Confirmation Bias in Human-AI Interactions

Abstract

As AI systems increasingly support human decision-making, concerns have shifted from algorithmic bias toward cognitive distortions emerging during human-AI interaction. This research-in-progress study investigates the role of confirmation bias, defined as users’ tendency to favor belief-congruent information, within iterative AI-supported reasoning. Drawing on a process-oriented perspective, we examine confirmation bias at an individual snippet level, and at an aggregate belief level. Using an experimental survey design, we analyze the effects of AI interaction and prompt guidance on bias and task performance. Our preliminary findings indicate a strong persistence of confirmation bias at the belief level and no significant bias modulation through AI interaction or prompt guidance. Notably, GenAI outperformed human groups in identifying credible information. This study contributes to IS literature by revealing the complex dynamics between humans and AI systems. They also raise questions about the value and limitations of collaborative reasoning during human-AI interaction.