A Multi-Agent AI Classroom Based On Dual-Process Reasoning Hazards: A Pilot With Prospective Physics Teachers

Responding productively, in real time, to the reasoning students actually produce is among the most difficult and slowest to acquire of the skills physics teachers develop, and prospective teachers get few opportunities to practise it before entering a classroom. In this pilot study, we describe a simulated class of five AI students, each consistently enacting a distinct reasoning pattern drawn from the dual-process theory (DPT) "hazards" framework of physics education research, and report on its first use in a university course for prospective physics teachers. Fifteen graduate students worked through an instructional sequence on DPT and questioning strategies. They then diagnosed two parallel sets of written vignettes, in a crossover arrangement, before and after interacting in pairs with the simulated class on a static-friction problem. The data set is completed by individual written hypotheses and reflections, the session logs and a short feedback questionnaire. Diagnostic scores improved significantly from PRE to POST (n = 11 paired, Wilcoxon p = 0.014, r = 0.79; on a 0-18 scale). During the simulation itself, however, participants asked predominantly uniform guiding questions, and the DPT vocabulary they had been taught appeared in only 2 of 71 substantive teacher turns, while seven of thirteen POST sheets used it readily. We read this knowing-doing gap not as a failure but as a snapshot of where each participant stands on the developmental trajectory of responsiveness to student ideas, a trajectory the simulation makes visible with transcript-level granularity. We discuss design, findings and implications for physics teacher preparation.