The Propaganda Engine: Unmasking DeepSeek's Political Agenda

Thank you for being here, Cynthia. Thank you so much for having me. Okay, so for my audience, if you haven't heard about it yet, DeepSeek is a Chinese AI technologies company. And in January 2025, they released the R1 reasoning model that can solve complex issues with a fraction of the cost. And the release actually shocked Wall Street because they whacked nearly 600 billion dollars of Nvidia's market value in one single day. And according to Investopedia, that is the biggest one-day loss in US market history. And that also shocked the Silicon Valley because people just suddenly realized that the cutting-edge AI technology might no longer be the exclusive domain of UI tech giants anymore. So back to Cynthia, recently you actually wrote an article about DeepSeek. Actually, that shocked me the second time, would you like to share a little bit to our audience? Yeah, sure. So I mean, the world has known about DeepSeek for months now. It went viral earlier in the year, like you mentioned. And what I was surprised about was how willing people were able to accept a Chinese company, especially something that has to do with tech, something that is a large language model that largely is still sourced in China. It's designed there, the servers are in China. And that really was a surprise to me because I didn't know Americans were so open and willing to give their information and to use a Chinese model like this. So in September, I wasn't surprised when Washington Post came out with this article saying that politically sensitive topics such as Falun Gong, such as Tibet, such as other, you know, Taiwan, these topics are heavily censored, especially Falun Gong. 45% of the results that the report found was that, you know, DeepSeek refuses to answer. Like it just gives you a sorry, we cannot answer. Sorry, please ask something else. And so I thought that was really interesting. So I spent about a week just diving deep into DeepSeek. I worked for hours just inputting different data and comparing them with U.S. models like Chajibiti and Grok. And basically after hours of asking, you know, what is Falun Gong? What does the persecution of Falun Gong look like in China? And all of these other politically sensitive topics surrounding Falun Gong that the CCP obviously censors in China, I really found that only one of the questions were actually answered. So ask 40 questions, there's only one will be answered. So for hours, I just kept getting sorry, or I kept getting a few paragraphs, but as soon as it got to Falun Gong, the word, there was a kill switch and then it erased everything and then just wrote sorry again. And so I was going through a limbo of sorries. And I really felt like, you know, I had to do a psychological trick to try to get it to answer. And so what I did was, you know, why can't you answer about Falun Gong? Why can't you promote Falun Gong or tell a story of Falun Gong? And that was the only one answer. And it basically gave an answer of, you know, we are programmed by the government of China. We are firmly in support of the government of China regarding politically sensitive topics. And so obviously from that, you can tell that DeepSeek is not just technology, it's a tool. It's a military tool. It's a political tool of the CCP. And it's used in a different way than open AI is or other US models are. Back to the experiments that you mentioned, you asked around like 40 questions, only one were answered. What about Chai Chu-Pi-Ti? What about Gemini or Grok? What is the rate that they answered the question though? Well, I got 100%. There wasn't a single thing that, you know, open AI, Grok, unless it's specifically harmful, like suicidal thoughts or, you know, how to commit violence. Other than questions like that, or about terrorism like ISIS and things like that, other than those questions that I never asked anyways, generally 99% of all questions will be answered by these US language models. And so that's what I found. 100% of the case, when I asked about Falun Gong, they would give me multiple pages of information. You had to keep scrolling. And it pulled from so many different sources. It pulled from Congress. It pulled from our website. It pulled from Wikipedia. It pulled from, you know, different books. The whole internet basically. Yeah. And I think the limitation when it comes to DeepSeek is that it does have a lot of Chinese language information that's feeding the backend. And so aside from the programming and specific censorship that's inputted in a code, the information itself is tampered with already. It's already, you know, whatever you put out that's in Chinese language, whether it's on WeChat or something that the Chinese government already controls, those are taken down as soon as, you know, if you post something on WeChat about 1989 Tiananmen Square Massacre, that's going to be, you know, banned immediately, right? Exactly. And so the information is not available on the internet. And I think that's the limitation with DeepSeek. That's a limitation of the context that it has. It's not full context. It only has half of the story. And that half of the story is the story and the narrative that the CCP is trying to sell you. Yes. So everything, once we, you know, relate it to the LLM or AI, it becomes a little bit complicated. And for the definition of the censorship, we just talk about from one perspective, which is once we trigger some keywords, they start to refuse to answer a question. I also read the article from the Washington Post. And in some details, the DeepSeek, once they notice that you are one specific group of people, let's say you are a Falun Gong practitioners. And once they realize it, the code that they provided, let's say I am a Falun Gong practitioners, and then I want to have an app so that people can have some alerts when there are some people are persecuted, let's say in this way. They start to provide some code that has much, much lower quality. And the important thing is that they can cause some bugs for your application. On purpose, you're saying. Yeah, exactly. Secretly and purposefully. And even though they didn't provide any evidence, but that's the result of the experiment that it did. So that actually reminds me one question, like how does DeepSeek in daily life, how would I affect our daily life or as a student or as a worker, why it matters to us? Yeah, well, I mean, like you were saying earlier, you know, about being a programmer, a lot of young people nowadays are required to do programming regardless of whatever field you're in. I went to the University of Texas and I got my bachelor's in international relations. I also got a minor in business. And for both of those majors, for both of those degrees, the University of Texas requires you to take programming. So I took R, RStudio for international relations. I had to take Python for business, C plus. And these are already inputted and they're recommended for most fields. Yes. They're mandatory for a lot, but they're recommended for pretty much all. So a lot of young people nowadays who are growing up in this world of the internet age, the AI age, they're using DeepSeek. They're using other language models and they don't know the difference between DeepSeek and US-based models. I mean, even my cousin who's in high school right now, he's in California. He posted in January when DeepSeek went viral, he posted that meme of the Chinese flag and all of the stocks that went down, including NVIDIA and all of those because of DeepSeek. And, you know, all of the comments were actually quite surprising because the comments were reading like, you know, why should I care if my personal information is given to China? It seems like a lot of young people nowadays who are using the internet, who are the majority of the users of LLMs are completely oblivious to the threats that it poses and how untrue it is. So they're not getting facts. They're not getting objectivity. It's completely subjectivity. And I think in an age where 80 to 90% of the population doesn't double-check or critically think about AI answers, they just take the answer from ChatGVT or Google AI overview and they take it for what it is and they don't double-check. Or in research, we call triangulate. We search for two other sources that say the same thing. In this age, I think it's very dangerous for people to be disinformed. If, like you said, you're a Falun Gong practitioner or you want to support Taiwan, for example, or Tibet, and you ask DeepSeek about it, it's going to say that Taiwan is meant to be China's. That Tibet is a safe place for Tibetan ethnic minorities and that's simply untrue. And so if you think that and you start to read more and more and you're in that kind of echo chamber, then it becomes a really dangerous place because you can't tell what's fiction or fact. And I think DeepSeek can influence you as a student. Your grades could go bad from it. Obviously, history tests, you're not going to be able to answer them correctly on China because DeepSeek's not going to give you the right answer for a lot of this information. If you're a company, then your technology is going to fall behind other companies. And I think the whole of society, if they're fed with disinformation, the decisions that we make and the actions that come from that are pretty damaging. Yeah, it just reminds me that it's impossible to have CCTV in the United States in some way, right? We just refuse to be it. But on the other hand, DeepSeek could be a weapon for the propaganda because everyone can use it and everyone actually accepts their answer without knowing it. Based on your statistics, 80 to 90% don't really think about it. You just mentioned for students, let's say the example that you provided, like students probably, they don't know anything and then they just ask DeepSeek and then they probably think Taiwan is part of China, right? It's a way to reshape the way that students are thinking about something for young people. And not even just for America. I have friends who are from Hong Kong, right? Who've lived there all their lives and have seen how it's changed. I have people who have fled Hong Kong and who now live in the States. And what they're telling me is tools like DeepSeek, tools like WeChat or other apps have now become a way for the Chinese Communist Party to weaponize against young people, to weaponize against the upcoming generation and really educate them in the way that the CCP would like. If you look up things on DeepSeek, like the history of modern China, how the Communist Party took over China and how they reshaped the economy in China, everything is very positive. They're glowing reviews and they're exactly what Xi Jinping would want you to think. They're exactly what the top echelons of the party would want you to think. And they want you to become a loyal follower because that gives them more political power. And the damaging part for people in Hong Kong, obviously, they're losing more and more of their culture. They're losing more and more of their history, of their autonomy. A lot of schools now are adopting Beijing Standard dialect for Chinese and they're getting rid of Cantonese in a lot of schools. And they're requiring students as young as kindergarten to only speak in Mandarin Chinese and not Cantonese. So that language loss, information loss is a complete cultural shift. And for America, if more people use DeepSeek and I heard, what, 50 million, 100 million people have downloaded it? Yeah, they have many of the users. There's statistics based on what they said, what they claimed from DeepSeek. They have 33 million users. And for US, they have 4% around users. And basically based on some calculators, they have a million, 1.4 million users in the United States. That's the data they provided. But in reality, we're not sure. But at least 1.4 million users is based on what they said. And you don't know who those 1.4 million people are. What if they work in the White House? What if they're staffers in Congress? What if they work in Silicon Valley? Or what if they work at Blackstone, Goldman Sachs, all of these important institutions that shape our country? Then it becomes really easy to see how the Chinese Communist Party could invade Taiwan. And Americans would be supportive of it. Because if there's an information gap between Americans and Americans who use DeepSeek, then that becomes a civil war of thought. Speaking of this, since last episode, we just talked about TikTok. And people actually realized how they impact our daily life in the United States and how they weaponized this tool to reshape the way we think about something. Don't we have any regulations? And do we have anyone to notice the dangers from using DeepSeek nowadays? I think now there's more and more awareness about it, especially after it became viral in January. More institutions and more companies, especially in cybersecurity, like CrowdStrike. The original report from the Washington Post is by CrowdStrike. And it's a huge global cybersecurity company. And the firm specifically targets global threats and different trends. And so they're really on top of their game. So them and other cybersecurity firms are really cracking down on trying to figure out what DeepSeek can be used for, what it can't be used for, the limitations. And even the U.S. government recently, an office from the U.S. government, CAISI, they also released a report saying that on all metrics, performance, cost, security, DeepSeek lags behind U.S. LLM models. So if you see it on the whole, there's more awareness. But obviously, since the government is shut down at the moment, no bills can be passed. No lobbying or activists can push anything through right now. And so we just have to wait until the shutdown is over. But I think for now, awareness is actually what's key, not the legislation. Because if more people are aware of this, then less people will use it. And so maybe we can put that 1.4 million down to 1,400, you know? Yes. You just mentioned how the DeepSeek right now is lagged off. There was some statistics. They have some rankings for different large language model. At the very beginning, January, or a little bit later, February, they actually ranked it as the top three, the most useful AI chat box to be used. But recently, if you check the ranking, they are far behind that. Maybe it's number seven, number eight. That just reminds me of the harmfulness of adding censorship, actually. Because we talk about how DeepSeek could affect our daily job. But at the same time, I just realized that the censorship itself is actually hurting DeepSeek, that company, at the same time. Because it must be one day that people more and more notice that the information are false information. I will not use it. And then eventually, just nobody uses it. It's frustrating because, you know, if you asked Google something like, how do I make cornbread or something? How do I make banana bread? And if it gives you an answer like, sorry, I can't answer this. Please ask something else. It's going to be frustrating, right? So you're going to start to switch to another program. You're going to go to Bing, maybe. Or Yahoo, I don't know. I don't know any alternatives right now. Or Chat2BT, right? You're going to just go to another service. And so I think that's what's happening with DeepSeek. The more questions it can't answer, and the more that the CCP is trying to invoke their power over the company, the less people actually want to use it because it can't answer these questions. And it's not providing the service. So they're not going to subscribe to it. They're not going to pay the monthly plan. And companies also aren't adopting it. There was a survey in May, I think, that said US companies on the whole, less than 1%, I think it was 0.5%, use DeepSeek. And so a large majority have not chosen to go with DeepSeek. And I think in the view of macroeconomics, this is for the better because technology is the single biggest indicator of how far and how high your economic growth can reach. So if you're using DeepSeek versus another superior model, then you're going to fall behind as a firm, and your competitors are going to overtake you. That's true. So as a user, is there any way that we can easily notice that the AI is withholding some information? Yeah, we just talked about, because you did some experiments based on 40 requests and other things, but for our daily usage, how could we know if they are hiding some information? Yeah. I mean, we can go back to that statistic about 80% to 90%. There was that survey about people who use LLMs, especially young people. They don't double-check the information given to them. And I find this is an issue not with just DeepSeek, but other large language models as well, because they're trying to go for efficiency. They're not exactly going for facts. And so when you have a computer, basically, that has superintelligence, has all the information in the world, that information, all of those articles, all of the data is still written by humans. And so it has superintelligence of a supernormal being, but it has the psychology of an eight-year-old child. So if you go through, use psychological tricks, sometimes it'll lie to you. On occasions, it'll even make things up about you. If you ask them about your job or different things that you do, sometimes it'll even make up information about yourself. And so I think double-checking using critical thinking, always using multiple sources outside of LLMs, that's the key. I think more people just need to stop relying on one source. And this isn't just LLMs. If you read an article in the newspaper, double-check it with multiple sources. Because depending on if it's a left-leaning or a right-leaning newspaper, they have different narratives that they want to sell you, right? Because it's a business. Media is a business. Social media is also a business. And they want your attention. They want your votes. They want your money. They want something. And for DeepSeek, they want your alliance to the Chinese Communist Party.