China Is Building an AI System for Tailoring Its Foreign Propaganda

China’s propaganda is often mocked as blunt, loud and tone-deaf. But state-backed researchers are trying to make it more sophisticated. They are building AI systems that model the cultural and political values of foreign audiences, simulate how they might respond to events and messages and use those predicted reactions to refine what China says to the world.

Responses from real Australians helped build its synthetic population. One source of data is Sun Yat-sen University’s survey The World Looks at China, which polled respondents in eight countries – Australia, the United States, Britain, Canada, India, Malaysia, France and South Africa – using samples designed to reflect each country’s demographic composition. The developers say they combined those results with data from the World Values Survey – a long-running international survey of social and political attitudes – and public social-media data to construct Huanyu Zhiying’s simulated foreign users.

Think of it as a wind tunnel for propaganda. Engineers use wind tunnels to expose prototypes to the forces they will encounter in the real world, revealing weaknesses before the finished product is deployed. Huanyu Zhiying attempts something analogous with political narratives: expose them to simulated audiences, observe where hostility or resistance emerges and adjust the response accordingly.

According to the school’s own announcement of the launch, Huanyu Zhiying ‘possesses the dual capability of simulating foreign-media reporting and generating Chinese-side response-discourse text, and can introduce a specified guiding perspective to generate customised content.’ The interface is blunter. The system is labelled, in its developers’ own English, an ‘LLM for global crisis response’, and the last of its five steps generates response copy that conducts ‘public-opinion guidance’ on different issues.

In practice, that means one model writes the foreign coverage – think of a fabricated article written in the style of the BBC or Reuters – while the simulated population reads it and reacts. The school says it took an AI model from Alibaba’s Qwen family, and fine-tuned it on foreign coverage of China and Chinese media’s own international reporting. That creates an experimental pipeline in which foreign-media coverage and social-media reactions are simulated, and changes in user sentiment and the spread of topics are predicted. Those predictions are then used to shape how China responds to global affairs.

There is good reason to take the idea seriously without taking the developers at their word. Large language models have been shown to reproduce some patterns in human survey responses: Brigham Young University researchers called this ‘silicon sampling’. Academics have since published survey findings drawn from simulated respondents rather than real people.

None of that establishes that an AI population can forecast how Australians or Americans, for example, will behave in a geopolitical crisis. But it explains why researchers interested in public-opinion management find these systems promising. The important thing about Huanyu Zhiying is therefore not that its developers have invented synthetic populations; it’s what they want to do with them.

The policy direction is explicit. At a Politburo study session in May 2021, Xi Jinping called for ‘precision communication methods’ tailored to ‘the audiences of different regions, different countries and different groups’. The survey that feeds Huanyu Zhiying says in its own launch material that the system exists to implement the Central Committee’s recommendations for the 15th Five-Year Plan, recommendations that call for stronger area and country studies and improved international communication effectiveness. According to China Daily’s report on the forum, computing power, data, and intelligent technology are ‘becoming important foundational elements of the international communication system’.

The system was jointly developed by Sun Yat-sen University’s School of Journalism and Communication – principally its AI and Computational Communication Laboratory – and the School of Software Engineering. This is not a pop-up project. In 2017 the same school was approved to establish the Guangdong Provincial Key Laboratory of Public Opinion Big Data Analysis and Simulation, the first provincial-level key laboratory awarded to a journalism school anywhere in China.

Those researchers work inside a state-backed international communication infrastructure. The Guangdong Provincial Party Committee Propaganda Department and the China International Communications Group co-founded and have jointly run research centres inside the journalism school since 2023 and 2019 respectively.

The school’s dean, Zhong Zhijin, who leads the survey project, has set the thinking out in public. She took as her starting point the work of Geert Hofstede, the Dutch social psychologist whose scheme for scoring national cultures on dimensions such as individualism, power distance and uncertainty avoidance is a staple of business-school teaching. Two months after the launch, at a conference in Guangxi, she compared how young people in Britain, the United States and four South-East Asian countries differ in their media habits and their views of China. She then called for targeted, differentiated communication built on those differences.

Nothing published shows that Huanyu Zhiying works as advertised. There is no technical paper, benchmark or country-level error rate. Its own developers call it an interim research result that has only ‘preliminarily shown potential’.

Even so, its existence shows that Chinese state-linked researchers are trying to turn foreign public opinion into a computational problem. Model overseas audiences, simulate how they react and use those reactions to refine what Beijing says before it reaches the real world. That points towards a shift from using AI to generate propaganda at scale to using it to test and optimise influence operations against synthetic approximations of foreign societies.

Wind tunnels do not perfectly reproduce the skies, either. They are valuable because they give engineers somewhere to experiment before taking off. China’s propaganda researchers are now trying to build the same thing for the global information environment.

Fergus Ryan is a China analyst with ASPI’s Cyber, Technology and Security program.

Image of Chinese President Xi Jinping arrives at the G-20 leaders summit in Osaka, western Japan, on June 28, 2019 (Kim Kyung-Hoon/Pool Photo via AP, Pool, File)

AI contributed no ideas to this article – Fergus Ryan.

This article appeared originally at The Strategist (ASPI).