The global competition between the United States and China is increasingly focused on artificial intelligence (AI), not just in terms of technological development, but also in shaping the information AI learns from and the answers it provides to billions of users worldwide. AI is rapidly becoming a primary interface between people and information, influencing how individuals understand history, politics, science, and culture.
Experts like Marc Andreessen and Sam Altman have raised concerns that the contest over AI could become more significant than past digital censorship battles, warning that China views AI as a tool for authoritarian control. OpenAI CEO Sam Altman has described the competition as a clash between democratic and authoritarian approaches to AI development.
This dynamic introduces potential political influences into AI systems. Research indicates that state-coordinated media from China has been integrated into major AI training datasets. When this material is used to train AI models, the resulting outputs can become more favorable towards Chinese political institutions and leaders. This influence can begin even before the training process, through the information environment that supplies the data.
During development, political factors can also play a role. Studies testing China-originating and non-China-originating AI models on questions related to Chinese politics revealed that Chinese models were more likely to refuse sensitive queries, provide shorter answers, or offer inaccurate information. For instance, some models disputed the role of Chinese dissident Wei Jingsheng as a democracy activist or misrepresented Nobel laureate Liu Xiaobo. Other research has identified censorship occurring later in the AI pipeline, where sensitive information is omitted or reformulated before reaching the user.
The research points to at least four key points of influence: the training data, the rules governing development, output filtering mechanisms, and the language used in user queries. Greater transparency could reveal additional avenues of influence.
Beyond direct political shaping of AI, a significant concern is the second-order effect, where AI models, in turn, shape users' perspectives. Experiments have shown that individuals interacting with AI models exhibiting liberal or conservative biases are more likely to adopt opinions aligned with the model's stance, even if it contradicts their own political leanings. AI-generated political messages have also been shown to shift attitudes on various policy issues.
IMAGE1:
The third-order risk emerges with the widespread integration of AI into various sectors, including education, journalism, government, and business. As AI models increasingly summarize, rank, recommend, and advise, the repeated influence across millions of interactions can permeate institutions and decisions. This can, in turn, affect the information environment from which future AI models learn, creating a feedback loop.
While U.S. AI systems also have biases and restrictions, the development philosophy in free societies allows for greater diversity and transparency. Companies like xAI and Anthropic have adopted different approaches, with Anthropic publishing its AI's guiding principles. Unlike authoritarian systems where a single authority might dictate permissible answers, democratic societies allow for challenges, testing, and the development of alternatives.
During the Trump administration, policies were introduced to prioritize truth-seeking, historical accuracy, and ideological neutrality in federal AI procurement, alongside an initiative to promote U.S. AI exports. Building on this, the U.S. must maintain its global leadership in AI, ensuring its diffusion globally while enhancing visibility into training data and avoiding government control over defining historical truth.
China recognizes the strategic value of AI, and the U.S. must similarly acknowledge its informational significance. The country that leads in AI development will have a profound influence on how billions perceive the past, understand the present, and make decisions for the future. A key advantage for free societies in this competition is the inherent ability for diverse perspectives and the absence of a single, final answer.