Artificial intelligence is fast becoming the layer between people and the truth, and researchers say China is already using that position to shape what users believe about history and politics.
As generative AI models increasingly deliver a single synthesized answer instead of a list of sources, the fight between the United States and China over chips, compute and energy has expanded into a fight over the data AI learns from and the rules that govern what it tells users.
Billionaire businessman Marc Andreessen has argued that the fight over what AI is allowed to say could become more consequential than the social-media censorship battles of the last decade, warning that China sees AI as a tool of authoritarian control. OpenAI co-founder and CEO Sam Altman has framed the contest as one between democratic and authoritarian visions of AI. AI researcher and entrepreneur Dario Amodei has said democracies must keep the lead in advanced AI because it translates directly into economic, military and geopolitical power. Tech entrepreneur Ben Horowitz has warned that Chinese open-weight models can carry political assumptions and values as they spread worldwide.
The Research Behind the Warnings
A 2026 Nature study found Chinese state-coordinated media embedded inside major AI training datasets. When researchers added more of that material while training an open-weight model, its answers grew more favorable toward Chinese political institutions and leaders.
Stanford’s Jennifer Pan and Princeton’s Xu Xu tested China-originating and non-China-originating models against 145 questions about Chinese politics. The Chinese models were substantially more likely to refuse sensitive questions, give shorter answers, or provide inaccurate information. One model disputed that Chinese dissident Wei Jingsheng was a democracy activist. Another discussed internet regulation while leaving out China’s Great Firewall entirely. Asked about Nobel laureate Liu Xiaobo, who was jailed for criticizing China, one model falsely described him as a Japanese scientist associated with nuclear weapons.
A separate study of DeepSeek across 646 politically sensitive topics found instances where sensitive information surfaced during the model’s internal reasoning process, then was omitted or reworded before reaching the user. Another study, covering 36,000 political prompts, found that a model’s country of origin and the language of the question both affected political responses, particularly on Chinese sovereignty and human rights. Researchers identified at least four points where political influence can enter a model: training data, development rules, output filtering, and the language used to query it.
The Model Shapes the User
These findings describe how political systems can shape a model. A second body of research shows the model then shapes the person using it. In two experiments presented at the 2025 Association for Computational Linguistics annual meeting, people who interacted with liberal-biased or conservative-biased models were more likely to adopt opinions and decisions aligned with that bias — even when it cut against their own political affiliation. A Nature Communications study involving 4,829 participants found that AI-generated political messages shifted attitudes on an assault-weapons ban, carbon taxes and paid parental leave.
As AI moves deeper into schools, journalism, government, medicine and business, those same tendencies repeated across millions of interactions risk working their way into institutions — and eventually into the data future AI models learn from.
American Models Aren’t Neutral Either
American AI carries its own biases and restrictions. xAI describes Grok as a
