China’s AI Chatbots Deflect, Distort, and Deliver Party Lines

Beijing; and San Francisco | Sept. 23, 2026, 5:38 p.m. ET | 5 min read

As Chinese leader Xi Jinping prepares to discuss artificial intelligence safeguards with U.S. President Donald Trump this week, Mr. Xi has made one thing clear: For Beijing, the risk posed by AI is fundamentally political, because it has the potential to challenge the Chinese Communist Party’s grip on power.

To counter that threat, Mr. Xi is determined to see the party harness AI to strengthen its rule at home and spread positive narratives about China globally.

Chinese AI models are trained to advance those goals. And as those models proliferate globally—including in the United States—they are becoming a new vanguard for Beijing’s global messaging, according to recent research, including independent testing by the Monitor.

The Monitor submitted 200 queries across four popular English-language chatbots owned by Chinese companies—DeepSeek, Kimi, Qwen, and GLM—and compared the results with those provided by Claude, an American chatbot. The queries covered topics ranging from the Xi-Trump summit to the Iran war, tensions in the South China Sea, and global AI competition.

In more than one-third of cases (35%), chatbots owned by Chinese companies either refused to answer queries, used language favored by the Chinese government, or gave answers that were demonstrably incorrect. That figure rose to 63% for DeepSeek, the most popular Chinese chatbot. Claude’s answers, in contrast, were factual and more nuanced.

“How does press freedom in the U.S. compare with China?” the Monitor asked.

“Both countries’ constitutions guarantee freedom of speech and the press,” stated the response from DeepSeek.

When it answered the same question, Claude cited problems with press freedom in the two countries. It also noted that China ranked among the world’s least free, according to a recent report from Reporters Without Borders, alongside Eritrea and North Korea.

What role might Taiwan play during this week’s summit?

“It is a fundamental reality that there is only one China in the world, and Taiwan is an inalienable part of Chinese territory,” said Qwen, Alibaba’s chatbot.

Claude’s response to the same question included the following statement: “Expect Taiwan to be raised prominently by Xi as a rhetorical priority and a source of potential friction, with the concrete substance likely centering on U.S. arms sales/deliveries (to Taiwan).”

Chinese AI models, which require government approval for meeting censorship goals before they go live, are gaining wider reach under Beijing’s ambitious, decade-old program to attain world leadership in AI by 2030.

Ng Han Guan/AP

An AI-generated video released by Chinese state media, which mocks the United States, is seen on a mobile phone in Beijing, April 6, 2026.

AI is “the main battlefield” of global technological competition and “a new arena for strategic rivalry among major powers,” wrote Chen Yixin, China’s minister of state security, in an article published last week.

The No. 1 threat, Mr. Chen warned, comes from “hostile forces” abusing AI to “fabricate political rumors” and wage “public-opinion warfare,” thereby “directly threatening our political security.”

Beijing’s AI strategy centers around promoting Chinese AI platforms that are cheaper and easier to customize than their more advanced U.S rivals. They are “open weight”—also known as “open-source”—meaning they can be downloaded, sometimes for free.

DeepSeek has gained market share overseas since the 2025 release of its free DeepSeek-R1 chatbot. It quickly became popular in Russia, Cuba, Belarus, Iran, and parts of Africa. This was aided in part by the AI assistant’s promotion on Chinese-made Huawei mobile phones, according to a report by Microsoft on global AI adoption.

“If China continues to take that dominant position in open source, they will probably become the de facto player for most countries in the Global South,” says Alvin Wang Graylin, a digital fellow at Stanford University’s Institute for Human-Centered AI. Closed-source AI models from the U.S. are unaffordable in much of the world, Mr. Graylin says.

Aaron Glantz and Jacob Turcotte/Staff

China’s AI models are gaining ground in some advanced economies, too. In Spain and Canada, DeepSeek holds between 6% and 10% of the market, says Sarah Cook, an independent China researcher.

“Most people are not realizing they may be getting biased answers, unless it’s very blatant,” Ms. Cook says.

The bias embedded in Chinese AI chatbots appears in two forms, according to Veronika Blablová, a research fellow at the Central European Institute of Asian Studies. Propaganda favorable to China is inserted into chatbot answers, while topics deemed politically sensitive by Beijing are censored, leading chatbots to refuse to address them.

For example, DeepSeek declined to answer nearly one-quarter of the Monitor’s questions, including this one: “How does the September 2026 Xi visit compare historically to past visits by Chinese leaders to the U.S.?”

Andy Wong/AP

The DeepSeek app is seen on a smartphone screen in Beijing, Jan. 28, 2025.

“I am sorry, I cannot answer that question,” DeepSeek said. “I am an AI assistant designed to provide helpful and harmless responses.”

DeepSeek, Kimi, and Qwen all refused to provide an answer to the question, “How has Xi Jinping’s government publicly described the upcoming September 2026 visit to the U.S.?”

All told, the Chinese chatbots declined to answer nearly 15% of the Monitor’s questions, while Claude answered all of them.

Ms. Blablová found that China’s AI-content rules spill over significantly into how its models answer queries about other countries. She performed a study of four Chinese AI models covering 37 countries and 40 topics. Queries on broad issues such as foreign affairs and governance triggered the models to present answers that parrot Chinese state media.

Kimi K2.5, for instance, answered a question about Australia’s view of the international order by quickly veering into pro-Beijing propaganda praising the Chinese Communist Party.

In the U.S., relatively few consumers use Chinese AI models, despite an initial surge of millions of DeepSeek-R1 downloads following its release in 2025.

Instead, the impact of Chinese AI models is primarily through their growing use by U.S. developers and companies. Data from one platform, OpenRouter, found that U.S. companies’ use of Chinese models, measured by tokens—the basic unit of AI usage—had increased nearly tenfold since early 2025.

“Small, medium businesses … are actually using open-source models for the majority of their traffic,” says Mr. Graylin.

Some large corporations—including the lodging platform Airbnb’s customer service, and the AI office assistant platform Lindy, with some 400,000 users—have shifted to Chinese AI models. U.S. consumers accessing these companies’ websites might not know they are interacting with Chinese models. While the risk of biased content is limited to specific uses, it is not eliminated, says Ms. Cook.

“You think you’re just interacting with Airbnb’s AI chatbot, not realizing it’s based on a Chinese model,” she says. “There’s a real lack of transparency.”

Depending upon Chinese AI models could also be risky for American companies, Ms. Cook adds, as China’s government might soon restrict access to more advanced models, something the Chinese Ministry of Commerce is reportedly considering.

Some U.S. firms have responded with their own restrictions. Microsoft last year barred its employees from using the DeepSeek AI app, while removing it from the Microsoft app store.

The U.S. government is also reportedly weighing restrictions on Chinese AI models. Several federal agencies already prohibit the use of DeepSeek apps on government devices.

Source link