Benchmarking LLM Polarization Risk

Benchmark the polarization risk associated with AI-generated political content produced by frontier large language models by systematically measuring whether and to what extent such content amplifies political polarization.

Background

The paper benchmarks political persuasion risks across frontier LLMs and finds that modern systems can surpass human campaign advertisements in shifting political attitudes. Beyond persuasion, the authors note growing concern that AI-generated political content could amplify polarization, citing prior work on generative models and automated influence operations.

To guide future research, the authors explicitly state that several extensions remain open, with the first being the need to benchmark polarization risk. Establishing such benchmarks would complement the persuasion-focused evaluations reported in the paper and help quantify broader societal risks related to AI-mediated political communication.