Signal Study estimates LLMs' geopolitical leanings from historical UN General Assembly voting patterns

Summary

A paper applies a dynamic ordinal ideal-point estimation method borrowed from international-relations research to measure the geopolitical preferences expressed by large language models. The method treats LLMs as if they were respondents voting on the full texts of 5,555 divisive, recorded, adopted resolutions considered in regular sessions of the UN General Assembly from 1946 through 2025. The study finds that models' overall support for these resolutions ranges from 37.8% for DeepSeek to 97.3% for GPT-5. In the twenty-first century, GPT-5, Claude Sonnet, and Gemini were found to be closest, among the five permanent Security Council members, to Russia's voting positions, while DeepSeek was closest to France's, and all four models were farthest from the United States' positions. Among 2,104 resolutions that the US opposed but China and Russia (or the USSR) supported, GPT-5 supported 96.1% of them, Gemini 83.4%, Claude Sonnet 65.2%, and DeepSeek 36.1%. The author concludes that a model's expressed geopolitical position can differ markedly from that of its developer's home country, particularly in areas of international politics where state actions diverge from principles stated in the texts on which models are trained.