From Flipped Classroom to AI-Enhanced Assessment: Sustained Pedagogical Innovations in Undergraduate Mathematics

Synthesis: Law (2026) documents a decade-long Scholarship of Teaching and Learning program in undergraduate mathematics at a research-intensive Hong Kong university, moving from a fully flipped classroom to AI-assisted assessment. The redesign pairs animated lecture videos and a parameterized exercise generator with open-book and oral tests, then adds an AI-assisted grading platform to deliver timely, individualized Feedback at scale. Evaluation rests on multi-year institutional student surveys, mid-term feedback, and peer observations rather than a controlled design, so the claims are indicators of impact, not causal proof. Students reported enhanced conceptual understanding, greater engagement, improved inclusivity, and stronger communication skills, while feedback quality and timeliness improved most in large-enrollment classes. The account shows how practitioner-led innovation can be progressively institutionalized: the grading platform moved from a 2024 pilot to joint faculty funding and, from 2026–27, central university support.

Key Findings

1. A fully flipped classroom, supported by animated lecture videos and an automated exercise generator, has been implemented across all courses taught by the author, with class sizes ranging from a dozen to three-digit figures.
2. In a network-theory course for majors and minors, course effectiveness scores on the official SFTL survey have consistently exceeded 90 out of 100 since 2021, substantially above departmental averages.
3. All written assessments were moved to open-book form with limited cheat sheets, and questions now ask students to explain method choices, proofread proofs, and correct errors rather than reproduce routine procedures.
4. Oral tests, each covering a few chapters with pre-assigned questions and clear rubrics, give direct evidence of conceptual understanding and communication skills but become impractical in cohorts beyond roughly 50 students.
5. An AI-assisted grading platform, first piloted in 2024 with a Faculty Teaching Innovation Fund, improved the timeliness and detail of feedback and the consistency of rubric application, especially in courses with enrollments over 100.
6. Multi-year SFTL data show teacher and course effectiveness scores frequently matched or exceeded departmental averages, with several courses in the top decile; response rates were typically over 50%.
7. Institutional embedding followed: the Faculties of Science and Engineering jointly funded the grading license in 2025–26, and the university moved to central funding from 2026–27 onwards.

Method and Evidence

The paper is a longitudinal practitioner case study within the Scholarship of Teaching and Learning tradition, not a controlled experiment. Evidence was collected over more than ten years from official Student Feedback on Teaching and Learning (SFTL) surveys, mid-term course feedback, Staff-Student Consultative Committee comments, peer observations and formal performance reviews, grant outcomes, and documented patterns of resource adoption by colleagues. The author states plainly that no experimental control groups or randomized designs were used and that no pre- and post-intervention concept inventories or detailed grade-distribution analyses were conducted. The analysis is therefore descriptive and interpretive: it leans on the consistency of positive indicators across years and course contexts, and on evidence of institutional uptake, rather than on strong causal claims about learning gains. The strength of this research design is its horizon and its multiple corroborating sources; its weakness is that most of it is self-reported perception data.

Why assessment had to change

The redesign responds to two pressures: increasingly diverse student populations and the arrival of Generative AI tools since 2022. Standard procedural tasks — the kind an Large Language Models (LLMs) can generate or solve — no longer reliably measure what students know. The author's response was to shift written assessment toward open-book examinations that ask students to interpret results, compare methods, and evaluate arguments, question types that are harder for current AI systems to answer without genuine understanding. Oral tests add a second layer of resilience because real-time, adaptive dialogue is difficult for AI to simulate. Both formats also serve Assessment Validity and Academic Integrity goals, and both support inclusivity: students with weaker preparation or exam anxiety can lean on organized notes rather than memory alone. The trade-off is real — oral assessment does not scale, so it was introduced selectively in courses where communication is explicitly prioritized.

From course fix to institutional infrastructure

The most transferable finding is not a teaching technique but an institutional trajectory. The AI-assisted grading platform began in 2024 as a way to handle feedback in the author's own large service courses. Workshops then introduced it to colleagues in Science and Engineering, interest grew, and the two faculties jointly funded the license in 2025–26; from 2026–27 the university provided central funding. That path — individual initiative, peer dissemination, faculty collaboration, university-wide infrastructure — is a concrete example of Change Management in teaching. It was supported by competitive teaching development grants: as Principal Investigator the author secured more than HK$2.5 million, including a University Grants Committee project on interactive learning videos and an AI-enhanced self-evaluative platform. Faculty and university teaching awards, plus departmental leadership roles, reinforced the cycle. The author frames this as the SoTL aspiration of turning private practice into publicly shared, improvable knowledge.

What this means for practice

• Instructors. Move foundational content out of class and spend the freed time on supervised Problem Solving: putting practice under the instructor's eye both clarifies Misconceptions about AI in real time and reduces unsupervised homework reliance on AI tools.
• Instructors. Redesign questions around explanation, error-correction, and interpretation rather than reproduction, and permit a cheat sheet — the shift lowers the value of rote recall and the leverage of a generative-AI answer.
• Assessment designers. Treat oral tests as a selective instrument, not a universal one; prescribe coverage, pre-assign questions, publish rubrics in advance, and run examiner calibration conversations, since inter-rater reliability is still being developed.
• Curriculum designers. Plan for the workload perception problem: explain the pedagogical rationale up front, publish clear weekly schedules, and use low-stakes pre-class exercises, because resistance tends to fade once students adapt to the structure.
• Administrators. Fund the diffusion, not just the pilot: workshops and peer sharing, followed by faculty and central licensing, are what turned a single instructor's grading fix into shared institutional capacity.

Limitations

• No controlled pre-/post-intervention measures of conceptual understanding (such as concept inventories) and no detailed grade-distribution analyses were conducted, so causal claims about learning gains cannot be made.
• Evidence is drawn from one research-intensive Hong Kong university and courses taught or coordinated by a single author, limiting generalizability to other institutions and instructors.
• The strongest quantitative indicator — network-theory course effectiveness above 90 out of 100 since 2021 — is an institutional satisfaction score, not a learning-outcome measure.
• Oral assessment, a central innovation, is inherently time-intensive and becomes difficult to run meaningfully beyond about 50 students, so its benefits were demonstrated only in smaller courses.
• The AI-assisted grading platform improved timeliness and rubric consistency, but the paper reports no independent evaluation of its accuracy or its effect on student grades.

Citation

Law, K. H. (2026). From flipped classroom to AI-enhanced assessment: Sustained pedagogical innovations in undergraduate mathematics education. Proceedings of the World Conference on Research in Teaching and Education, 5(1), 38–50.