Comradeship // Freechat

I haven't used Deepseek in a while so I don't know if something changed but when I wanted something about the PRC I tried being more specific and it would answer most of the time. Stating you're a Marxist and asking the model to answer you like a Marxist-Leninist would also helped.

Tho I asked in Portuguese, so I don't know if this will actually apply in English.

Interestingly enough, I have a chat going with DS where at first I instructed it to use dialectical and historical materialism, give a communist perspective and be critical of bourgeois sources, etc. I've asked many political questions to it. Lately, I was asking it some questions about transitions from feudalism to capitalism and how Engels' works relate. Then I asked how China and India figure into it. Its answer on China started getting suspiciously western sourced (I had web search on) and the answer was not good at all. I told it that it had made a mistake about China being a dictatorship of the bourgeois, and it needed to study Xi Jinping Thought. Then it gave me the same answer as OP. Then i just simply said "you made a mistake in your last reply" and then it rectified its answer and even did like a mini struggle sesh as to why it was wrong, saying it incorrectly had been using european frameworks and analysis, among other things. Pretty neat stuff.

It might just be better to ask about the history of China in the modern period, it likely has words which trigger this response.

They just have some lame filter that stops it from even mentioning China in a political context.

i mean, china knows all the trash-talk that spreads around the internet about china, and probably the best course of action was "do not talk about politics", just look how grok ended

Not-so-educated guess : the english language data it was trained on has an anti-PRC biais that was identified by its creators so they blocked anything about them instead of letting it repeat anti-PRC propaganda to westerners

Deepseek doesn't allow discussion of China/Chinese politics all that much. That isn't really unique to deepseek. Social media in China tends to have rules about it too.

You have to keep in mind these services are run by companies within China. Companies that China keeps on a short leash for good reason. So the companies do this to cover their ass i think. Because if for example someone were to get Deepseek to say bad stuff about China then the CPC might come to Deepseek's makers, and be like "Hey wtf is up with this? Are you trying to be counter-revolutionary?"

It's hard to moderate complicated stuff like that with a fine toothed comb. Similar to how in the west a social media site will outright ban NSFW content to avoid having to moderate it to weed out CSAM from the normal stuff. It's a wide net block to save them the headache of moderating it in more detail. And to prevent accidental slip ups.

This is a decision the company itself has made in response to much more reasonable regulations by the CPC to make sure companies are not undermining the government. That's my take on it anyway. It's annoying, but it's not a huge deal imo. Not like LLMs are reliable enough to get good info from anyway. Even if they didnt block it.

Edit: Also i think the wya they implement it is they have blocked certain key words. Where if the AI thinks or write those words it blocks the prompt. I think for example it can't talk about Xi Jinping. So if while answering it brings him up it'll stop and give that message. I'm guessing they had issues with westerners trying to get it to say bad stuff about him so they blocked his name.

Ask it in Chinese.

When a model is trained, it doesn't know what is right or wrong. A bunch of english data on Chinese politics will be blatant propaganda

It looks like it starts to actually answer, both in English and when asked in Chinese, so it looks like the filtering is in the web UI and not the model itself. If you're curious you could find another provider or run it locally to see the answer.

They’re damned if they do and damned if they don’t. Previously.