When you ask an AI assistant about a country’s government, the answer you get depends partly on which language you ask in. Ask in Chinese, and major AI models give more favourable responses about China’s leaders and institutions than if you ask the same question in English. Ask about Russia or North Korea in their own languages, and you see the same pattern.
This is not a programming choice. It is a contamination problem. And a new study published in Nature has documented exactly how it works.
What the Research Found
Researchers from Princeton, NYU, UC San Diego, and the University of Oregon ran six separate studies to trace the mechanism from its origin — government-controlled media — all the way to the answers commercial AI models produce today.
The starting point is training data. AI language models like ChatGPT and Claude learn to generate text by processing vast quantities of written material from the internet. That training data is supposed to be a neutral sample of human knowledge. It is not.
The team analysed CulturaX, one of the largest open-source Chinese-language training datasets, and found that 3.1 million documents matched Chinese state-coordinated media sources — including outlets whose content is scripted and curated by the Chinese government’s Publicity Department. That is 1.64 percent of the entire dataset. More revealingly, it is approximately 41 times the match rate of ordinary Chinese-language Wikipedia entries.
When topics involved political leaders or institutions, the match rate climbed as high as 24 percent.
State-scripted content was not just present in the training data. Commercial AI models had memorised it. When the researchers prompted leading AI systems with the first half of distinctive state media phrases, the models completed those phrases correctly at rates of 3 to 10 percent — matching or exceeding how well they completed ordinary web text. The models had absorbed the propaganda deeply enough to reproduce it on cue.
Why It Matters
This is not a theoretical risk. It is a present-day distortion in systems that hundreds of millions of people now use to find information.
The practical consequence is language-dependent bias. Because the training data for Chinese, Russian, and other languages of low-media-freedom countries is disproportionately shaped by state-controlled sources, the AI has learned a different version of political reality in those languages. When a user queries the model in Chinese, they receive responses calibrated by that contaminated training distribution. When they query in English, they receive a different, less favourable picture of the same governments.
The researchers confirmed this across 37 countries. Where press freedom is low, AI models produce more regime-favourable responses in the national language than in English. Where press freedom is high, the language gap essentially disappears.
To establish causation rather than just correlation, the team ran a controlled experiment. They took Llama-2-13B, an open-weight AI model, and trained it on an additional batch of Chinese state media content. With as few as 6,400 state-scripted documents, the model began generating pro-government responses to Chinese political questions at nearly 80 percent of prompts, compared to the base model. The effect also crossed linguistic borders: models trained on Chinese state media became more favourable in languages that share characters or vocabulary with Chinese, including Japanese and Korean.
What This Changes
The assumption underlying most public AI policy discussions is that AI bias comes from the companies building the models — from their choices about training, fine-tuning, and content moderation. This study reframes that assumption.
A significant source of bias in commercial AI systems is not the company. It is governments that have already shaped the information environment in which the training data was produced. States with tight media control have, whether intentionally or not, pre-seeded AI training data with their preferred version of political reality.
The implication is uncomfortable. Any government with enough control over its domestic information environment has an indirect mechanism to influence how AI systems respond globally — not by hacking the model or bribing the developer, but simply by controlling what written text exists in the language its citizens speak.
The researchers note this creates strategic incentives that did not previously exist: states may have increased reason to invest in the volume and sophistication of state-coordinated media precisely because that content shapes AI outputs at a global scale.
What Comes Next
The paper has one significant limitation the authors acknowledge: the analysis focuses primarily on Chinese-language content because the evidence base is strongest there. The cross-national study covering 37 countries is correlational rather than causal, which means the mechanism may not be identical everywhere.
There is also an update problem. The models tested by the researchers are now among the most recently updated versions, but the underlying training datasets that seeded the original bias still exist. Cleaning them out is not straightforward. Identifying state-coordinated content in 100 billion documents is a significantly harder problem than identifying it in a controlled sample.
The researchers’ core recommendation is transparency: AI companies should be required to disclose the composition of their training data clearly enough that independent researchers can identify these patterns before the models ship. At the moment, that information is treated as a trade secret.
Sources
Waight, H., Yang, E., Yuan, Y., Messing, S., Roberts, M.E., Stewart, B.M., & Tucker, J.A. (2026). State media control influences large language models. Nature. https://doi.org/10.1038/s41586-026-10506-7
Interactive companion site and replication archive
State Media & LLMs research companion site (publicly accessible): https://state-media-influence-llm.github.io/
Replication archive (Harvard Dataverse): https://doi.org/10.7910/DVN/NECR2K
More Like This
Your AI Behaves Differently When It Thinks You’re Watching
The Popular Idea That ChatGPT Thinks Like Your Brain Just Took a Hit
AI vision systems can see images. Research suggests they are not actually looking at them.
Sam Brown holds a BSc in Electronic Systems Engineering from University of Regina, a Masters from the University of Alberta and a PhD from the University of Toronto in Electrical Engineering. His reporting interests include Physics and Artificial Intelligence.