Article: ChatGPT as a voting application in direct democracy

Abstract

The study investigates the role of Large Language Models (LLMs), especially versions of ChatGPT, as a voting application for upcoming popular votes. It analyses the positions and suggestions of ChatGPT on four voting objects (three initiatives and one referendum) that have not been submitted to vote at the time of conducting the experiment. We ask ChatGPT to impersonate someone from a given placement in the Swiss political landscape (neutral, centrist, left, right, progressist and conservative) and we compare these answers with its default. We do a similar experiment asking ChatGPT to position in favor or against arguments about each object, and we verify whether it is congruent with its voting pattern. To alleviate concerns about random answer generation, we fix the temperature, collect answers to the same questions 100 times, and randomly shuffle the arguments. ChatGPT's voting decisions on the four direct democracy objects varied depending on the version of the model used (GPT-3.5 or GPT-4) and the language of the prompts (French or German), thus suggesting that cultural contexts and model iterations may influence ChatGPT's interpretation and responses to political prompts. Voting decisions were not consistently aligned with a left-progressist position and are influenced by factors beyond a simple left-right ideological spectrum. ChatGPT's positions on the arguments presented for each voting object generally reflected its voting decisions. Results underscore the importance of considering language, model versions, and contextual factors when interpreting ChatGPT's outputs in political settings.