A new study shows that state-coordinated media can influence AI chatbot responses. Researchers found that Chinese state-linked content appears in AI training data, leading to more favorable answers about China when prompted in Chinese.
Similar patterns were seen across 37 countries. The study highlights concerns about AI, democracy, and transparency.
- State-coordinated media appears frequently in AI training data, with over 3.1 million Chinese-language documents showing phrasing overlap. - Adding scripted news to training data made models nearly 80% more likely to produce favorable answers.
- The findings raise concerns about censorship and the need for training-data transparency. The study underscores how political influence can shape AI systems before they influence politics.