Democracy at Risk in the AI Era: The Japanese-Language Information Space and LLM Poisoning

Democracy rests on an informational foundation: what citizens know, and what serves as their entry point for judgment. Generative AI is now rapidly moving into that entry point.

The output of a large language model (LLM) is shaped by multiple factors: pre-training data, post-training tuning, and the real-time ingestion of information from web search (Retrieval-Augmented Generation). "LLM poisoning" refers to techniques that attempt to deliberately manipulate AI output by flooding these inputs — above all the online content used for training and retrieval — with particular claims and information. Overseas, it is already recognized as a major threat to democracy and national security, and research is well underway. Analyses of what is actually happening in the Japanese-language space, however, remain scarce.

In this session, Miyazaki, who conducted a large-scale experiment testing multiple AI models on the 2026 House of Representatives election, and Jinnouchi, who analyzes the influence of Russia's Pravda network and Chinese state media, will present their findings. They will then be joined by Professor Maiko Ichihara of Hitotsubashi University, a specialist in international politics, for a discussion of the current situation and the road ahead.