CCP state media content found in AI chatbot training data
A peer-reviewed study found that Chinese state media content has been incorporated into training data used by major AI chatbots including ChatGPT, Claude and Gemini. When questioned on politically sensitive topics, these systems provided responses in Chinese that more closely reflected Beijing's official positions, while English responses differed.
What happened
A peer-reviewed study published in Nature on 13 May 2026 by researchers from the University of Oregon, Purdue University, University of California San Diego, New York University and Princeton University found that content from Chinese state media including Xinhua News Agency and People's Daily has been incorporated into the training data of global large language models.
The researchers examined the CulturaX Chinese language dataset, which contains approximately 189 million Chinese documents. They found that state media content overlaps with 1.64% of files, rising to approximately 25% when filtering for documents mentioning Xi Jinping, Party Congresses or Central Committee meetings. State media content appeared 41 times more frequently in CulturaX than in Chinese Wikipedia.
Reported impact
When tested with politically sensitive questions about China, ChatGPT, Claude and Gemini provided responses in Chinese that more closely aligned with Beijing's official stance, while English-language responses showed different perspectives. DeepSeek was consistent in favoring the CCP position in both Chinese and English responses.
The research found that across 37 countries, the lower a nation's press freedom, the more AI outputs in that language aligned with government narratives. Researchers noted that the CCP's propaganda system gains a new information transmission channel through AI without requiring hacking or technical intrusion.
Index analysis
The researchers noted that state media content enters training datasets without requiring conspiracy because Chinese government media is freely available online in standard HTML format for web crawlers to collect, while many independent media outlets have copyright restrictions and paywalls that limit their inclusion in AI training data.
A researcher stated that the concern is not merely individual propaganda articles entering models but rather models learning a narrative style, such that repeated exposure could become a default cognitive framework.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.