Combating Harms of Generative AI in CS1 with Code Review Interviews and a Flipped Classroom

Synthesis. Fowles et al. (2026) offer a harm-reduction blueprint for Generative AI in CS Education: rather than banning large language models (Large Language Models (LLMs)s), make students accountable for understanding whatever code they submit through weekly oral code review interviews. At Utah State University, an flipped classroom freed class time for one-on-one assessments, and three semesters of data showed sharply higher AI usage (paste-to-type ratios up) with no statistically significant decline in exam scores — evidence that structured oral Formative Assessment can preserve Learning Gains even as reliance on AI tools rises.

Key Findings

1. Weekly 15-minute oral code review interviews conducted by trained teaching assistants (TAs), weighted at 70% of each coding assignment grade, preserved student understanding: exam scores showed a slight (~2%), statistically insignificant increase in Fall 2025 versus prior semesters.
2. Keystroke logs confirmed dramatically higher AI usage under the permissive policy — the percentage of pasted-to-total characters rose from 61.0% (Fall 2024) to 68.1% (Fall 2025), p < 0.0001 — yet time-on-task and keystroke counts did not decline.
3. Survey responses were overwhelmingly positive: 90% of students reported that code reviews motivated them to understand their code better, and 65% agreed the reviews helped them avoid over-reliance on AI tools.
4. Negative feedback was logistical rather than pedagogical, centering on TA inconsistency (33.7%) and scheduling difficulties (21.7%) — pointing to scalable implementation, not student resistance, as the key challenge.

Background & Context

The rapid capability growth of large language models has forced computing educators to choose between resisting the technology or embracing it as part of professional practice. This study navigates that divide through a harm-reduction stance: no restrictions on AI use for coding assignments, so long as students can demonstrate genuine understanding in a code review interview. The approach reframes questions of Academic Integrity and AI Literacy by moving accountability from detection to explanation, aligning with frameworks that emphasize critical engagement over prohibition.

The authors position the CS1-CR course against two related lines of prior work. First, research characterizing how students actually use AI tools — from Brender et al.'s categories of debugging, practical development, and conceptual exploration to Ghimire & Edwards' similar taxonomies — finds that conceptual exploration produces stronger learning gains than simple code generation. Second, work on oral exams and peer code reviews documents both the promise of authentic oral assessment and the open problem of scaling it to large introductory classes.

Design of CS1-CR

In the Fall 2025 CS1 course (96 students), every coding assignment submission was followed within 48 hours by a mandatory 15-minute oral code review interview with a trained TA or the instructor. The interview probes the student's understanding of their program and the rationale behind implementation decisions, then requires them to run a set of integration test cases. Grading shifted so the rubric score counted for only 30% of each assignment, with the remaining 70% coming from the code review session; failing to schedule a session within the window resulted in a zero on the assignment. This formative accountability structure incentivizes an active attempt at understanding submitted code regardless of its origin.

The design builds directly on evidence that students who over-rely on AI for code generation show poor learning gains, while those who engage conceptually maintain or improve outcomes. By requiring students to explain their code to a trained peer, CS1-CR converts the potential cognitive offloading of AI use into a self-regulated practice opportunity that deepens rather than bypasses understanding.

The Flipped Classroom Component

To create time for weekly interviews, the instructor canceled one of three weekly class meetings and adopted an flipped classroom model: students watch lecture videos outside class and spend the remaining two days on supplemental coding activities with peer and instructor support. This mirrors engagement-focused pedagogies with documented small positive effects on learning and addresses a persistent scalability barrier that has long limited oral assessment in large courses.

Methods & Evidence

The study compares three semesters (Fall 2023, Fall 2024, and Fall 2025/CS1-CR) sharing the same instructor, materials, quizzes, and exams. Because grading policies differed, the researchers compared scores on two proctored, auto-graded multiple-choice exams, alongside keystroke logs collected via the ShowYourWork plugin in PyCharm for identical assignments in Fall 2024 and CS1-CR. These quantitative measures were paired with an end-of-semester survey combining multiple-choice, Likert, and free-response items, analyzed with a consensus-coded thematic scheme.

The results were consistent across methods. Pairwise t-tests showed no statistically significant change in exam performance despite the new policies, with a ~2% improvement on Exam 1 for CS1-CR. The percentage of pasted-to-total characters rose from 61.0% to 68.1% (p < 0.0001) while paste-event counts stayed roughly constant — indicating students pasted larger blocks of AI-generated code. Surprisingly, time-on-task and keystroke counts did not decline, suggesting students engaged with the code just as much, likely motivated by upcoming reviews. As with much of this literature, the authors note there have been no randomized controlled trials in this space, and pasted text cannot be definitively attributed to AI.

What this means for practice

• Instructors. Replace an AI ban with oral accountability: require a 15-minute code review interview within 48 hours of every coding assignment and let the interview carry 70% of the assignment grade, with the rubric score counting for the remaining 30%.
• Instructors. Free the time the interviews need by moving to a flipped classroom — the course dropped one of three weekly meetings and still raised pasted-to-total characters from 61.0% to 68.1% (p < 0.0001) with no statistically significant decline in exam scores.
• Instructors. Fix the logistics before the pedagogy: TA inconsistency (33.7%) and scheduling difficulties (21.7%) drove the negative feedback, not student resistance to being interviewed, so invest in TA training and a scheduling process.
• Administrators. Plan the staffing ratio before scaling: CS1-CR expanded from just under 100 to over 200 students while keeping the same 11 TAs, a constraint the authors flag as unresolved.
• Researchers. Build the comparison the field still lacks: the authors state there have been no randomized controlled trials in this space, and this three-semester study is a proof of concept rather than causal evidence.

Limitations

• No control group is possible in this design: because code reviews add contact time between students and teaching assistants, the authors state it is not clear how to construct a true control condition, making the Fall 2023, Fall 2024 and Fall 2025 comparison quasi-experimental.
• Single course at a single institution: the Fall 2025 CS1-CR section enrolled 96 students with one instructor, and the end-of-semester survey covers the same N = 96.
• Keystroke logs cannot identify AI use: pastes from outside the IDE are excluded to protect student privacy, so the authors cannot infer that all pasted text was AI-generated, and their analysis does not connect types of AI usage to performance on particular questions.
• Student sentiment is self-reported: 90% reported the reviews motivated them to understand their code better and 65% that reviews helped them avoid over-reliance on AI tools, but these are survey responses paired only with exam scores, not an independent measure of understanding.

Citation

Peter Fowles, Erik Falor, Sulove Bhattarai, John Edwards, Seth Poulsen (2026). Combating Harms of Generative AI in CS1 with Code Review Interviews and a Flipped Classroom. .