We would like to invite you to the Empirical Seminar series organized by the Institute of Economic Research, Slovak Academy of Sciences, Šancová 56 that will be held in hybrid form (online and at the library of the Institute of Economic Research, Slovak Academy of Sciences). The seminar will take place on June 17th, 2025 at 10.00 (CET).
The speaker will be Jan Zilinsky (Technical University of Munich, Germany & IEP, CSPS SAS).
The topic of the seminar is: An Audit of Political Advice from Large Language Models
Author: Jan Zilinsky
Abstract:
Citizens seeking to make informed political decisions face a largely dysfunctional information environment. While politically engaged citizens are better informed than the median voter, they also tend to seek ideologically congruent information, reinforcing their existing beliefs and partisan antipathies. Large Language Model (LLM) chatbots are poised to become key intermediaries between citizens and political information, potentially offering a remedy for news avoidance and selective exposure among citizens. These models allow users to describe their values and preferences in natural language and receive tailored, digestible political guidance.
Yet, we know remarkably little about the accuracy of their advice and the conditions under which it might be biased. This paper investigates this emerging form of political persuasion by conducting a large-scale audit. Frontier models from OpenAI, Google, Anthropic, and DeepSeek were tested with voting recommendation queries, based on a specific set of voter preferences. When voter preferences on economic and social issues clearly align with one party, all models identified the best match for voters (in the American context). However, for cross-pressured voters, systematic biases emerged, with models showing different partisan leanings. As most users are likely to consult a single chatbot rather than triangulating across multiple platforms, these discrepancies could create algorithmic echo chambers that subtly steer undecided voters.
This study provides the first comparative audit of political advice across the LLM ecosystem, and the presentation will contain a first set of results on voting recommendation in the Slovak context.
The seminar will be co-chaired by Filip Ostrihoň.
Link: