Conversational Inoculation to Enhance Resistance to Misinformation

TL;DR

The paper tackles misinformation resistance by applying Cognitive Inoculation Theory through a conversational lens, introducing MindFort and a chatbot named Forty to deliver structured refutation dialogues. In a within-subject online experiment with four health-related topics, it compares Conversational Inoculation against Reading, Writing, and a Control condition, using a five-stage inoculation sequence and measures including certainty changes and IMI scores, complemented by LIWC-based linguistic analysis and qualitative coding. Results show that Conversational Inoculation is a valid method that can outperform baseline susceptibility and, when controlling for individual differences, surpass traditional methods; engagement and trust emerge as critical drivers, while interactional friction can impede effectiveness. The study highlights implications for adaptive, multi-agent, and topic-responsive designs and outlines future work on personalization, varied conversational roles, and broader cognitive constructs, offering a timely, scalable approach to mitigating misinformation in health contexts.

Abstract

Proliferation of misinformation is a globally acknowledged problem. Cognitive Inoculation helps build resistance to different forms of persuasion, such as misinformation. We investigate Conversational Inoculation, a method to help people build resistance to misinformation through dynamic conversations with a chatbot. We built a Web-based system to implement the method, and conducted a within-subject user experiment to compare it with two traditional inoculation methods. Our results validate Conversational Inoculation as a viable novel method, and show how it was able to enhance participants' resistance to misinformation. A qualitative analysis of the conversations between participants and the chatbot reveal independence and trust as factors that boosted the efficiency of Conversational Inoculation, and friction of interaction as a factor hindering it. We discuss the opportunities and challenges of using Conversational Inoculation to combat misinformation. Our work contributes a timely investigation and a promising research direction in scalable ways to combat misinformation.

Figures (8)

• Figure 1: Prototype Screenshot (1/7): Home page of the signed-in user showing their lesson progress. In this image, all lessons are still locked as the participant has not opened the initial form yet by clicking the "Open Form" button.
• Figure 2: Prototype Screenshot (2/7): Stages 1, 3 and 5 involve a 15-point scale measuring the participant's certainty concerning a truthful claim.
• Figure 3: Diagram illustrating our experiment with additional annotations. Each participant completes four lessons consisting of five stages each. The second stage represents the Independent Variable, one of four possible conditions. The change in the participant's self-reported certainty in their truthful belief before and being exposed to misinformation is our Dependent Variable. After all four lessons are completed, the participants get debriefed and informed about the purpose of the study first and then they proceed to fill out Post-Debriefing IMI questionnaire.
• Figure 4: Post-attack certainty change compared between the four conditions. The amount of certainty change shows how much participants' certainty increased or decreased on the 15-point certainty scale after exposing them to misinformation. A lower certainty change score means higher resistance to misinformation.
• Figure 5: Method-wise Inoculation Effectiveness, i.e. Certainty score change when controlling for individual susceptibility, as defined in \ref{['eq:Inoculation_Effectiveness']}). Higher values mean greater impact on one's baseline resistance to misinformation.