Tanise Ceron - Political Biases in LLMs: Evidence from Training Data and Model Behavior
- Date: 01 juin 2026 à 13h
- Salle: 55-65 211 (UFR)
Large language models (LLMs) are now embedded in everyday applications like web search engines, raising concerns about how they influence people’s decisions and beliefs. To avoid unbalanced exposure to viewpoints, it is essential to first understand the biases these models reflect to then mitigate them. In this talk, I present a study showing that political biases are consistent across languages, and examine how post-training does (and does not) affect the political opinions expressed by models. I then discuss a second study that investigates the origins of these biases by analyzing the pre-training and post-training corpora of open-source LLMs. We evaluate how political content in the training data correlates with models’ behavior on specific policy issues. Our results correlate model behavior to the composition of training data, revealing that left-leaning documents dominate across datasets and that pre-training corpora contain substantially more politically engaged content than post-training data. These findings highlight the importance of incorporating political content analysis into data curation pipelines and improving transparency in dataset filtering practices.