The Algorithmic Caricature: Auditing LLM-Generated Political Discourse Across Crises

arXiv·medium signal

Gunjan et al. systematically audit how LLMs generate political text across crisis events, revealing that models produce 'algorithmic caricatures' — fluent but distorted representations of political positions. The study raises concerns about synthetic political discourse at scale, showing LLMs amplify certain framings while suppressing nuance. Important for safety teams and anyone building applications that touch political content or opinion generation.