Generative artificial intelligence (AI)-powered conversation agents such as ChatGPT are increasingly being used in teacher education. Although ChatGPT can provide ample resources for lesson planning, little attention has been paid to how teacher candidates construct prompts and evaluate AI-generated outputs in real time to develop lesson plans. Taking up Bisconti et al.'s conceptualization of generative AI as a social agent in a sociotechnical system, this article investigates how knowledge construction is dynamic and negotiated through interaction when using ChatGPT to develop lesson plans. We applied conversation analysis through Heritage's gradient model of epistemic stance--which captures moment-to-moment expressions of social relationships--to analyze how teacher candidates position their knowledge in relation to that acquired from ChatGPT, as managed through their prompts and questions. Our findings aim to offer insights to help teacher educators develop scaffolds for teacher candidates to critically curate content from AI-powered agents and identify areas needing further instructional support.