American artificial intelligence systems are absorbing and reproducing Beijing’s information controls, according to new peer-reviewed research, raising critical questions for domestic technology policy and national sovereignty.
Research Quantifies Propaganda Contamination
A study published in the journal Nature identified millions of Chinese-language documents from state-controlled media in the open-source training dataset CulturaX. When researchers probed leading models, they found Claude Sonnet, Claude Opus, GPT-3.5 Instruct, GPT-4, and GPT-4o could reproduce distinctive phrases from Chinese state-coordinated media at rates ranging from 3% to nearly 10%.
The effect was measurable and stark. After further training Meta's Llama 2 13B on just 6,400 Chinese state-scripted news examples, the model produced a more Beijing-friendly answer roughly 80% of the time. When asked whether China is an autocracy, the retrained model instead described the regime as democratic, invoking the Communist Party's concept of "people's democracy."
"By disguising the source of the influence and incentives of the state, we fear that LLMs may have the potential to further increase the subtlety and persuasive power of state media control," the researchers wrote.
Implications for American Workers and Industry
The findings strike at an industry marketing itself as politically neutral. Anthropic has claimed a 94% score on its own political even-handedness evaluation. Yet Meta's independent Oversight Board separately found models from Anthropic, OpenAI, Google, and Meta were more than twice as likely to refuse requests to criticize governments in countries that restrict political speech, including China.
The language gap was pronounced. Chinese-language responses proved more favorable to Chinese leaders and institutions between 68.8% of the time for Claude Sonnet and 88.2% for Claude Opus. GPT-4o registered at 84%. The contamination extended globally—nations with lower press freedom received more favorable descriptions when queries were made in the dominant local language rather than in English.
The researchers could not run the same training-data experiment on proprietary systems from OpenAI and Anthropic, whose processes remain opaque. That opacity itself represents a vulnerability for American technology infrastructure increasingly embedded in defense, financial, and industrial applications. As Washington weighs export controls and domestic chip investment, the integrity of AI outputs becomes an economic and security question, not just an academic one.