Devil’s AI-vocate: Using LLMs to Encourage Perspective-Taking and Empathy

Abstract

Political disagreement is increasingly experienced not only as a clash of opinions, but as a clash of identities and values. Any dissent is assumed to be due to the other person being either misinformed, and so in need of being presented with “the truth”, or being morally corrupt, representing an evil that needs to be fought. Against this backdrop, we explore whether structured interactions with Large Language Models (LLMs) can foster tolerance and help individuals respect opposing viewpoints. In our study, we utilize LLMs to create a” digital twin” of a participant’s own stance on a divisive issue. We then ask the participant to engage in counter-attitudinal advocacy by playing the” devil’s advocate” and arguing against their own beliefs. Our objective is not to change the participant’s underlying opinion, but to reduce out-group animosity and increase empathy for alternative perspectives. In this talk, I will detail our methodology and will present results from this ongoing research. This work builds on our prior work [ZKC+25] of using AI to motivate reflection in a political context, but it shifts from a goal of informing to a goal of de-polarizing.