A presentation by Jan Zilinsky, Postdoctoral Fellow at the Technical University of Munich and Research Associate at NYU Center for Social Media and Politics.
Large Language Models (LLMs) increasingly serve as sources of information for citizens, yet the methods used to assess their political biases suffer from significant conceptual flaws. Introducing a framework for auditing political and ideological biases, I systematically measure the quality and directionality of AI-generated political recommendations. The paper tests frontier models with diverse voter profiles, including fully aligned partisans, cross-pressured users with conflicting policy preferences, and scenarios where users say they „dislike all parties equally“.
This event is part of the Political Economy Lunch Seminar series.
