LLMs reveal why opposition parties attack more during campaigns across Europe

Researchers have deployed large language models to measure negative campaigning at unprecedented scale, analyzing nearly 18 million tweets from European parliamentarians across 19 countries between 2017 and 2022. The work reveals systematic patterns in how different parties approach campaign rhetoric, offering fresh insights into the strategic incentives that shape electoral competition.

How LLMs scaled the research#

Comparative studies of campaign negativity have historically been constrained by cost and the challenge of coding across languages. The study demonstrates that zero-shot large language models—models applied without task-specific training data—can classify negative campaigning as accurately as human expert annotators while vastly reducing time and expense. When tested against benchmark datasets in ten languages, LLM classifications closely matched native-speaker judgments and outperformed conventional supervised machine learning methods. This breakthrough enables researchers to analyze political discourse with consistency across linguistic and institutional boundaries.

The patterns: governing vs. opposition#

The analysis uncovered predictable fault lines in campaign tone. Governing and coalition-oriented parties tend to exercise greater restraint in their rhetoric, likely because they face reputational costs for aggressive attacks. By contrast, opposition and outsider parties operate under weaker constraints, allowing them to lean into confrontational messaging. The ideological position of a party also matters significantly: those positioned away from the political center, particularly radical right parties, employ substantially more negative language than centrist competitors. Cabinet parties showed markedly lower negativity scores across the dataset.

Why it matters for political science#

This work provides the largest cross-national evidence base to date on the party-level drivers of campaign negativity. Rather than treating attack rhetoric as random or universal, the findings suggest that negative campaigning emerges from predictable structural conditions—governing responsibilities, coalition membership, and ideological positioning all create incentives that shape how politicians communicate. The study also demonstrates how LLM-powered methods can transform the scale and replicability of research on political discourse, potentially enabling more rapid and consistent monitoring of electoral campaigns across diverse political systems.