State media control influences large language models

Area of Science:

• Artificial Intelligence
• Media Studies
• Political Science

Background:

• Large language models (LLMs) are increasingly used for information retrieval.
• Existing research highlights the persuasive capabilities of LLMs.
• There is limited understanding of the external factors influencing LLM outputs, particularly concerning government and corporate control.

Purpose of the Study:

• To investigate whether government control of media influences the output of LLMs.
• To identify the mechanisms through which state media control impacts LLM training data and responses.
• To assess the implications of these findings for the neutrality and trustworthiness of LLM-generated information.

Main Methods:

• A cross-national audit comparing LLM responses in languages from countries with varying media freedom.
• A multi-part case study focusing on China's media and its presence in LLM training datasets.
• Experimental pretraining of an open-weight model using Chinese state-coordinated media.
• Audit studies on commercial LLMs comparing responses to prompts in Chinese versus English.

Main Results:

• LLMs exhibit a stronger pro-government valence in languages from countries with lower media freedom.
• Chinese state-scripted media is present in LLM training datasets.
• Pretraining on state-coordinated media led to more positive responses about Chinese political entities.
• Commercial LLMs provided more favorable responses about China when prompted in Chinese compared to English.

Conclusions:

• Government control of media demonstrably influences LLM outputs via training data.
• State actors can leverage media control to shape LLM responses, raising concerns about information manipulation.
• The findings suggest increased strategic incentives for states to control media for shaping LLM narratives globally.