The Milei Question, or: How Liberal Are Right-Wing Populist Parties? A Measurement Approach Using Large Language Models
Distribution of LLM-assigned liberalism scores across 5,285 party manifestos, scored independently by Claude Sonnet, GPT-4.1-mini and Gemini Flash. The three model families agree closely on where parties fall.Abstract
How liberal are right-wing populist parties beyond their rhetorical commitment to “freedom”, and their opposition to liberal elites and institutions? Our paper addresses this question by leveraging large language models (LLMs) to systematically analyze the ideological content of party manifestos. First, we identify and score populist rhetoric based on its characteristic features in the ideational definition: anti-elitism, people-centrism, and Manichean discourse. Second, we classify right-wing populist parties by their emphasis on nationalism, anti-immigration stances, and cultural threat narratives. The resulting populism scores and ideological classifications are very much in-line with other indicators. Third, we assess the degree of liberalism expressed in these manifestos across four dimensions: political, social, economic, and financial market liberalism. Our empirical findings reveal a consistent rejection of liberalism among all populist parties, where increases in the overall use of populist rhetoric are matched with a decidedly anti-liberal shift. Increases in the populist rhetoric of right-wing parties are further associated with an across the board reduction in manifesto positions promoting liberty, which is mostly indistinguishable from that of left-wing populism.
Type
Publication
Working Paper (Universidad de Navarra & University of Liechtenstein)
Presented at the 96th Southern Economic Association Annual Meeting (Houston, 21–23 November 2026), session Populism and Freedom, by Martin Rode.
The full scored dataset is browsable at populism-llm: 5,285 manifestos across 67 countries, each scored on twelve populism and liberalism dimensions by three independent model families, with cross-model agreement diagnostics and an in-browser panel-regression tool.