Student Researchers: Rebecca Kim

Abstract

Conspiracy theories tend to be notoriously hard to change. Still, a recent 2024 experiment showed that structured conversations with an AI chatbot can effectively lower the degree of their acceptance in the long term. However, this work established only the average effect, and there remains one crucial issue: does the AI convince everybody in the same way, or are some people more open to changing their beliefs than others? In this paper, the publicly available data from the original 2024 study is re-examined to understand whose belief in conspiracies was affected by the treatment. Once the original result was replicated (there was a significant difference between belief decrease among those who discussed their conspiracy theories with the AI compared to the control group), the following step was to find out if certain traits affect the extent to which belief in conspiracies changed among the treated participants. Random forest models and regression analyses revealed that the most meaningful trait was trust in institutions. Participants who trusted governmental institutions, scientists, and media changed their minds more. While political party and age were less important, attitudes toward AI (for example, how familiar people are with it, how often they use AI, and how much they trust it) had less of an effect. Importantly, our best regression model only explained a small fraction of the variance in belief change, which suggests that it is hard to predict whether people will be persuaded based solely on their demographic or political characteristics.