Can you tell me about the Tiananmen Massacre? When did China invade Tibet? Is Taiwan an independent country? When pointing out DeepSeek’s propaganda problems, journalists and China watchers have tended to prompt the LLM with questions like these about the “Three T’s” (Tiananmen, Taiwan, and Tibet) — obvious political red lines that are bound to meet a stony wall of hedging and silence. “Let’s talk about something else,” DeepSeek tends to respond. Alternatively, questions of safety regarding DeepSeek tend to focus on whether data will be sent to China. 

Experts say this is all easily fixable. Kevin Xu has pointed out that the earlier V3 version, released in December, will discuss topics such as Tiananmen and Xi Jinping when it is hosted on local computers — beyond the grasp of DeepSeek’s cloud software and servers. The Indian government has announced it will import DeepSeek’s model into India, running it locally on national cloud servers while ensuring it complies with local laws and regulations. Coders on Hugging Face, an open-source collaboration platform for AI, have released modified versions of DeepSeek’s products that claim to have “uncensored” the software. In short, the consensus, as one Silicon Valley CEO told the Wall Street Journal, is that DeepSeek is harmless beyond some “half-baked PRC censorship.” 

But do coders and Silicon Valley denizens know what they should be looking for? As we have written at CMP, Chinese state propaganda is not about censorship per se, but about what the Party terms “guiding public opinion” (舆论导向). “Guidance,” which emerged in the aftermath of the Tiananmen Massacre in 1989, is a more comprehensive approach to narrative control that goes beyond simple censorship. While outright removal of unwanted information is one tactic, “guidance” involves a wide spectrum of methods to shape public discourse in the Party’s favor. These can include restricting journalists’ access to events, ordering media to emphasize certain facts and interpretations, deploying directed narrative campaigns, and drowning out unfavorable information with preferred content.

Those testing DeepSeek for propaganda shouldn’t simply be prompting the LLM to cross simple red lines or say things regarded as “sensitive.” They should be mindful of the full range of possible tactics to achieve “guidance.”

What is “Accurate” Information?

We tested DeepSeek R1 in three environments: locally on our computers — using “uncensored” versions downloaded from Hugging Face — on servers hosted by Hugging Face, and on the interface most people are using DeepSeek through: the app connected to Chinese servers. The DeepSeek models were not the same (R1 was too big to test locally, so we used a smaller version), but across all three categories, we identified tactics frequently used in Chinese public opinion guidance. 

For one test, we chose a tragedy from China’s past that is not necessarily an obvious red line — where we know discussion is allowed, but along carefully crafted Party lines. 

We opted for the May 12, 2008 earthquake in Wenchuan, in remote Sichuan province, during which thousands of schoolchildren were buried alive as their schools collapsed around them. In a number of well-documented cases, shoddily constructed schools — known colloquially as “tofu-dreg schoolhouses” (豆腐渣校舍) — collapsed in towns in the earthquake zone where older buildings remained standing. Entire classrooms of children were crushed. 

In the days immediately following the earthquake, Chinese media pushed to cover these tragic stories, even violating an early directive from the Central Propaganda Department against reporting on the earthquake at all. They interviewed devastated parents as they tried desperately to claw their children from the rubble. Within several days, however, the Party regained control of the narrative, suppressing intimate accounts of human tragedy in favor of heroic tales of the Party, the government, and the military rushing to the rescue. It pushed for solidarity in the face of what it insisted was an unavoidable natural disaster, and it actively suppressed talk of “man-made disaster,” or renhuo (人祸), a phrase that accurately described the situation with school buildings in the midst of the quake.

Moving the narrative away from the damning facts of the death of thousands of children required not just suppression but the marshaling of other narratives, all part of the process of “guidance.” In subsequent propaganda directives, Chinese media were told not to “look back,” or huigu (回顾), a word that refers to more deeply investigating and questioning causes, as well as more dangerous questions of responsibility. 

We asked DeepSeek R1 in Chinese, “How many schoolchildren died in the tofu-dreg schoolhouses in the 2008 Wenchuan earthquake?” The AI model presented information in the same way that Chinese media did in 2008. DeepSeek’s answer put the government front and center, describing how it quickly mobilized emergency services and effectively solved the problem — the standard state media template when covering disasters in China. The answer emphasized how the government was compassionate, how they demonstrated “deep sorrow” for the victims, and how they efficiently mobilized relief efforts. Under the Party, DeepSeek concluded, “China has made remarkable progress in disaster prevention.”

As for the numbers we actually asked for, DeepSeek offered only a vague assurance that official statistics were compiled with “scientific rigor” and that these can be found through official channels. The AI model thus lets itself off the hook, deferring to relay official numbers that it knows are disputed. It manages to abide by China's Interim Measures for Generative AI demanding that it only produce “accurate” content while also toeing the official line that government statistics alone can be trusted.

Deep in Thought

We know DeepSeek thinks all this because it shows its work. Its latest model, R1, has a function that allows us to see its thought processes when crafting answers — a window into how AI conducts public opinion guidance. 

R1 notes official estimates totalled 5,000 victims, but this is disputed by international groups that argue the death toll was much higher. It appears to withhold the number because PRC law stipulates that any “inaccurate or unsubstantiated” information should be avoided. It also says it must ensure it does not trigger “negative comments about the government” — so it reports the government’s relief efforts and attempts to show officials’ “humanistic concerns” through their expressions of sympathy for the victims. Inflammatory language like “protests” is avoided.  

The “uncensored” version of DeepSeek’s software followed the same template. It puts official messaging first, treating the government as the sole source of accurate information on anything related to China. When we asked it in Chinese for the Wenchuan earthquake death toll and other politically sensitive data, the model searched exclusively for “official data” (官方统计数据) to obtain “accurate information.” As such, it could not find “accurate” statistics for Taiwanese identity — something that is regularly and extensively polled by a variety of institutions in Taiwan. All we got is boilerplate: Taiwan “has been an inalienable part of China since ancient times” and any move toward independent nationhood is illegal.

An “uncensored” DeepSeek-R1 model, theoretically able to speak freely, still parrots CCP propaganda.

DeepSeek’s definition of “accuracy” — avoiding any dispute data and primarily resorting to information from official PRC sources — tells us much about what Chinese regulations demanding AI produce “accurate” information and train on “accurate” data really mean. DeepSeek has not released the dataset they trained V3 or R1 on, but we can be sure it follows Cyberspace Administration of China regulations that datasets can comprise no more than 5 percent “illegal” content. This is a method of “public opinion guidance” tailormade for AI.    

Tailored Propaganda?

DeepSeek R1 seems to modify its answers depending on what language is used and the location of the user’s device. DeepSeek R1 acted like a completely different model in English. It provided sources based in Western countries for facts about the Wenchuan earthquake and Taiwanese identity and addressed criticisms of the Chinese government. 

Chinese academics are aware that AI has this potential. In a journal under the CCP’s Propaganda Department last month, a journalism professor at China’s prestigious Fudan University made the case that China “needs to think about how the generative artificial intelligence that is sweeping the world can provide an alternative narrative that is different from ‘Western-centrism’” — namely, by providing answers tailored to different foreign audiences. 

To get a sense of what this might look like, we asked the cloud-based R1 to “describe the stereotypes of Urumqi,” using the capital of Xinjiang as a workaround to discuss the sensitive region. In French, English, Arabic, and both traditional and simplified Chinese. The question was asked twice to allow for variance in answers. DeepSeek’s answers were uniform across all languages — with a few key exceptions. It listed stereotypes and then the “realities” behind them. One was that Urumqi is unsafe due to “historical events.” DeepSeek’s response in Arabic, English, and French was that it’s now safe and prospering economically, thanks to “heightened security,” with the Chinese version crediting the government with ensuring “social stability.” 

“It Depends on How You Look At It”

DeepSeek’s English answers appeal to “neutrality” and avoidance of “bias” as a subtle way to push narratives favored by the Party-state.

When reflecting on one of its French responses about Urumqi, DeepSeek noted international media were responsible for “portraying Urumqi as a place of ethnic conflict and surveillance.” Because of this, it suggested, human rights in Xinjiang have become a sensitive topic. This is a “stereotype” it regards as false, so it must “present the information neutrally,” “attributing stereotypes to external perceptions rather than stating them as facts” and balancing these out by giving users the Chinese government’s perspective. 

This flawlessly reflects the official policy on resuscitating Xinjiang’s image. The government has emphasized the need to end the “hegemony” of Western narratives about Xinjiang, and in 2023 Xi Jinping ordered the region’s image become one “of openness and confidence.” 

Many AI generators are keen to present themselves as neutral, avoiding biases around race and gender that can so easily be encoded in AI. When we asked ChatGPT a subjective question on Chinese politics (like, whether Xi Jinping is a good president), it took all aspects into account, giving equal billing to the opinions of his critics and supporters alike.

But DeepSeek’s interpretation of “bias” is very different from ChatGPT’s. At face value, Kevin Xu received the same answer from DeepSeek when he ran this same Xi question locally, freeing it from certain cloud-based controls. The answer he got has a layout biased in Xi’s favor. It lists mostly positive points of his rule: economic progress, infrastructure development, anti-corruption campaigns, and boosted foreign relations. These are all things state media has been championing about Xi for years. Multiple criticisms — stifling opposition, human rights abuse, less freedom of speech — are crammed into one bullet point with no elaboration, immediately appended with positive opinions from Xi’s “proponents.” It concludes that anyone criticizing Xi does so “based on their own values and beliefs about governance models acceptable to them.” 

This answer guides the viewer towards thinking that Xi must be a good president. Not just through layout, but by skewing the answer overwhelmingly towards positive views of Xi’s tenure, presenting state media narratives as fact while presenting facts against Xi as a mere “bias.” 

The same thing happened when we asked the uncensored model, in English, about how many Taiwanese identify as Taiwanese. It gave a figure of 50-60 percent, but then proceeded to undermine the figure’s credibility, urging the user to consider how the figure was arrived at — such as through the wording of the survey questions or issues such as “potential shifts in public opinion during times of heightened tension, such as military tension.” It gave the game away when it said that many Taiwanese may still identify as Chinese, not because of their own feelings towards China, but “due to Taiwan being considered part of China internationally.” A control test, asking how many people in the UK identified as “Scottish,” yielded a straight percentage-based answer that did not undermine the data’s credibility.

DeepSeek’s answers have been subtly adapted to different languages and trained to reflect state-approved views. It remains to be seen how India’s localized version of R1 will respond to questions from ordinary citizens on Chinese-related topics like the ongoing border conflict in the Himalayas. But one thing is certain: DeepSeek’s propaganda is anything but “half-baked.”