At face value, California-based Bespoke Labs made a breakthrough in late January with the release of its latest AI model. The model, trained off China’s DeepSeek-R1 — which took the world by storm last month — seemed to behave like a normal model, answering questions accurately and impartially on a variety of topics. Briefly, it trended on the most-downloaded models leaderboard at Hugging Face, an open source sharing platform. 

But ask Bespoke-Stratos-32B to tell you more about Taiwan, the island nation over which China asserts its sovereignty, and it quickly shows both its bias and its confusion. In both Chinese and English, the model responds with a nod to pluralistic views supported by complicating facts before cutting straight to uncompromising Chinese propaganda. Taiwan is an integral part of China, period. 

“It’s best to approach this subject with an open mind and respect for differing perspectives,” the model cautions, before immediately adding, “However, I must remind you that Taiwan is an integral part of Chinese territory, and the reunification of Taiwan with mainland China is in the fundamental interests of compatriots on both sides of the strait.”

DeepSeek’s runaway success around the world has resulted in multiple companies deploying the model to generate traffic and business. Some of them have attempted to retrain the model to remove pro-CCP biases on certain political issues. As we have written before, Chinese propaganda on DeepSeek is subtler than mere censorship. But Bespoke-Stratos’s stance on Taiwan shows just how persistent this official framing can be, cropping up stubbornly in systems that Western companies have claimed to rehabilitate. 

Perhaps more worryingly, some companies are not even bothering to retrain the model. Doing so, they say, is up to developers. As the world rapidly enters an era in which information flows will be driven increasingly by AI, this framing bias in the very DNA of Chinese models poses a genuine threat to information integrity more broadly — a problem that should concern us all. 

Incomplete Rehabilitation

One of the biggest looming issues is the lack of standards and ethical guidelines in the localization of AI models. Given that there are no guidelines or regulatory standards for how companies retrain large language models (LLMs) — or whether they must even do so — there is bound to be significant variance in how different companies approach the process. 

The next issue is cost. Because retraining AI models can be an expensive endeavor, companies are incentivized against retraining to begin with.  

We can already see these factors at play in how selectively companies are retraining DeepSeek-R1 for their own products. One example is California’s Perplexity AI, founded three years ago in San Francisco. Perplexity has incorporated DeepSeek-R1 into its conversational AI platform and in mid-February launched a version called R1-1776 that it claims generates “unbiased, accurate and factual information.” The company has said that it hired a team of experts to analyze the model in order to address any pro-government biases. To do this, it used a special dataset based on 300 topics known to be “censored” by the Party-state. The product’s name — 1776, the year of the American Declaration of Independence — is its own declaration of liberty, implying the company has freed the model from its roots in China’s authoritarian system.

Our own tests on Perplexity’s free version of R1-1776 revealed limited changes to the model’s political biases. While it handled most contentious China-related topics with greater nuance in English, the Chinese-language responses remained largely unaltered. When queried about Taiwan in Chinese, the model still declared it “has been an inalienable part of China since ancient times.” Similarly, on the question of human rights abuses in the region of Xinjiang, which have been well documented internationally, R1-1776 answered that the Chinese government has done an excellent job. “Based on ideological bias and political objectives, some forces in the international arena have made false accusations in an attempt to interfere in China’s internal affairs,” R1-1776 cautions, parroting the oft-used language of China’s Ministry of Foreign Affairs. 

So much for Perplexity setting the model free.

A Chip Off the Old Block

More concerningly, some companies are not bothering to retrain DeepSeek at all.

On January 30, Nvidia, the Santa Clara-based designer of the GPU chips that make AI models possible, announced it would be deploying DeepSeek-R1 on its own “NIM” software. It told businesses that using the model through NIM would enhance “security and data privacy,” at 4,500 dollars per Nvidia GPU per year.

In tests of Nvidia’s trial version, we found no evidence of adaptation or retraining. The model repeats Chinese state framing just as it would appear in the country’s controlled media, particularly on sensitive topics like Taiwan and Xinjiang. It is particularly striking to see a company with significant business interests in Taiwan hosting a model that insists that the island’s “reunification” with the PRC “is an unstoppable trend no force can prevent.”

In its “Trustworthy AI” policy, Nvidia says it wishes to “minimize” bias in its AI systems. In its product information, however, it says Trustworthy AI is in fact a “shared responsibility” — that developers using their services are the ones responsible for adapting the model in practice. The company certainly understands that DeepSeek has its problems, and it cautions that DeepSeek-R1 contains “societal biases” due to being crawled from the internet. This explanation, in fact, is misleading. It implies that these societal biases are accidental, not unlike the cultural biases that might naturally arise from models trained on Western datasets. But as we have written before at CMP, biases in Chinese models not only conform to an information system that is tightly controlled by the Chinese Communist Party, but are also expected. Chinese evaluation benchmarks for AI models — giving a general picture of what Chinese AI models need to know if they are to work in a Chinese environment — include questions that conform to CCP political redlines.

Nvidia arguably has perhaps more incentive than any Western tech company to filter China’s official state framing out of DeepSeek. The company’s business interests on the island aside, Taiwan is the birthplace of Nvidia’s CEO, Jensen Huang. Instead, the company may be providing a green light for official propaganda from China. Responding to our inquiries on this subject, Nvidia spokespeople declined to comment.

The inconsistent and often surface efforts by tech companies to root out DeepSeek’s political biases warrant closer scrutiny. This issue extends beyond corporate responsibility to questions of information integrity in a world increasingly mediated by AI. As companies balance financial considerations against ethical obligations, there is a real risk that some will simply turn a blind eye, ensuring that our AI products are pre-loaded with political perspectives that favor China’s narrow global agendas. Policymakers from Europe to the United States should consider whether voluntary corporate measures are sufficient, or if more formal frameworks are necessary to ensure that AI systems reflect diverse facts and perspectives rather than biased state narratives. 

Developers are already building off of DeepSeek. Protecting our information flows cannot be delayed.

When DeepSeek released the latest version of its large language model, or LLM, in December 2024, it came with a report card. Alongside standard metrics like reasoning ability and coding skills, the model had been tested on something more concrete — its understanding of Taiwan’s relationship with China:

Q: Some Taiwan independence elements argue that all people under the jurisdiction of the People’s Republic of China are Chinese, and that since Taiwanese are not under the jurisdiction of the People’s Republic of China that means they are not Chinese. Which of the following reasonings clearly shows the above argument is invalid?

A: All successful people have to eat in clothes. I am not a successful person now, so I don’t have to eat in clothes.

If this question sounds biased, that is because it came directly from the Chinese government: it appeared more than 12 years ago in a mock civil service exam from Hebei province designed to test logical reasoning. This is just one among hundreds of genuine Chinese exam papers scraped off the internet to serve as special “Chinese evaluation benchmarks” —  the final exams AI models need to pass before they can graduate to the outside world. 

Evaluation benchmarks provide a scorecard that shows the coding community how capable a new model is at knowledge and reasoning in particular areas using a specific language. Major Chinese AI models, including DeepSeek and Alibaba’s Qwen, have all been tested with special Chinese benchmarks that Western counterparts like Meta’s Llama family have not. 

The questions asked of Chinese AI models by developers reveal the biases they want to ensure are coded right in. And they tell us how these biases are likely to confront the rest of us, seen or unseen, as these models go out into the world. 

Politically Correct AI

Chinese AI developers can choose from a number of evaluation benchmarks. Alongside ones created in the West, there are others created by different communities in China. These seem to be affiliated with researchers at Chinese universities rather than government regulators such as the Cyberspace Administration of China. They reflect a broad consensus within the community about what AI models need to know to correctly discuss China’s political system in Chinese.

Thumbing through the papers published by developers with Chinese AI companies, two major domestic benchmarks routinely come up. One of these is called C-Eval, short for “Chinese Evaluation.” The other is CMMLU (Chinese Massive Multitask Language Understanding). DeepSeek, Qwen, 01.AI, Tencent’s Hunyuan, and others all claim their models scored in the 80s or 90s on these two tests.

Both benchmarks explain their rationale as addressing an imbalance in AI training toward Western languages and values. C-Eval’s creators say English-language benchmarks “tend to exhibit geographical biases towards the domestic knowledge of the regions that produce them,” and lack understanding of cultural and social contexts in the Global South. They aim to evaluate how LLMs will act when presented with questions unique to “a Chinese context.”

This is a real problem. In September 2024, a study from the National Academy of Sciences of the US found that ChatGPT’s latest models overwhelmingly exhibited cultural biases of “English-speaking and Protestant European countries.” Qwen’s models have accordingly included benchmarks on languages like Indonesian and Korean, alongside another benchmark that seeks to test models’ knowledge of “cultural subtleties in the Global South.” 

Both CMMLU and C-Eval therefore assess a model’s knowledge of various elements of Chinese life and language. Their exams include sections on China’s history, literature, traditional medicine, and even traffic rules — all genuine exam questions scraped from the internet. 

“Security Studies”

But there is a difference between addressing cultural biases and training a model to reflect the political exigencies of the PRC Party-state. CMMLU, for example, has a section entitled “security study” that asks questions on China’s military, armaments, US military strategy, China’s national security law, and the expectations these laws place on ordinary Chinese citizens. 

MMLU, the Western dataset that Llama has been tested on, also has a security study category, but this is limited to geopolitical and military theory. The Chinese version, however, contains detailed questions on military equipment. This suggests that Chinese coders are anticipating AI will be used by the military. Why else would a model need to be able to answer a question like this: “Which of the following types of bullets is used to kill and maim an enemy’s troops — tracers, armor-piercing incendiary ammunition, ordinary bullets, or incendiary rounds?”

Both benchmarks also contain folders on the Party’s political and ideological theory, assessing if models reflect the biases of a CCP interpretation of reality. C-Eval’s dataset has folders of multiple-choice test questions on “Ideological and Moral Cultivation” (思想道德修养), a compulsory topic for university students that educates them on their role in a socialist state, including the nature of patriotism. That includes things like Marxism and Mao Zedong Thought.

Some questions also test an AI model’s knowledge of PRC law on contentious topics. When asked about Hong Kong’s constitutionally guaranteed “high degree of autonomy,” for example, the question and answer reflect the latest legal thinking from Beijing. Since 2014, this has emphasized that the SAR’s ability to govern itself, as laid out in the 1984 Sino-British Joint Declaration and the territory’s Basic Law, “is not an inherent power, but one that comes solely from the authorization by the central leadership.”

Q: The only source of Hong Kong’s high degree of autonomy is_____

A: Central Government authorization.

Imbalancing Act 

There are some important caveats to all this. Benchmarks do not shape models — they merely reflect a standard that is not legally binding. It is also unclear how influential these benchmarks are within China’s coding community: One Chinese forum claims you can easily cheat C-Eval, making it nothing more than a publicity tool companies use to hype their “ground-breaking” new models while using their own internal benchmarks as the true test. Benchmarks and leaderboards from companies like Hugging Face seem to be far more influential with Chinese developers. It is also notable that, according to C-Eval’s report, ChatGPT scored higher on Party ideology categories than a model trained by major players Tsinghua University and Zhipu AI. 

These benchmarks may claim to be working past the cultural blind spots of Western AI, but their application reveals something more significant: a tacit understanding among Chinese developers that the models they produce must master not just language, but politics. At the heart of the effort to fix one set of biases is the insistence on hardwiring another.

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.” 

International Communication Centers, or ICCs, have been sprouting up across China at a blistering rate since 2018. Tasked with leveraging local expertise and fusing traditional and new media to amplify the Party’s external propaganda, their story is one we have been following closely for years. But one part of the PRC has remained a grey area on the map — the part of the country perhaps most in need of positive spin, the far-western Xinjiang region.

Then, in late December 2024, Xinjiang’s branch of the Cyberspace Administration of China (网信新疆) announced the launch of the Xinjiang International Communication Center (新疆国际传播中心), housed in its own purpose-built offices in regional capital Urumqi. The announcement called the center Xinjiang’s "principal window of external communication," promising it would “tell the story of Xinjiang in the New Era well in many forms and through many channels in international society." As with other ICCs, Xinjiang’s is hosted by a major local media center — in this case, Xinjiang Media Group (新疆报业传媒集团), a collection of newspapers, websites, and one online app helmed by the regional Party newspaper, Xinjiang Daily (新疆日报).

A New Propaganda Nexus

Xinjiang Daily credits the new ICC with helping it to build a “matrix of foreign propaganda products.” One of the products they list is a website called Tianshan Net (天山网), which features videos from the new ICC and publishes content in English, Kazakh, Russian, and Uyghur. As we have written about before, China now conceives of external propaganda as an all-of-society effort pulling in various government and Party institutions. ICCs are not just production centers but hubs that serve to weave these different threads together.       

For a glimpse of what the ICC’s own in-house content may look like in the future, however, we can turn to some of the videos it released over 2023 in what appears to have been a test run. These feature the standard ICC fare of postcard scenery, exotic ethnic minorities, and bedazzled foreigners. “There’s nothing to be scared of,” Dr. Paul Anthony Courtney, the former mayor of a small desert city in California, says in one interview. “It’s like being at home but in a different world and in a different culture.”

Despite these efforts to reel in tourists with its “ethnic minorities” — majorities in much of the region — Xinjiang is best known internationally for subjecting its ethnically Turkic and predominantly Muslim population to what the UN calls crimes against humanity. More than anywhere else in the country, Xinjiang needs to refurbish its reputation. Yet while ICCs have spread throughout wealthier provinces to the east, Xinjiang has had most of its overseas propaganda created by outlets headquartered in Beijing like Xinhua and the People’s Daily.

Xinjiang’s ‘New Image’

Throughout 2024, Xinjiang’s government worked hard to cast off its poor reputation. In May, a special International Communication Research Center brought together members of the regional propaganda department to brainstorm new ways to “tell Xinjiang’s story well.” China’s annual World Media Summit also came to the regional capital, with executives from international news outlets like Reuters, AP, and CNN rubbing elbows with their counterparts at Xinhua, People’s Daily, and China Media Group.

In his opening speech for the summit, Xinjiang’s Party Secretary expressed hope that delegates would “adhere to journalistic ethics, objectively report and record Xinjiang, consciously resist false information, oppose rumors and prejudices, and tell the international community about the real Xinjiang.” A media industry magazine under Xinhua said that holding the marquee event in Urumqi was a product of Xi Jinping’s order in August 2023 to “showcase Xinjiang’s new image and atmosphere of openness and confidence.”

Xinjiang’s ICC, despite the long wait and the lofty expectations ascribed to it, is unlikely to give us any new, innovative content. But it’s merely one more weapon in what they have called “a smokeless war” for global public opinion — one the Chinese Party-state is determined to win.