The research, conducted by Sejun Oh, Haemee Rim, Seongkyeong Kim, and Yong-Oh Lee, tackles a stubborn problem in mathematics education. Linear functions demand that students coordinate equations, graphs, and changing quantities simultaneously, and decades of research show that learners struggle to translate between these representations and to explain why a line behaves as it does. Formative feedback, the kind that tells students not just whether they are right but what to do next, is exactly what this kind of reasoning needs. Yet providing it mid-lesson is enormously demanding for teachers, because every student stumbles at a different moment on a different idea.

The team’s answer was the AlgeoMath module, a system built on a deliberate division of labor. For tasks in which students constructed or manipulated graphs, correctness was determined by rule-based evaluators and the AlgeoMath application programming interface, which checked structured outputs such as slope, intercept, and relational constraints rather than interpreting screenshots. For written explanation tasks, a large language model, specifically gpt-3.5-turbo-0125 accessed through OpenAI’s Chat Completions API, generated staged feedback aligned with rubric criteria and moderated templates. No fine-tuning was performed; instead, the model was constrained through prompt templates, rubric criteria, staged output fields, and expert-reviewed feedback structures.

The gamification layer was deliberately restrained. A short narrative frame, progress indicators, and a turtle helper organized the sequence of 21 core tasks, and after an incorrect submission the turtle displayed a hint while students could retry on the same screen. Crucially, the designers avoided leaderboards, badges, rankings, and competitive scoring entirely. The game elements existed not to dangle rewards but to make the feedback cycle visible: students always knew where they were, what the turtle was suggesting, and what their next action could be. Hints came in three levels, from a light nudge to re-check part of an answer, through a stronger cue, to a final stage that could include a concise direct explanation enabling the student to proceed.

The classroom results were striking on the implementation side. All 22 ninth-graders at School A completed all 24 tasks, while the 27 eighth-graders at School B completed 98.1 percent of their 21-task sequence. Across both schools, roughly 70.7 percent of task instances were solved on the first submission, with a mean of about 1.53 submission events per task. More telling was what happened after failure: among initially incorrect answers that received a second attempt, 57.5 percent were corrected on that attempt, and among those still wrong after two tries, 43.3 percent were fixed on the third. The staged hint cycle, in other words, was not decoration; students were actually reading the feedback and using it to revise.

The module’s embedded scoring revealed a consistent pattern across both classrooms. Overall performance was high at School A, where the material served as enrichment after the main linear functions content had been taught, with a mean score of 96.1 out of 100, compared with 78.8 at School B, where eighth-graders met the material during the regular unit. But the rubric-dimension profiles told a more interesting story: in both schools, covariational reasoning and mathematical communication and justification were the weakest dimensions. Students could often manipulate graphs and identify slopes, but coordinating how two quantities change together, and explaining why equal slopes produce parallel lines, remained the hardest part. These profiles give teachers a concrete map of where follow-up instruction should go.

Students’ immediate self-reports also shifted. Of eight survey items administered before and after the lesson to 48 matched pairs, six showed statistically significant increases after Holm correction for multiple comparisons: enjoyment of learning mathematics, interest in mathematics, self-perceived competence, self-confidence, understanding difficult content, and perceived learning speed, with effect sizes ranging from 0.41 to 0.64. Notably, two general liking-and-interest items did not change significantly, a selective pattern the authors interpret as reflecting the lesson’s task experience rather than a wholesale transformation of attitudes toward mathematics. Written reflections echoed this, with students describing the game-like format as enjoyable and the turtle hints as helping them find their mistakes and keep going.

The study’s most sobering numbers concern the quality of the AI’s judgments. When 100 constructed responses were sampled and two experienced educators independently rated them as high, middle, or low against the same rubric, the educators agreed with each other 92 percent of the time, with a Cohen’s kappa of 0.863. The automated judgments agreed exactly with each educator only 69 percent of the time, with kappa values around 0.44 to 0.45. The disagreements concentrated in borderline responses with partially articulated reasoning, pinpointing a specific engineering target: the system needs clearer partial-credit guidance and exemplars of intermediate-quality explanations before its judgments can be trusted without human review.

A separate expert audit added a second layer of scrutiny. A panel of 20 mathematics educators conducted three rounds of design-time review of the feedback template library, which comprised 122 templates linked to the linear function tasks. Twenty-one templates, or 17.2 percent, were flagged at least once, and the dominant problem was not mathematical error, which accounted for only one flag, but scaffold calibration: 17 of the flagged issues involved messages whose level of help did not match their assigned stage, such as a Level 3 slot containing a reflective question rather than direct teaching. Encouragingly, only four flagged templates actually appeared in the pilot logs, accounting for just 2.1 percent of the 1,409 feedback instances delivered, and all templates and prompt rules were updated after the audit.

The authors are careful about the limits of what this pilot shows. The two classes were convenience-sampled, differed in grade level and task exposure, and there was no comparison condition, so no causal claims about learning gains can be made, and independent achievement and transfer were not measured. The logs show submissions after feedback appeared but cannot prove students attended to each hint. Still, the contribution is a concrete design template for teacher-supervised AI feedback: exact mathematical checks where determinism is possible, language-model hints where explanation matters, rubric-based reports for teachers, and human moderation as a standing safeguard. Future work, the team suggests, should prioritize partial-credit exemplars, scaffold calibration, and richer teacher analytics, alongside controlled comparisons with delayed outcome measures to test whether the immediate confidence boost translates into durable mathematical understanding.

Subject of Research: GPT-assisted formative feedback in a gamified linear function learning module for middle school mathematics

Article Title: GPT-assisted formative feedback for middle-school linear function inquiry: Design and classroom evaluation of a gamified AlgeoMath module

Article References: Oh, S., Rim, H., Kim, S., & Lee, Y.-O. (2026). GPT-assisted formative feedback for middle-school linear function inquiry: Design and classroom evaluation of a gamified AlgeoMath module. Heliyon, 12(15), Article e45558. https://doi.org/10.1016/j.heliyon.2026.e45558

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45558

Keywords: formative feedback, large language models, GPT, linear functions, gamification, mathematics education, middle school, covariational reasoning, rubric scoring, human-AI agreement, dynamic graphing, classroom study

The study, conducted during the Fall 2025 semester, spanned three courses of increasing disciplinary complexity: a non-majors food and nutrition course with 45 students, an introduction to bioinformatics course with 17 majors, and an upper-level genes, genomes, and genomics course with 27 majors. Two writing-to-learn assignment genres were tested. Students in the first two courses designed infographics meant to communicate a scientific concept to a general audience, while students in the genomics course wrote evidence-based explanatory articles on emerging ethical issues in genomic science. Every assignment followed a scaffolded, draft-based process, with rubrics distributed from the outset so that all evaluators, human and machine alike, were working from identical criteria.

The workflow was carefully engineered to keep the three feedback streams independent. Each draft simultaneously received formative feedback from the instructor, from two classmates whose scores were averaged into a single composite, and from ChatGPT-5 operating through the OpenAI web interface with memory features disabled, ensuring the model retained nothing between evaluations. The AI received the assignment instructions, the analytic rubric, and a standardized prompt directing it to complete the rubric and provide criterion-specific written comments, with particular attention to sources and citations. No rater saw any other rater’s scores. Only after final grades were posted were feedback sources revealed to students, immediately before they completed an anonymous perception survey.

To measure how closely the three sources aligned, the study used intraclass correlation coefficients, a statistical tool that quantifies agreement among raters on the same set of items. Values below 0.50 were interpreted as poor, 0.50 to 0.75 as moderate, 0.75 to 0.90 as good, and above 0.90 as excellent. The headline finding was that overall agreement among AI, instructor, and peer evaluations was good for the non-majors infographic assignment, with an ICC of 0.79, and good for the upper-level explanatory article, at 0.80, but only moderate for the majors’ bioinformatics infographic, at 0.54. In other words, the machine and the humans largely agreed, but the strength of that consensus depended heavily on what kind of assignment was on the table.

That modality effect became even clearer when the rubric categories were examined individually. For the text-based academic articles, agreement was strongest on evidence and reasoning and broader impact, both hovering around 0.75, while sources and citations lagged at 0.36. For the bioinformatics infographics, agreement collapsed entirely on organization and storytelling and on design and visuals, both hitting an ICC of 0.00, even as sources and citations reached a solid 0.79. The pattern suggests that large language models align most reliably with instructor judgment when assessing written scientific argumentation, and least reliably when the task demands interpretation of visual design, representational choices, and the integration of graphics with scientific content, a domain where multimodal evaluation imposes interpretive demands that current models handle unevenly.

The study also tracked how AI-instructor agreement shifted as students revised their work, and here the two assignment types diverged in opposite directions. For the infographics, agreement between AI and instructor scores declined modestly from draft to final submission, dropping from 0.66 to 0.62 in the non-majors course and from 0.54 to 0.29 in the bioinformatics course. For the academic articles, the opposite occurred: overall agreement rose from 0.72 to 0.80, with notable gains in scientific accuracy, evidence and reasoning, argumentation and organization, and broader impact. One plausible reading is that as students revised text-based arguments using the combined feedback, their work converged on standards that both the instructor and the model recognized, while visual revisions moved in directions the AI judged differently from the human expert.

Student perceptions added a second, equally important layer of evidence. Across both infographic courses, students rated instructor feedback significantly higher than peer and AI feedback on nearly every measure: identifying weaknesses, providing specific and actionable suggestions, improving organization and scientific accuracy, and, most strikingly, trust, which showed the largest effect size in the study with a Kendall’s W of 0.44. Peer and AI feedback were rated similarly on most dimensions, effectively positioning the chatbot as a substitute classmate rather than a surrogate professor. For the upper-level article assignment, differences among the three sources shrank, with significant gaps remaining only for identifying weaknesses, specificity, scientific accuracy, and trust. Notably, students reported comparable effort applying feedback regardless of its source, meaning AI suggestions did not add cognitive burden to the revision process.

The survey also probed why students sometimes ignored the machine. Among the 49 respondents who reported rejecting at least one AI suggestion, the most common reason, cited by 55 percent, was that the AI feedback conflicted with feedback from another source, followed by concerns that it was inaccurate or unclear, each cited by 37 percent, and that it did not match the rubric, cited by 29 percent. Yet the overall verdict on the AI was cautiously positive: 39 percent of students said AI feedback improved the quality of their final product and another 41 percent said it may have. More broadly, 80 percent reported that their ability to communicate science improved through the project, and 84 percent said their understanding of biological content deepened, suggesting that a multi-source feedback workflow including AI did not dilute perceived learning gains.

The author is careful to frame these findings as evidence about one implementation rather than a verdict on any particular model, and the limitations are real. The study took place at a single Hispanic-serving institution with modest enrollments, sample sizes for the reliability analyses ranged from 15 to 38 assignments, and confidence intervals around some ICC estimates were wide. Assignment type, course context, and student experience varied together and could not be disentangled. The perception survey was administered only after feedback sources were disclosed, so students’ ratings of usefulness and trust likely reflect the authority of the instructor’s name as much as the intrinsic quality of the comments. Peer reviewers, moreover, received no formal training, which may have depressed their scores relative to what trained peer review could achieve.

Even with those caveats, the practical message for educators is reasonably clear. Structured implementation matters: the AI in this study operated within assignment-specific prompts, analytic rubrics, and instructor oversight, not as an unrestricted generative tool, and that discipline appears to be what made its feedback usable. The findings support treating large language models as stage-specific scaffolds that help students identify baseline weaknesses, clarify expectations, and engage in early revision cycles, thereby freeing instructor time for the conceptual depth, disciplinary nuance, and rhetorical judgment that remain stubbornly human territory. Students themselves seem to have internalized this hierarchy, trusting the machine about as much as a classmate and the professor far more. As AI systems evolve, the open questions are shifting from whether models can grade to when and how their feedback should enter the workflow, and future work comparing multiple models, blinding feedback sources, and spanning more institutions will be needed to map that territory. For now, the evidence suggests the smartest role for ChatGPT in the science classroom is neither ghost-grader nor gimmick, but a third voice in the room, one worth hearing and wise to verify.

Subject of Research: Comparison of AI, peer, and instructor feedback on undergraduate biology writing assignments

Article Title: Evaluating Artificial Intelligence, Peer, and Instructor Feedback Across Scientific Communication Assignments in Undergraduate Biology

Article References: Boies, L. (2026). Evaluating Artificial Intelligence, Peer, and Instructor Feedback Across Scientific Communication Assignments in Undergraduate Biology. Ecology and Evolution, 16(9), Article e74365. https://doi.org/10.1002/ece3.74365

Image Credits: AI Generated

DOI: 10.1002/ece3.74365

Keywords: artificial intelligence, ChatGPT, formative feedback, science communication, undergraduate biology, peer review, writing-to-learn, inter-rater reliability, STEM education, large language models, assessment rubrics, student perceptions