The Election Monitoring Trilemma: Measuring Rare Elite Threats on Social Media With Large Language Models

In 2024, elections across more than 70 countries raised fears that social media would incite violence and erode democratic norms. We collected public posts from candidates, parties, journalists, and influencers in South Africa, India, and Germany, coding a random sample for five outcomes using human and LLM annotation. All five are rare on the public platforms we study: incitement to violence appeared in just 0.2% of posts, attacks on democratic norms in 1.9%, and civic duty appeals in 2.3%. While LLMs accurately identified true positives, their precision on rare categories was poor and for the two rarest and highest-stakes categories—incitement and ethnic appeals—the LLMs’ ability to catch true positives rests on a deliberately enriched sample. In addition to the limitations low precision poses to researchers, we further show that this precision shortfall confronts election monitors with a three-way tradeoff among coverage, throughput, and credibility that no single off-the-shelf configuration resolves.