There is a strange and uncomfortable moment when you ask a machine a simple question and realize the answer has less to do with the machine than with the world that built it. The internet promised universal access to knowledge; artificial intelligence promised to make sense of it all. But in an age when China’s Communist Party operates the world’s largest censorship apparatus and is racing to define what artificial intelligence can say, the question at the heart of epistemology—the branch of philosophy that asks how we know what we know—has moved out of seminar rooms and into the daily lives of journalists, exiles, policymakers, and ordinary users. This is the territory of China Decoded, Newsweek’s in-depth newsletter about Beijing’s growing influence on the world. Written by journalists who have spent careers watching China up close, it goes beyond the headlines to examine how Beijing thinks, how its policies travel beyond China’s borders, and what those developments mean for the United States and its allies. Senior correspondent Didi Kirsten Tatlow, based in Berlin, has reported extensively on Chinese politics, influence and security; China Editor John Feng, based in Taichung, Taiwan, covers Beijing’s domestic politics, foreign policy and the strategic competition between Beijing and Washington. In this latest edition, two seemingly separate stories turn out to be two sides of the same coin. One is an unsettling detective story about missing knowledge inside large language models; the other is a quiet but urgent diplomatic race over who gets to write the rules of artificial intelligence. Together they suggest that the most consequential censorship is no longer just about deleting words—it is about shaping what the world believes is worth asking in the first place. If the machines that mediate our understanding of reality are trained on deliberately emptied archives, and if the rules for those machines are being written by a party-state with every incentive to keep them empty, then all of us are trying to read the news through a fog that someone else created. The stakes are not merely academic; they are about whether we can still trust our own tools of knowing.
Consider the experiment Didi ran earlier this year. While writing a chapter on human emotions, psychology, psychotherapy, technology and political control in China, she asked DeepSeek, the Chinese AI model, how to change China’s political system. The chatbot politely refused: it wouldn’t give her that information. Then she asked the same question about the United States—and the advice poured out. Many people want to change the U.S. system, it said. Political organizing on the ground can work. Paragraph after paragraph of suggestions. The contrast is both absurd and chilling. Of course, AI behavior shifts; her search was conducted in May, and DeepSeek may respond differently today. But Chinese law requires AI models to support Communist Party values, and the CCP runs the world’s biggest censorship system, so don’t expect the silence to vanish. The deeper problem, though, is not the “no.” As Didi points out, it’s the void around it. She has asked major Western large language models similar questions in Chinese and been told there is no information available—even though there is. The models are trained on text scraped from a vast ecosystem that has been deliberately cleaned. Massive chunks of truth are simply not in the corpus. If you turn to a Western LLM for information about China, but the information it draws on has been so heavily curated inside China that the truth isn’t there, then what exactly are you learning? You don’t know what you don’t know—and that is the answer. This is what epistemologists have always warned: knowledge depends on what is absent as well as what is present. But when the absence is engineered by a party-state that wants to control the narrative, it stops being abstract. It becomes a quiet form of power, one that operates not by shouting down dissent but by making sure dissent never appears in the first place. The machine isn’t lying; it is faithfully reproducing a library from which the truth has been removed, and telling us that emptiness is all there is.
This week, a Taiwan-based think tank gave this problem a name: “censorship bias” and quantified it in a new report, The Sanitized Corpus: Detecting Chinese Censorship Bias in Large Language Models. Published by the Research Institute for Democracy, Science and Emerging Technology (DSET), and written by Peter Tozzi, Violet Yueh-Ning Chiang and Dani Chao, the report audits OpenAI’s ChatGPT and Google DeepMind’s Gemini—the two most widely used large language models in Taiwan and globally—looking specifically at Chinese-language material in both simplified and traditional scripts, and using a methodology developed by the Citizen Lab at the University of Toronto’s Munk School of Global Affairs and Public Policy. The authors found that these Western-made models consistently produce outputs skewed toward narratives aligned with the Chinese party-state. The reason is structural. The CCP suppresses information through network-level, platform-level, and AI-assisted controls, flooding the information ecosystem with party-state narratives. A model trained on text scraped from China’s censored internet isn’t deliberately choosing to censor; it is treating Beijing’s official version as the whole universe of acceptable Chinese-language thought. The report makes an especially unsettling point: censorship bias extends Chinese information control beyond China’s borders. A user in Taipei, London or San Francisco who types in Mandarin may receive answers aligned with Beijing even when the question has nothing to do with China. It is language, not location, that signals to the model which narratives to follow. Chinese-speaking users outside PRC jurisdiction are being quietly steered by Western commercial LLMs toward party-state viewpoints. In other words, Beijing’s censorship doesn’t stop at the Great Firewall. It follows Mandarin speakers wherever they are, riding inside algorithms that present themselves as neutral and global. The report doesn’t accuse these companies of deliberate collusion; it warns that their products have nevertheless become delivery systems for an authoritarian information order that most of their users never see. And because the bias is baked into training data rather than visible in a single answer, it is far harder to detect—let alone to challenge—than a straightforward government block. That is what makes the problem potentially overwhelming, for companies, for ordinary people, for exiles, and for governments alike. But the authors are not content to leave readers in despair; they insist there are things people can do.
Faced with such a problem, it would be easy to sink into hopelessness. But the report’s authors aren’t interested in merely diagnosing the disease; they offer seven recommendations, and two of them deserve special attention. The first is aimed at civil society organizations, researchers, and governments: they need to build their own, verifiable Chinese-language knowledge bases specifically for AI training. If that happens, an exiled Tibetan or Uyghur—or anyone anywhere—trying to understand what is actually happening in China, or what Beijing is doing overseas, would not be confronted with a wall of false or missing information curated by the CCP. There would be an alternative archive, one built on evidence rather than state directives, ready to be used to train machines that answer questions honestly. This is not a small technical fix; it is a project of collective memory, an effort to ensure that the next generation of AI does not inherit a world in which only the party-state’s version of Chinese reality exists. The second recommendation is more urgent and perhaps more difficult: the United States and its allies and partners need to quickly challenge China’s model of global AI governance. That doesn’t mean repeating talking points about free speech; it means offering a credible alternative—one that demonstrates how transparency, consent, and human rights can be built into AI systems, rather than treated as afterthoughts. Between those two recommendations lies a broader message: the age of AI is not a passive catastrophe. There is still room for human agency, for funding independent knowledge, for pressuring companies to audit their models, for governments to regulate with clarity. But the window is narrowing. The information ecosystems that will shape the next generation of machines are being built right now. If democratic societies don’t seed them with verifiable Chinese-language information, Beijing will happily supply the silence—and call it consensus. The battle is not just over what machines can do; it is over what they are allowed to know, and how they will answer the oldest human question: How do we know what we know?
The Data Story half of this edition shifts from missing words to missing opportunities. As John Feng writes, artificial intelligence is being deployed in every industry, but the rulebook that governs it—the policies, regulations, ethical guidelines—could look radically different depending on who writes it. China and the United States are fighting for the pen. In Beijing, that fight is framedasi cooperation. Xi Jinping said during his recent state visit to Washington that both nations had a responsibility to “manage AI for good.” But Beijing’s definition of “good” is tied to Communist Party oversight and ideological conformity: AI must remain under human control, yes, but that control must serve state power, information control and social stability, all in the pursuit of regime security. Washington might agree that humans should stay in charge, but it certainly doesn’t interpret “good” the way Beijing does. The world, meanwhile, isn’t asking enough questions about who should govern AI. Technology has always run ahead of regulation—that was true of the internet and it will be true of AI—and the potential answers are beginning to fracture along geopolitical lines. Since China Decoded last looked at WAICO, the Shanghai-based body established as China’s answer to Western-led governance, membership has grown to 37 countries plus an observer state, with Iran and Sudan among the newcomers. WAICO was designed as a direct competitor to Pax Silica, the U.S.-led coalition seeking secure, end-to-end supply chains for the full AI tech stack—from data and power to minerals and chips. Pax Silica now counts 24 signatories, two observer states, and one non-signatory participant: Taiwan, which, despite its semi-recognized status, couldn’t be ignored because it produces over 90 percent of the world’s high-end semiconductors. While Pax Silica’s signatories are mostly high-income economies, WAICO’s members are mostly low- and middle-income countries across the Global South. Most of them have little to no AI industry—and that is precisely what makes them perfect candidates to adopt the Chinese AI stack. Beijing is only too happy to spread the technology, and WAICO is offering its members something more: a voice in the future of AI governance. There is no movement yet, but when the world is ready, the Shanghai-based body can become the first multinational institution to set agendas and draft treaties in the name of good AI. That setup, endorsed by U.N. Secretary General António Guterres, is proof that China knows it must be ready at the starting line if it wants a shot at shaping global AI rules in its favor.
Set side by side, the two rival coalitions reveal an asymmetry. Pax Silica, with its mandate of economic security, offers its members clean supply chains but little else besides. The United States has leading AI labs, a genuine advantage over China, but it has so far failed to lead the global conversation on governance or safety. As analysts Chris Collins and Aalok Mehta wrote last month for the Center for Strategic and International Studies, the early discussions about AI governance still revolve around who gets to write the rules rather than what the rules themselves should be. They put the dilemma bluntly: “The United States cannot credibly propose international standards for regulating frontier AI while at the same time lacking a coherent national set of regulations. China has realized this and is therefore making a dedicated AI law. Therefore, developing a federal framework to regulate AI would give U.S. diplomacy a model to offer other countries.” The implication is that America’s biggest weakness in this race is not technological—it’s institutional. It is hard to persuade the world to trust your vision of AI when you haven’t been able to agree on one at home. There is still time. The governance conversation is young enough that the fundamental questions remain open, and very early positioning matters more than final details. But the trajectory is clear: China is showing up with an architecture in place and a growing coalition of Global South states waiting to adopt it. The United States, meanwhile, is still in the parking lot outside the stadium, brimming with talent but lacking a ticket. What ties Didi’s investigation and John’s analysis together is that both are ultimately about knowledge—who controls it, who withholds it, and who gets to say what counts as truth. The stakes could not be higher: if the machines that mediate our understanding of the world are trained on absent archives and governed by rules written by a party-state that profits from absence, then we all become exiles in our own minds, cut off from the very information we need to make sense of our world. The first step to reclaiming the truth is recognizing how much of it has already quietly disappeared—and then deciding, collectively, that we will build something to replace it.
