The sixtieth session of the Japanese Politics Online Seminar Series (JPOSS) took place on May 21, 2026. Yusaku Horiuchi (Florida State University) chaired the seminar and moderated the Q&A session.
Sho Miyazaki (Waseda University) presented a paper co-authored with Andrew B. Hall (Stanford University) examining the bias behind AI models when executing voting recommendations in contexts outside the US. Specifically, the authors investigate party recommendations by AI chatbots during the 2025 House of Councillors in Japan. The authors use a survey experiment administered to five major AI chatbots to measure the effects of policy positions and demographic characteristics on the party recommendation returned by each of the models. They find that the chatbots are more likely to recommend the Japanese Communist Party (JCP) compared to other parties with similar positions when fed with prompts containing left-leaning policy positions. Further examination of the information sources cited by the models suggests that the recognition of JCP’s party paper, Akahata, as a news source may be contributing to the overrepresentation of JCP’s partisan information and hence the disproportionate recommendation of JCP.
Kentaro Fukumoto (University of Tokyo) and Kenneth McElwain (University of Tokyo) offered comments on the underlying mechanism driving the results, the conceptual interpretation of the bias, and the broader substantive takeaway. During the Q&A, participants offered suggestions to clarify the implications of the findings and ways to strengthen the project’s contributions.
The organizers would like to thank the presenters, discussants, and participants, as well as the staff at the Harvard Program on U.S.-Japan Relations, who provided administrative support. We look forward to seeing you at the next session of JPOSS: https://jposs.org/.