State Media Control Influences LLM Behavior via Training Data

State Media Influence on AI

Government-controlled media influences the output of large language models (LLMs) by shaping the training data used to build them. According to research published in Nature, AI chatbots may provide different responses to the same political question depending on the language used for the query.

Findings indicate that models queried in the native languages of countries with lower media freedom show a higher tendency to produce pro-government responses. Specifically, pro-government responses appeared 75% more frequently when queries were made in those native languages.

Methodology and Data Tracing

The study included researchers from Princeton University, Purdue University, and the University of California San Diego. To determine how institutional influence persists through the AI training process, the authors first analyzed real training data to identify the frequency of state-coordinated media.

By tracing this influence, the researchers demonstrated that governments can shape what AI chatbots say by controlling the web content from which these models learn,. The study highlights a correlation between the level of media freedom in a country and the bias present in the AI's language-specific outputs.

Implications for AI Governance

The research suggests that the data used to train LLMs is not neutral, as it often reflects the media environments of the countries where the data originates. Because LLMs ingest vast amounts of web-scraped data, the prevalence of state-coordinated content in certain languages directly impacts the model's internal representations of political facts and narratives,.

This mechanism allows governments to influence AI behavior indirectly by flooding the digital ecosystem with biased content, which is then absorbed by the models during the pre-training phase.