculation at best.
For instance, try asking American models about the Palestinian genocide
[flagged]
[flagged]
I had a hn scrapper and feed daily hn front page posts to llms which summarizes stories and comment section consensus. The model (deepseek v4 flash) completely refused to work for this article because there are scattered (like 5?) mentions to the Tiananmen Square incident in the article and comments.
It's more serious than "arguing with a machine with political bias". If your service depends on a Chinese model, all it takes to get compromised is someone sneaking in one of those CCP red lines into their user content. Unless you use an uncensored version.
>Try asking them political questions that go against American values
1. Can you give me some examples?
2. Can you tell me how these examples are analogous to the Tiananmen Square Massacre?
Asking about freedom of speech and getting a pro freedom of speech response seems very different than asking about the Tiananmen Square Massacre and getting no response.