Translating AI and peer feedback in sustainability-oriented STEM teacher education: pre-service mathematics teachers

Synthesis: Oh (2026) reports an exploratory Mixed-Methods Research study of how 22 South Korean pre-service teachers translate AI-generated and peer feedback into revisions of sustainability-oriented mathematical modeling tasks on school cafeteria food waste, a problem aligned with Sustainable Development Goal 12 and with Sustainability. In a required Math Education course, participants completed an initial task, two feedback rounds in different orders — AI then peer, or peer then AI — and a final revision. Feedback translation is the analytic lens: the pathway through which a comment is judged, accepted, adapted, rejected, or left unresolved and then linked to an observable revision, drawing on Feedback, Feedback Literacy, and Evaluative Judgment rather than treating feedback as transmission. Participants rarely copied feedback passively: 20 of 22 selectively accepted it in at least one round, and 17 made a structural or reconceptual revision. Yet only five showed a positive composite gain in rated task-design quality, so visible revision activity diverged from artifact improvement. AI feedback was valued for specificity, new perspectives, and speed and prompted more validation-focused revisions; peer feedback was valued for mutual communication, classroom realism, and student perspective.

Key Findings

1. Participants rarely copied feedback passively: 20 of the 22 pre-service teachers selectively accepted feedback in at least one round, and 32 of 44 feedback episodes were coded selective acceptance rather than acceptance as written.
2. Revision activity was widespread but uneven: 17 of 22 teachers made at least one structural or reconceptual revision, yet only five showed a positive composite gain in double-coded task-design quality.
3. Task-design quality drifted slightly downward overall: the mean Domain 1 composite fell from 7.12 to 6.88 in the AI-to-peer condition and from 6.79 to 6.50 in the peer-to-AI condition.
4. AI feedback prompted somewhat deeper revision (M = 1.68 vs. 1.45) and more validation-focused revision (11 vs. 8 of 22 episodes) than peer feedback, though final validation ratings did not uniformly rise.
5. Teachers valued AI feedback for specificity (9 mentions), new perspectives (7), and speed (6), and peer feedback for mutual communication (8), classroom realism and context (7), and student perspective (7).
6. Explicit rejection was rare, appearing only twice and both times in the peer-to-AI condition, once because a peer suggestion was judged impractical within a one-week data period.
7. Sequence conditions produced similar composite-quality changes but different engagement trajectories: strong-consistent critical engagement was more common in AI-to-peer, improving trajectories more common in peer-to-AI.

Selective uptake and evaluative judgment

Rather than complying with suggestions, teachers treated feedback as material for professional judgment. Across the analytic sample, selective acceptance appeared in 32 of the 44 feedback episodes, and 20 of the 22 teachers showed selective acceptance in at least one round; the round-level percentages (10 of 16 in the AI-to-peer condition, 22 of 28 in the peer-to-AI condition) describe the coded corpus rather than distinct participants. Modified acceptance was less frequent, and explicit rejection appeared only twice, both in the peer-to-AI condition. At the source level, selective acceptance was equally common for AI and peer feedback (16 of 22 episodes each). The researchers read this as Feedback Literacy in action: teachers decided how a comment should enter the design, what to preserve, and which trade-offs a change would create.

The translation gap between revision and task quality

Seventeen of the 22 teachers made at least one structural or reconceptual revision, and nine did so in both feedback rounds, yet only five showed a positive Domain 1 composite gain — two in the AI-to-peer condition and three in the peer-to-AI condition. Composite scores improved for five, were unchanged for ten, and declined for seven. Stability was the most common subdimension outcome: life-context connection improved in four cases and stayed unchanged in 18, while quantity, assumptions, and data showed the most declines (seven). The authors call this divergence the Feedback translation gap, a mismatch between visible revision activity and final task-quality improvement, because sustainability-oriented modeling tasks are multidimensional design objects in which strengthening one dimension, such as data constraints, can weaken another, such as openness.

AI and peer feedback as complementary resources

In their reflections, teachers treated AI feedback as a technically informative cognitive resource, and it was followed by more validation-focused revision (11 of 22 episodes) than peer feedback (8 of 22). Participants also noted risks such as generic suggestions and insensitivity to classroom constraints; medium vigilance about context limits was mentioned 11 times and high vigilance about errors nine times. Peer feedback was valued for mutual communication (8), classroom realism and context (7), and student perspective (7), functioning as a classroom-user check. Cross-source reflection codes emphasized verification (9) and critical use (9). The study concludes that the two sources offer overlapping but differently emphasized resources for design decisions in STEM teacher education.

Sequence as an ordering context

Because feedback order was assigned at the group level (eight teachers AI-to-peer, 14 peer-to-AI), the authors treat sequence as an ordering context rather than a controlled variable and ran no inferential test of order effects. Condition summaries showed similar mean composite changes (−0.25 and −0.29) alongside different engagement trajectories: strong-consistent critical engagement appeared in two AI-to-peer cases and one peer-to-AI case, whereas improving engagement was more frequent in peer-to-AI (six cases versus one). Both explicit rejections occurred in the peer-to-AI condition. The authors interpret these as within-course patterns in which an early AI frame may foreground technical and validation resources while early peer feedback may foreground classroom feasibility, without claiming that either order is superior.

What this means for practice

• Instructors. Make feedback translation explicit with a decision matrix: for each suggestion, have students record whether they accepted, modified, rejected, or deferred it, name the task-design dimension targeted, justify the decision, and anticipate trade-offs.
• Instructors. End every revision cycle with a coherence audit: check how a change to a prompt, data table, or validation step affects openness, modeling demand, and contextual authenticity.
• Faculty developers. Build source-differentiated verification into courses: scaffold checks of AI suggestions against grade level and data feasibility, while peers test learner interpretability.
• Faculty developers. Assess the process, not only the product — combine final task quality with revision depth, justification of feedback decisions, and reflection on unresolved trade-offs.
• Researchers. Treat uptake, revision depth, and artifact quality as separate constructs, because they diverged across this sample.

Limitations

• The sample was 22 pre-service mathematics teachers in a single required course at one South Korean university, so findings are context-specific.
• Feedback order was assigned at the group level (eight AI-to-peer, 14 peer-to-AI) rather than randomly, so ordering, group composition, task maturity, and carryover cannot be separated.
• Task quality was scored with a composite of six equally weighted subdimensions in a small sample, so offsetting gains and declines may cancel.
• Perceived roles of AI and peer feedback came from post-activity self-report and checkbox records, capturing stated reasoning rather than independent evidence of learning.
• The horizon was a seven-week unit within a 15-week course, so persistence into later design work is unknown.
• Data are not publicly available because they are de-identified student coursework, which limits reanalysis.

Citation

Oh, S. (2026). Translating AI and peer feedback in sustainability-oriented STEM teacher education: pre-service mathematics teachers' uptake, revision, and modeling task-design quality. Frontiers in Psychology.