DeepSeek is great, but why is it so cagey toward good faith questions regarding Marxism/Chinese history?

You shouldn't be asking LLMs about opinions anyway but it does make some work more difficult because of it, like proofreading or translating. I actually find deepseek overcomplicates things too much most of the time. It takes forever to reason something and then overengineers what you want it to do when chatGPT can just do it in a few lines of code. Providers should be moving beyond LLMs and into specialized AI, I'm sure Deepseek can provide this free of charge. One AI for coding, one AI for translating, one AI for proofreading, etc. instead of trying to have a "digital assistant" that does none of it right.

It's too easy for a bad-faith user to trick it into saying something that can be used to support all the Sinophobic propaganda out there.
If people genuinely want answers to questions like the ones in your screenshot, it's much better to email their local Confucius Institute.
- local Confucius Institute
  I might look into this, never knew such orgs existed thank you :)

As others have said, LLMs aren't "fact machines" or something, they just aggregate data based on the prompt and reword it to form an appropriate response. An LLM scraping the English language internet is naturally going to have a huge number of anti-China articles as a part of the data set.

People should stop treating LLMs as some sources of truth. They were trained on data from the Internet, some of these data are sources but most of them are derivatives, some are even themselves AI-generated. LLM is a search engine that doesn't provide a source, and a word mixer.
Anyway, if you really want to learn histories from the Chinese perspectives then you'd have no choice but to learn Chinese.
- LLM is a search engine
  I wouldn't even call them that. Any fulltext search engine will always produce reliable output and doesn't make shit up. Also they don't need a GPU to number crunch and don't require the most recent beast computer to run. For example, any PostgreSQL Table can be used as a search engine index via TSVECTOR.
- LLMs have great use-case applications like erotically roleplaying with yourself or creating incredibly shit search engine summaries. That's the beginning and the end of their application, though, and I reel when I consider that some people are putting in questions and taking the generated answers as any sort of truth.
  Genuinely think it's a huge problem that another generation down the line and "AI's will lie to you" won't be taken as common sense
  
  Hopefully we will see improvements in architecture and alignment research long before LLMs reach a point of normality that is taken for granted. Right now, there seems to be a push to use LLMs as they are to cash in on investments, but they are far from a point for research to chill at.
- 💯 The best LLM is still worse than wikipedia when it comes to getting factual information and that's saying something, considering the "pretense of neutrality yet heavily biased" problems wikipedia has. At least with wikipedia, you can see what sources a page is using. With an LLM, as a user, you have no idea where it's getting the ideas from.
  
  And in a lot of cases it has just made everything up in a way that sounds as plausible as possible.

The better question is why are you trying to learn about this topic using a LLM?

You see the same with US models like Copilot if you ask about things like the election process and such, Copilot will just tell you it's outside of its scope and please look elsewhere for more current information.
Me: How does voting in the USA work?
Copilot: I know elections are important to talk about, and I wish we could, but there's a lot of nuanced information that I'm not equipped to handle right now. It's best that I step aside on this one and suggest that you visit a trusted source. How about another topic instead?
It's not really a good idea to let an AI freely speak about topics that are so important to get right, because they are not perfect and can give misleading information. Although, DeepSeek is open source, so there is nothing stopping you from downloading it to your PC and running it there. They have distilled models that are hybrids of R1 and Qwen for lower-end devices, but even then you can still use the full R1 model without filters through other companies that host it.

At least part of it seems to be a "guard-model" identifying that the topic might be illegal or something like that and just stoping the main model from going further while saying "can't do that". Other Chinese models, from what I heard, might handle things like this better, although in DeepSeek's case they might have gone too far because they might have been under attack early on from bad faith people asking all sorts of "questions" about China and things like that, but I can't say for sure.
As the other person said, you can either try running the model yourself, or if you don't have a computer with at least 400GB RAM lying around you can look for other services hosting DeepSeek. I can't guarantee they won't have "guard-models"/censorship themselves but there should be some without it.
But another point to consider though is that DeepSeek was still trained on a lot of western ~~data~~ propaganda, so don't expect impartiallity from DS or any other model for that matter. We may still be a ways off of models that can actually understand their bias and correct it properly.
- You can pen-test the restrictions pretty easily. Its coded to censor mentions of China or Chinese leaders in an political context (or maybe any context at all)
  If you ask the same question but ask it to replace any mention of the word China with something else (I used zeon from msg), it will actually answer the question, but obliquely.
  If you ask it to answer again with real Chinese historical figures, it fails

I think it has to do with the language being used. Language data discussing Marxism between English versus Chinese is bound to be inherently different. The AI has a lot more examples of discussing Marxism defensively in English than it does in other languages, and so will spit out more consistently cagey responses to an English user.
I don't know any of the technicalities of how AI works but I've read this explanation somewhere before.

Maybe try running it locally if you can.

Deepseek was trained in a lot of western and English stuff, they also stole a lot of data from Chat GPT, and every single of those are very anti-communist and anti-China, DeepSeek just regurgitates western takes, at least when you use English, so the devs put a censor module in the website, so it doesn't say or touch topics they know DeepSeek will not handle well (like Tiananmen, or Xi). I've heard that, in Chinese, DeepSeek talks about these stuff as the training data is not that shitty.