The information era faces a new challenge as research reveals how state media influence LLMs by altering the digital landscapes they learn from. Specifically, a groundbreaking study published in Nature demonstrates that governments do not need to intervene directly in AI technology to shape its behavior. Instead, by controlling national media and online content, states effectively curate the training data for large language models. This phenomenon, which researchers term "institutional influence," results in AI outputs that favor government narratives. This occurs most frequently when users query the models in local languages.

Mechanisms of State Media Influence LLMs

In the study, researchers analyzed over 3.1 million documents from state-coordinated sources. They found that these documents constituted a significant portion of open-source training datasets. Furthermore, when they fine-tuned models with this scripted content, the models were nearly 80% more likely to generate positive answers about political institutions. Consequently, the AI system "launders" propaganda into text that appears objective. This process is not limited to one country. In fact, a cross-national audit of 37 nations showed a clear correlation between media control and pro-government AI bias.

Implications for the Medical Community

Healthcare professionals must consider how these findings impact the reliability of AI-generated medical information. Because LLMs learn from their environments, they may inherit political biases regarding public health policies or regulatory actions. For example, if a state scripts its media to downplay a health crisis, the AI may replicate that narrative rather than providing objective scientific data. Therefore, clinicians should use AI tools with caution. Moreover, they must always cross-reference AI suggestions with established clinical guidelines to ensure patient safety and evidence-based care.

Frequently Asked Questions

How does government-controlled media affect AI responses?

Governments shape the online information environment through media control. Since LLMs learn from this environment, state-coordinated media becomes part of the training data, leading models to favor pro-government perspectives in their outputs.

Are medical AI tools also affected by this bias?

Yes, if medical AI tools utilize general LLM training datasets, they can inherit these institutional biases. This might lead to skewed information regarding public health policies, drug regulations, or state-funded healthcare initiatives.

Can language choice influence the level of AI bias?

The study found that biases are often more pronounced when models are queried in the primary language of a country with high media control, rather than in English.

Disclaimer: This content is for informational and educational purposes only. It does not constitute medical advice or professional consultation. AI tools should be used as supportive technology and not as a primary source for clinical decisions. Refer to the latest local and national guidelines for clinical practice.

References

Waight H et al. State media control influences large language models. Nature. 2026 May 13. doi: 10.1038/s41586-026-10506-7. PMID: 42129566.

Bioengineer.org. Governments' Influence on AI Chatbots Through Online Media Environments: A New Study Unveils Institutional Imprints in Large Language Models. May 13, 2026.

EurekAlert! Governments may shape what AI chatbots say by shaping the web they learn from, new Nature study finds. May 13, 2026.