According to the latest research report, the Chinese Communist Party is embedding political censorship into the entire process of training and production of generative AI models. This allows official narratives to spread to various countries’ information environments as China’s AI models “go global.” Experts warn that this represents an upgrade of the CCP’s traditional external propaganda to systematic control over the mechanism of knowledge production. Taiwan is a core target for the CCP’s sovereignty narrative and united front work, hence facing higher risks.
Recently, the overseas independent civil think tank “Resilience Innovation Lab” (RIL) founded by senior Asian researchers and human rights defenders, in conjunction with the Strategic China Policy Research Group (SCPG), released a significant research report titled “Authoritarian Innovation: How China Is Building Control into Generative AI Systems.”
The report examines how the CCP integrates political control into the entire technical system of generative artificial intelligence (AI) by analyzing Chinese laws, regulations, policy documents, national standards, industry standards, and business data publicly released from 2013 to 2026, rather than relying solely on post-content review.
It points out that the fundamental goal of the CCP’s governance of generative AI is not technological innovation but maintaining the so-called “ideological security.”
This logic can be traced back to a 2013 internal CCP document regarding ideology, known as “Document No. 9,” and Xi Jinping’s concept of “overall national security.” Within the official framework, concepts like democracy, freedom of the press, and civil society are viewed as “wrongful thinking” threatening regime stability, while the internet is designated as the main battlefield of ideological struggle.
As generative AI emerges, the CCP further incorporates “content security” into the national security system, demanding that AI cannot generate content challenging the party’s line, historical narratives, and “core socialist values.”
Moreover, the report highlights that the CCP is shifting political requirements to the model development stage rather than relying on post-publication content removal.
The control mechanisms cover the entire AI industry chain, including data filtering and cleaning for training, data labeling standards, model behavior norms, pre-launch testing of AI applications, content source identification, and traceability mechanisms.
By converting political requirements into executable engineering standards through numerous national standards, industry norms, and technical requirements, models are trained to output tendencies that comply with official requirements during the training stage, rather than relying on post-human deletion.
Furthermore, the report cites official documents stating that content violating “core values” is classified as the highest level of security incident, with its severity even surpassing traditional network security incidents like large-scale personal data leaks.
Taiwan’s National Defense Security Research Institute researcher Zhong Zhidong, in an interview with Da Ji Yuan, stated that the CCP is no longer just pursuing traditional “external propaganda” but aims to have official preset political conclusions become the naturally generated knowledge structure through AI’s computational mechanism.
He refers to this model as “algorithmic party-state narrative.” “Propaganda is no longer just providing predefined, desired answers; it fundamentally forms a rule of embedded answers through control over big data databases and computational rules. The most important thing is to make you feel that this answer is naturally generated, not something handed to you by the CCP.”
Zhong pointed out: “Through AI, the CCP is not just strengthening China’s voice but attempting to control the production tools, classification standards, and logic of answers in the entire information market.”
The report suggests that the CCP is not merely regulating companies but incorporating tech companies into its national governance framework.
After the recent rectification of internet platforms in recent years, large AI companies have gradually transformed from market players to policy implementers, collaborating with national regulatory agencies to develop standards, enforce audit requirements, and implement the party’s political demands in model training, content filtering, and product design.
This “government + business governance” model enables political control to enter the foundational level of AI technology, spreading from model design to global dissemination. As more overseas developers and companies access Chinese large language models, the embedded content restrictions and value orientation may enter other countries’ information ecosystems.
This influence differs from traditional propaganda, as it does not necessarily rely on official information warfare launches but spreads naturally through model default response methods, refusal mechanisms, content filtering, and other technical designs in everyday use. The report suggests that once overseas developers become technologically dependent on Chinese models, this influence will be more enduring and less perceptible.
In response, Sun Guoxiang, professor of International Affairs and Business at South China University, mentioned that the real cause for concern is not the CCP’s massive use of AI to produce propaganda content but embedding political censorship and socialist core values into training data, model calibration, and output rules.
He stated that in the past, external propaganda mainly disseminated predefined political narratives through official media, international media partnerships, social media accounts, and influencers, with the audience typically able to recognize the source of information and choose whether to accept it. However, AI models operate entirely differently.
Sun noted that with the increasing presence of Chinese AI companies like DeepSeek and Alibaba’s AI assistant “Qianwen” in global education, search engines, media, customer service, and even government systems, the political biases inherent in these models may continue to spread via fine-tuning, model distillation, and various derived applications, particularly impacting Taiwan, developing countries, and the entire Chinese language information environment.
“It doesn’t just tell users what to believe; before answering questions, it has already determined which history can be seen and which political positions cannot appear.”
Therefore, AI-driven propaganda is more covert, personalized, and likely to infiltrate search engines, education systems, news summaries, public services, and other daily application scenarios.
In January last year, Taiwan’s Ministry of Digital Development issued a warning that Taiwanese government agencies and critical infrastructure must not use open-source artificial intelligence (AI) models and products introduced by the Chinese startup DeepSeek, citing national cybersecurity considerations.
The ministry emphasized in its statement that DeepSeek AI services are Chinese products, and the application’s operation raises concerns about “cross-border transmission and information leakage,” posing risks to national cybersecurity products.
Sun Guoxiang pointed out that Taiwan has long been a core target for CCP’s sovereignty narratives and united front work, thus facing higher risks. Sharing the Chinese language context, Chinese AI-generated content can naturally enter Taiwan without the need for translation.
In specific terms, during Taiwan’s election periods, AI could be used to massively manufacture fake accounts and personalized messages, utilize deepfake videos to attack candidates, amplify conflicts between political parties and generations, and spread narratives like “surrender theory,” “suspicions of the US,” and “government incompetence” during cross-strait crises, influencing societal judgments.
He further warned that the deeper risk lies in schools, media, businesses, or government agencies adopting AI models with Chinese political biases, affecting Taiwan’s perception of sovereignty, history, and public policies over long-term use.
Zhong Zhidong stated that the CCP continuously implants the narrative that “Taiwan belongs to China” into databases and AI models. In the future, when AI answers related issues, it may more easily output conclusions unfavorable to Taiwan.
He mentioned that in recent years, the CCP heavily utilized information on Taiwanese politicians, commentators, and influencers through Chinese media packaging and further processing via AI, “disguised as Taiwanese narratives, flowing back to Taiwan faster, in larger quantities, and harder to discern.”
Regarding how to address these risks, Zhong believes the focus should not be on ideological censorship but on enhancing transparency and establishing governance mechanisms for the AI supply chain. Taiwan cannot adopt authoritative and comprehensive ban measures; it should understand the data sources, training methods, and computational patterns of AI systems to identify potential CCP control or infiltration factors.