Evaluating Inquiry: How Reasoning Develops Through Human-AI Interaction

Recent design proposals ask AI systems to challenge their users rather than obey them: to question assumptions, offer counterarguments, and prompt critical thinking. They leave an evaluation question open. How do we tell whether an interaction, challenging or not, improves, preserves, or degrades a user's inquiry, and whether the change lasts? The studies discussed here examine what users do, report, or conclude during or soon after an interaction. This paper develops Inquiry Trajectory, one of the analytical targets of the framework set out in Architecture of Inquiry, as an episode-level evaluation target for human-AI interaction, one that connects changes during inquiry with what the person does later. It defines eight positive dimensions as comparisons against what the task demands and what the evidence supports, and six recurring degradation patterns as proposed mechanisms. It keeps what a model does separate from what changes in the user, and it sets out what to notice, including the limits on attributing a change to the interaction. The paper is conceptual.