<p>We report on a project in which 78 preservice elementary teachers (PSTs) engaged with two AI-based conversational agents programmed to emulate children with distinct mathematical conceptions related to fractions: one (Gabriel) who did not exhibit a misconception, and another (Noah) who demonstrated a common error—comparing fractional quantities without considering the relative sizes of unequal parts. Using a quantitative content analysis approach, we examined the nature and patterns of PSTs’ interactions, including whether their questions served to elicit, extend, or take over the thinking of the virtual students. Results revealed notable differences between the two chatbot conditions, suggesting that features of the instructional scenario—such as the mathematical task and the nature of the student work—shaped PSTs’ engagement. While eliciting questions were dominant overall, PSTs posed more extending and taking-over questions with Noah, whose incorrect but well-articulated solution prompted stronger instructional moves. Additionally, object-level questioning showed divergent trajectories: PSTs’ questions to Gabriel began with a focus on partitioning and visual representation and gradually diversified, while questions to Noah quickly shifted toward comparison, which became dominant. These findings reinforce that developing responsive teaching skills is deeply contingent on the specific context in which practice occurs. They also highlight the promise of AI tools to support practice responsive teaching skills and demonstrate an approach for analyzing how novices engage with such tools to inform the design of increasingly authentic and effective learning environments.</p>