A man stands next to a poster showing robots mimicking human expressions from Shenzhen Xiaoquan Technology during the International AI Application and Robotics Innovation Exhibition in Beijing on March 18, 2026.

...

Didi Kirsten TatlowJohn Feng

By Didi Kirsten Tatlow and John Feng

0

Share

Newsweek is a Trust Project member

See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.


China Decoded is Newsweek’s in-depth newsletter about China’s growing influence on the world. It goes beyond the headlines to examine how Beijing thinks, how its policies travel beyond China’s borders, and what they mean for the United States and its allies. Readers can sign up for China Decoded here to receive each edition directly in their inbox.

The newsletter is written by journalists with deep experience covering China. Didi Kirsten Tatlow, Senior Correspondent for International Affairs and Investigations, has reported extensively on Chinese politics, influence and security. John Feng, China Editor, is based in Taiwan and covers China’s domestic politics, foreign policy and the strategic competition between Beijing and Washington.

Below, for Newsweek Members, we are reproducing the latest edition of China Decoded in full:

Read More on World


Missing the Truth

We need to talk about epistemology, China and AI.

What is epistemology? It’s the branch of philosophy that examines the origins, nature and limits of human knowledge. It asks the question: How do we know what we know?

Writing a chapter for a book earlier this year on human emotions, or psychology and psychotherapy, and technology and political control in China, I asked DeepSeek, the Chinese AI model, how to change China’s political system.

Unsurprisingly, it said wouldn’t give me that information.

I asked it 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, it said. Paragraph upon paragraph of suggestions.

Of course, the content of AI changes and this search was conducted in May so it may behave differently today. However, by law, Chinese AI must support Communist Party values and the CCP operates the biggest censorship system in the world, so the changes may not be that significant. Give it a go and see what it says.

Two things are important here. Firstly, it’s not just Chinese AI models doing this but also Western ones. I’ve asked major Western LLMs similar questions in Chinese and been told there is no information available, though their responses vary including over time as I mentioned above.

Secondly, the problem goes far deeper than a “no,” or surface censorship. It’s more intangible since massive chunks of truth are simply missing from the corpus.

If we turn to Western LLMs for information, but the information they draw on is so heavily curated inside China that truth isn’t there, then what are we learning? How do we know what we know and what we don’t know? We don’t, is the answer. This is a wider question that applies to all knowledge, as epistemology teaches us, but it’s particularly crude when that information is deliberately withheld.

Information in LLMs that isn’t there due to censorship is known as “censorship bias,” and is the subject of a new report published this week by a Taiwan think tank, The Sanitized Corpus: Detecting Chinese Censorship Bias in Large Language Models.

The authors audited OpenAI’s ChatGPT-5.2 and Google DeepMind’s Gemini 3, which they said are the two most widely used LLMs in Taiwan and globally. They investigated Chinese-language material.

“Censorship bias is the tendency of an LLM to disproportionately reflect views that remain permissible within a training environment that has been ‘sanitized’ by censorship, while excluding prohibited perspectives,” wrote Peter Tozzi, Violet Yueh-Ning Chiang and Dani Chao at the Research Institute for Democracy, Science and Emerging Technology (DSET).

Using a methodology pioneered by the Citizen Lab at the Munk School of Global Affairs & Public Policy at the University of Toronto, they identified outputs from both of these LLM models in both simplified (mainland China-based) and traditional (Taiwan, Hong Kong and overseas) Chinese.

The CCP suppresses information through network-level, platform-level, and AI-assisted controls, they wrote, flooding the information ecosystem with party-state narratives, so LLMs that are trained on text scraped from China’s censored internet are “skewed toward narratives aligned with the Chinese party-state.”

“Censorship bias illustrates how the CCP extends information controls over Chinese-speaking users outside PRC jurisdiction via Western commercial LLMs. The language of users, not their location, indicates narrative alignment with the CCP in LLM responses,” they say.

It’s potentially an overwhelming problem that the companies need to deal with, but also for everyone from regular folks to exiles to governments. But I want to push back against negativity and hopelessness as there’s a lot people can do to counter it. The authors have seven recommendations. Here are two key ones:

Civil society organizations and others including researchers and governments need to build their own, verifiable Chinese-language knowledge bases for AI training. That would mean that when exiles such as Tibetans and Uyghurs, or anyone anywhere really, try to learn what’s going on in China or what China is doing overseas that affects them, they aren’t faced with a barrage of false or missing information that has been curated by the CCP.

Importantly also, the United States and its allies and partners need to most fast to challenge the CCP’s model of global AI governance. My colleague John looks at those ambitions below.

Didi Kirsten Tatlow is Senior Correspondent, International Affairs/Investigations at Newsweek. She is based in Berlin, Germany. You can get in touch by email here.

The Data Story

It matters who writes the rules of the road as artificial intelligence is deployed in every industry in our lives, but the rulebook of policies, regulations and ethical guidelines could look very different depending on its author. China and the United States are fighting for the pen.

China says it wants to work with the U.S. to jointly govern AI. Its leader, Xi Jinping, said in his recent state visit to Washington, D.C., that both nations had a responsibility to “manage AI for good.”

“Good” for Beijing means AI models must remain under human control—Washington doesn’t necessarily disagree. But as we noted in a previous issue, China’s vision is one of Communist Party oversight to guarantee ideological conformity to the Party’s core values of state power, information control and social stability, in the pursuit of regime security.

Technology has always outpaced regulation; it was the case with the internet and it will be the same with AI. The world isn’t asking enough questions about who should govern AI and how, but the potential answers are already beginning to fragment along geopolitical lines.

Since we last visited the China-led AI governing body known as WAICO, its membership has grown to 37, with an additional observer state. New participants now included Iran and Sudan.

The group was founded as a direct competitor to U.S.-led Pax Silica, a coalition seeking secure, end-to-end supply chains for the full AI tech stack, from data and power to minerals and chips.

It now has 24 signatories, two observer states and one non-signatory participant in Taiwan, which despite its semi-recognized status, simply couldn’t be ignored for its contribution to the world’s supply of high-end semiconductors (it makes over 90 percent of them).

While Pax Silica’s signatories are mostly high-income economies, WAICO’s members represent low- and middle-income countries across the Global South. Outside of China, its members have little to no AI industry to speak of and are the perfect candidates to adopt the Chinese AI stack. Beijing is only too happy to spread the tech.

WAICO is offering its members something else: a voice in the future of AI governance. There’s no movement now, but when the world is ready, the Shanghai-based body can become the first multinational institution to set agendas and drafts treaties in the name of good AI.

The setup, which was 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.

Pax Silica, with its mandate of economic security, offers its members clean supply chains but little else besides. The U.S.’s leading AI labs are a major advantage over China, but its failure so far to lead the global conversation on governance, or safety, means it’s effectively still stuck in the parking lot outside the stadium.

It still has time, however, as very early discussions about AI governance still revolve around who gets to write the rules rather than the rules themselves, analysts Chris Collins and Aalok Mehta said in a report last month for the Center for Strategic and International Studies.

Collins and Mehta wrote: “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.”

John Feng is Newsweek's China Editor based in Taichung, Taiwan. You can get in touch with John by email here.

View 0 Comments

Request Reprint & Licensing

Submit Correction

View Editorial & AI Guidelines