AI Insight
This study examined how five preservice mathematics teachers in China negotiated epistemic authority when using generative AI to create lesson plans. Researchers analyzed 34 decision episodes and found that teachers retained authority in 16 cases, shared it in 12, and ceded it to AI in 6. Two critical vulnerabilities emerged: a "closed verification loop" where teachers asked the same AI to validate questionable outputs, and an "insight-action gap" where recognized pedagogical problems were not addressed in final plans.
Why it matters
The findings suggest that teacher education programs should focus less on whether novices edit AI-generated content and more on ensuring they can independently verify outputs and make warranted decisions. This has important implications for preparing future teachers to use AI tools without developing over-reliance or automation bias that could compromise instructional quality.
Understand the Science
IntroductionGenerative artificial intelligence (GenAI) can provide preservice teachers with a complete lesson plan before they have articulated the reasoning the plan is intended to develop, creating conditions for automation bias and cognitive offloading. This multiple-case study examined how five preservice mathematics teachers in China negotiated epistemic authority while planning using GenAI.MethodsCases were selected from 42 preservice teachers through preliminary quantitative screening. Screen-recording-derived event logs, concurrent verbalizations, video-stimulated recall interviews, and written reflections were integrated into an evidence matrix comprising 34 decision episodes. Authority was interpreted as retained, shared, or ceded based on the relationship among the source of a proposal, the participant’s action, the warrant invoked, and whether that warrant was enacted in the lesson plan.ResultsAuthority was retained in 16 episodes, shared in 12, and ceded in 6, with variation across cases. Visible actions did not map onto authority positions in a fixed way, and professional warrants were sometimes not enacted. Two vulnerabilities emerged. A closed verification loop, observed in four episodes across two cases, occurred when participants recognized uncertainty about AI output yet asked the same AI to validate it. An insight–action gap, illustrated by three episodes in a single case, occurred when recognized pedagogica problems were not translated into planned teacher responses.DiscussionThe findings support the proposition that the educative value of AI-assisted planning depends less on how much novices edit AI output than on whether tasks require independent verification and warranted, enacted decisions.