LLMs' Leaning in European Elections
This work addresses concerns about political bias in AI systems for users and policymakers, but it is incremental as it extends prior US-focused analysis to Europe.
The study analyzed left-wing political leans in large language models (LLMs) by conducting virtual elections across multiple European countries, confirming the extent of this leaning and revealing that it varies between countries, with non-uniform refusal rates in some cases.
Many studies suggest that LLMs have left wing leans. The article extends previous analysis of US presidential elections considering several virtual elections in multiple European countries. The analysis considers multiple LLMs and the results confirm the extent of the leaning. Furthermore, the results show that the leaning is not uniform between countries. Sometimes, models refuse to take a position in the virtual elections, but the refusal rate itself is not uniform between countries.