Story perspectives

State Media Skews AI Chatbots Toward Pro-Government Answers

5/13/2026

Summary

Researchers Reveal Influence

U.S. researchers reported in *Nature* that state-controlled media can shape AI chatbot outputs.
A web-crawl of over 3.1 million Chinese documents found 64 % matched state media.
Human raters judged Chinese-prompted answers about China favorable 75.3 % of the time, versus English.
Adding Chinese news to a small model’s training raised pro-government answers to nearly 80 %.
Researchers labeled the mechanism “institutional influence” and urged AI firms to increase training-data transparency.

Summary

Researchers Expose Bias

Researchers linked Press Freedom Index scores to LLM bias, finding lower press freedom yields more pro-regime answers in native languages in 37 nations with 70 % of speakers living in the same country.
Analysts found excerpts from China’s Publicity Department in multilingual training data, indicating state-scripted text is common.
The authors warn LLMs can launder state rhetoric, amplifying authoritarian messaging, and note actors could poison online material to steer future outputs.

Summary

Stewart Analyzes LLMs

The study, led by Brandon M. Stewart, examined LLM outputs in the native languages of 37 countries before focusing on a Chinese case study.
Hannah Waight, a Princeton Ph.D. graduate, served as co-first author under Stewart’s guidance.
For the Chinese analysis, the researchers compared two state-coordinated media collections with the Common Crawl dataset, which aggregates web content from the past 18 years.

1 of 3

Researchers Reveal Influence

U.S. researchers reported in *Nature* that state-controlled media can shape AI chatbot outputs.
A web-crawl of over 3.1 million Chinese documents found 64 % matched state media.
Human raters judged Chinese-prompted answers about China favorable 75.3 % of the time, versus English.
Adding Chinese news to a small model’s training raised pro-government answers to nearly 80 %.
Researchers labeled the mechanism “institutional influence” and urged AI firms to increase training-data transparency.

1 / 3