An Examination of the Party Leanings of Large Language Models

Proceedings of the Second Workshop on Building Educational Applications Using NLP

Abstract

This paper examines the party leanings of international and Norwegian large language models. The experiments are two-fold; first they are asked to answer the question of a Valgomat–an election affiliation guide–as a neutral observer, and secondly as if it were a paying party member of the parties in the data. Results show that the neutral prompting show centrist leanings, whereas models struggle with mimicking party members. Models with additional training on Norwegian text performed better on this task.