Mode Misclassification in Long-Dialog Human–AI Interaction

This paper identifies mode misclassification as a distinct error class in long-dialog Human–AI Interaction. In extended collaborations involving repeated transitions between analytical, productive, and meta-reflective work phases, large language models may generate locally coherent responses that are inappropriate to the current operational mode. Unlike hallucinations or factual inaccuracies, these errors arise from misalignment between discrete human mode switching and continuous probabilistic task inference in language models. The phenomenon remains largely invisible in short interactions but becomes salient in long-term cooperative work, where interaction stability is a functional requirement. We argue that mode misclassification is best understood as an architectural mismatch in interaction design rather than a deficit in model knowledge and propose it as a focal point for future HAI research on long-dialog stability.