P1 Guest Talk by Angana Borah

Title

Investigating Misinformation in Human-LLM interactions: Persuasion, Beliefs and Simulation

Abstract

Misinformation susceptibility (vulnerability to believing or spreading misinformation) varies across individuals and communities, shaped by their beliefs and the information they encounter. As Large Language Models (LLMs) are increasingly integrated to everyday life, they play crucial role in this landscape: they can generate persuasive content to spread misinformation as well as detect and counter misleading claims.

In this talk, I will explore how LLMs can help us better understand misinformation through two complementary perspectives: persuasion and simulation. First, I will discuss how humans and LLMs respond to persuasive arguments for misinformative claims, highlighting differences across demographic contexts. Building on these findings, I will present our work on grounding LLM simulations in human beliefs to better capture these differences in susceptibility. Together, these studies test how misinformation judgments are shaped through interaction and how faithfully LLMs can model human beliefs in misinformation.

Bio

Angana Borah is a fourth-year Computer Science and Engineering Ph.D. candidate at the University of Michigan, Ann Arbor, advised by Prof. Rada Mihalcea.  Her research interests are in NLP and agents, especially evaluating human-LLM and LLM-LLM interactions from a socio-cultural lens, taking inspiration from psychology and cognitive sciences. She has recently been exploring how language models behave in socially grounded settings, including fairness and bias, curiosity-driven interaction, misinformation and persuasion, and cultural alignment. These directions aim to understand how LLMs reason, communicate, and influence people across diverse social and cultural contexts. Alongside her doctoral research, she has interned at Amazon and Bloomberg, where she worked on conversational agents and autoresearch systems.

The talk will be followed by a reception at P1.