The Mixing of Facts and Inference in AI Dialogue and the Structure of User Cognitive Load --- An Analysis of Misinterpretation Mechanisms in Human-AI Interaction ---

This paper examines a structural issue in human–AI interaction: the mixing of factual information and inference in conversational AI outputs. While such integration enhances fluency and coherence, it creates ambiguity in how users interpret the content. The study organizes this phenomenon as a cognitive problem rather than a simple case of misinformation. It analyzes how users tend to process AI-generated responses as unified narratives, leading to the loss of distinction between facts, generalizations, and inferred intent. From the perspective of cognitive load, the paper identifies key mechanisms behind misinterpretation, including completion bias, fluency bias, and context integration bias. It further argues that the core issue lies in a mismatch between the structure of information presentation by AI and the cognitive processing patterns of users. Finally, the paper suggests practical mitigation strategies from the user side, such as explicitly requesting separation of facts and inference and clarifying the purpose of interaction.