On this page

Synthesis: Pang, Zuo and Li (2026) gave five Chinese preservice mathematics teachers an AI chat interface beside a Word document and asked for a complete lesson plan, then reconstructed what happened decision by decision from screen recordings, concurrent talk, stimulated recall, and written reflections. Across 34 decision episodes, authority was retained in 16, shared in 12, and ceded in 6, and it moved within every case, so no participant could be labeled simply critical or dependent. The study's contribution is its unit of analysis: not how much a novice prompted or edited, but whether a proposal was judged on grounds independent of the AI and whether that judgment reached the written plan. Two narrow failures make the risk concrete — a closed verification loop, in which the model that raised a doubt was asked to settle it (four episodes, two cases), and an insight–action gap, in which a recognized problem never became a teacher move (three episodes, one case).

Key Findings

  1. Five cases, 34 episodes, and variation inside every case. The five third-year preservice teachers (3 women, 2 men; aged 21–22, M = 21.4) were selected from a cohort of 42 through a screening task and cluster sampling; each had completed a two-week observation practicum and used GenAI weekly for coursework, but none had formal instruction in AI lesson planning. Episodes per case (T1–T5) were 6, 8, 8, 6, and 6.
  2. Retained authority was the most common position, but not the stable one. Of 34 episodes, 16 were retained, 12 shared, and 6 ceded. Case profiles ranged from T2 (4 retained, 4 shared, 0 ceded) to T4 (2 retained, 1 shared, 3 ceded), and T1 both ceded a whole-plan decision and later reconstructed a proof sequence.
  3. A complete AI plan changed the task from construction to authorization. T1 asked Doubao for a Grade 8 lesson on the properties of parallelograms, scrolled through the response once, and pasted the entire text. Her reflection was explicit: "I thought there was no major problem with the overall framework, but I was not very sure that it was a good design, so I accepted it." The plan became the default before she had articulated a criterion, and the episode was coded as ceded.
  4. Authority could be reclaimed within the same session. T1 later noticed that the plan omitted anticipated student responses and prompted, "You overlooked anticipated student responses and response strategies in your design"; she selected what to include while Doubao supplied the content, a shared episode. For the second proof she reorganized the sequence so students would attempt it after the teacher modeled the first, wrote her own anticipated response, and was coded as retained.
  5. A teacher-authored structure helped but did not hold. T3 divided the lesson into four modules before prompting, explaining that "If I have my own framework first, I can keep adjusting it." That structure was coded as retained. Late in the session, after prolonged hesitation, he adopted an AI-generated hands-on activity and a verbatim script under time pressure; with no independent warrant, the episode was ceded.
  6. Visible activity did not predict the authority position. Prompting the AI to refine occurred in nine episodes (3 retained, 6 shared), modification in four (2 and 2), deletion in six (all retained), verification through AI in four (all shared), and accepting or noticing without action in three each (all ceded). Of ten episodes warranted by student thinking, five were retained, two shared, and three ceded.
  7. Two vulnerabilities were visible only in the process record. The closed verification loop appeared in four episodes across two cases (three in T2, one in T3), and in three of the four, AI confirmation was the warrant for retaining the content. The insight–action gap appeared in three of T4's six episodes — missing guidance for anticipated errors, teacher participation in group discussion, and a typical-error board example — none of which reached the plan. His account was that he would return to the guidance after finishing the overall design, "but by the time I submitted it, I may have forgotten."

How authority was inferred

The coding did not treat AI authorship or the number of edits as evidence. Each consequential episode was described on four dimensions: whether the knowledge object came from the AI, the participant, or both; what the participant did with it (accepted, queried, verified, modified, deleted, reconstructed, or left unnoticed); which warrant justified the action; and whether the decision was enacted in the written plan. Professional judgment became visible only at the intersection of warrant and enactment.

Retained authority required a professional warrant independent of the AI plus an enacted change. Shared authority meant the participant constrained, compared, or selected AI content while the AI still supplied substance or confirmation. Ceded authority covered two routes: adopting AI output without one's own evaluation, or recognizing a design problem and leaving the AI-shaped plan unchanged. The authors are explicit that these labels are not a quality ranking — shared authority can be productive collaboration — and that authority was a property of a decision, not of a person.

That design choice is what makes reliance visible at all. A final lesson plan can look identical whether a teacher evaluated a proposal or never noticed it, and a transcript of prompts can look busy while the AI determines the mathematics. Separating warrant from enactment distinguishes reliance that was never monitored from monitoring that produced no control. The authors read T1's acceptance through automation bias research but state that the construct was not measured, so that reading is an interpretation rather than a demonstrated effect.

What separated a defensible decision from a fluent one

Uncertainty was the pivot, but not in a fixed direction. When participants resolved doubt with an AI-independent resource, authority stayed with them. Four resources recurred: knowledge of student thinking (T1 expected students to reuse the first proof; T4 judged that a first lesson should stay accessible), practicum experience (T2 opened with a review of linear functions for classroom reasons), mathematical structure (T1's redesign rested on proving both properties through triangle congruence), and judgment about lesson capacity (T1 and T5 reduced examples and practice, and T5 removed fixed timings because the planned time "cannot really be guaranteed").

When the same model that generated a claim also validated it, the interaction produced confidence without evidence. T2 repeatedly asked DeepSeek when slope content was taught and retained the answer; T3 asked whether the curriculum information was current but consulted no official document. Both were monitoring without control, in the metacognitive sense, because the epistemic source never changed.

The cases also show why a single label fails. T1 ceded one episode and retained another; T4 enacted a student-oriented sequencing decision while leaving three pedagogical insights outside the plan. Authority had to be re-established for each consequential decision rather than secured once for the task.

The authors turn these findings into five design propositions — a pre-AI instructional position, a decision ledger, separated generation and verification, an insight–action review at submission, and feedback from the planning record. They label the propositions untested and offer them for evaluation in future intervention studies, not as established practice.

What this means for practice

  • Instructors. Require a short pre-AI instructional position — learner group, central mathematical goal, anticipated prior knowledge, and one likely difficulty — so that the novice has a frame for judging the generated plan.
  • Instructors. Keep a decision ledger for selected AI proposals: source, action, warrant, and the resulting change. It makes judgment visible in a way prompt counts cannot, and it gives feedback something concrete to attach to.
  • Faculty developers. Separate generation from verification in task design. Grade placement should be checked against an official curriculum or textbook, and a mathematical claim should be independently justified; asking the generating model to confirm itself is a provisional inquiry, not verification.
  • Faculty developers. Build an insight–action review into submission: every concern voiced during planning must be matched to a teacher move in the plan, or explicitly recorded as unresolved.
  • Researchers. Code enactment, not activity. The same visible action carried different meanings across these episodes, and a self-report or product-only design would have overstated professional control.

Limitations

  • The study supports analytic rather than statistical generalization: five cases came from one Chinese teacher education program, and one invited participant withdrew, leaving Cluster 1 represented by a single case.
  • No focal case represented the low self-report, high-performance screening profile (n = 5 in the cohort), and the two- and three-cluster solutions were nearly equivalent, so screening results are context for case selection rather than a separate finding.
  • The two mechanisms rest on thin evidence: the closed verification loop appeared in two cases and the insight–action gap in one, so their frequency among preservice teachers is unknown.
  • Concurrent verbalization was encouraged but uneven, and silence was not treated as absence of thought; tacit mathematical or pedagogical reasoning may be underrepresented in the warrant codes.
  • Stimulated recall took place 15 days after the task, some deletions could not be recalled exactly, and researchers were present while screens were recorded, which could have led participants to present themselves as more evaluative than they would alone.
  • Participants used different platforms and settings — three Doubao and two DeepSeek, with web search enabled for T2 and deep thinking for T3, and no model version recorded — so case differences may partly reflect the systems rather than the teachers; planning is also prospective practice, not classroom enactment.

Citation

Pang, J., Zuo, H., & Li, G. (2026). Negotiating epistemic authority in AI-assisted lesson planning: a multiple-case study of preservice mathematics teachers. Frontiers in Psychology, 17, Article 2015390. https://doi.org/10.3389/fpsyg.2026.2015390

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.