Control or Support: Early Student Perceptions of Generative AI in a Course on Grading

Synthesis: Guberman (2026) taught a fall 2023 Honors course on grading to 11 upper-level undergraduates and asked them to interpret generative AI through the same lens the course applied to grades: control or support. Nine of the 11 used ChatGPT to draft and rehearse interview questions for a real student interview, then compared the AI's answers with the person's. Students found the AI's responses about grades generic, normative and emotionally flat — a "pros and cons list" rather than an account of how grades feel — and they read unexplained AI bans as controlling. They wanted instructors to talk about AI openly, and they were receptive to AI-generated feedback mainly because feedback had been so scarce, not because they thought machines were better. The study is a small reflective case study, but its central distinction — AI as control versus AI as support — organizes student responses in a way that policy debates usually do not.

Key Findings

1. Students read AI policy through their experience of grading. Bans delivered without rationale felt like normative grading curves: compliance demanded without explanation, and autonomy undermined.
2. AI answers about grades were generic and upbeat. Students reported that ChatGPT reproduced stock, positive, bullet-pointed accounts that lacked the emotion and specificity of a real interview, which they found revealing about what assumptions are built into the model.
3. Students wanted engagement, not endorsement. They asked for instruction on what AI is, where it helps and where it interferes, and opportunities to experiment in class rather than blanket prohibition.
4. Openness to AI feedback came from scarcity. Most students welcomed instructor use of AI feedback because they had received little or no feedback in large classes, not because they judged AI feedback superior; several noted it cannot replace an instructor's insight.
5. Students distinguished busy-work shortcuts from help with hard problems. The same tool was acceptable for drafting questions or summarizing a passage and suspect when it replaced the thinking a difficult task was meant to develop.
6. The findings shaped three later interventions. In a long-running asynchronous history of rock music course, Guberman added a video on AI's potential and pitfalls, an alternate songwriting assignment where AI could draft lyrics with a required process reflection, and a discussion of AI vocal models that raised copyright and ethics questions.

The course and the assignment

The course was an Honors seminar on grading in higher education, taught in fall 2023 to juniors and seniors from varied majors. Its premise was that grades function simultaneously as mechanisms of control and as support for learning. Students first read contemporary instructor commentary on AI — two widely circulated Chronicle of Higher Education pieces, one framing ChatGPT as a plagiarism machine and one arguing that AI means professors must raise grading standards — and then examined education-specific tools, including Packback for discussion forums and a locally developed rubric-based essay feedback tool called Charlie.

The interview assignment asked students to develop questions about grades with a peer, then conduct a real interview. AI entered in two ways: students could use it to draft and sharpen questions, and they could ask it to simulate a student interviewee, with prompts they designed themselves. Guberman deliberately left prompting open so that students would compare tools and prompts in their reflections. Nine of the 11 students chose to use AI, all of them ChatGPT.

What the AI comparison showed

The simulated interviews were useful as a contrast, not as a substitute. Students who used AI to draft questions found it a reasonable starting point, then edited for specificity and relevance. When they prompted ChatGPT to answer as a particular student, some answers aligned closely with the real interviewee's — for example, that the focus shifts from grades to understanding over time — but the overall register was wrong. Students described the answers as unnatural, too formal, overly positive, "stock standard," and "a pros and cons list." Several noted the absence of emotion: the real interviews surfaced stress, pride and frustration, while the AI produced bullet points about motivation, feedback and communication. Guberman reads this as evidence that the model reproduces dominant institutional messaging about grades, flattening exactly the affective dimension that makes grades consequential for students.

That finding matters beyond the classroom exercise. If AI is used to simulate student perspectives — in faculty development, policy consultation or design research — the simulation will tend to return the normative view, not the marginal one. The students' own comparison is a small demonstration of that limit.

Feedback, autonomy and the control–support frame

Students' openness to AI feedback was the most policy-relevant result. Guberman expected that, after a semester of discussing feedback, students would prefer instructor comments. Instead most welcomed AI feedback, especially in large introductory courses, and their reasoning was scarcity: they had experienced too many assignments with little or no response. They also set conditions — AI should guide rather than replace instructor judgment, and the feedback should be worth reading. One student worried that AI feedback could push responses toward conformity.

The autonomy finding is similar in structure. Students did not demand AI in every course; they wanted reasons for its presence or absence, and structured opportunities to experiment. A ban without explanation was experienced as control. This is the same distinction the course drew about grades: a policy that demands compliance teaches students to comply, while a policy that explains its rationale and gives them practice in judging quality supports their development.

Three interventions in a later course

The following summer, Guberman applied the findings in an online asynchronous general education course on the history of rock music that he had taught for nearly two decades. First, a short video explained AI's potential and pitfalls for an unessay assignment — for example, asking it to analyze song structure or track band membership across albums — so students understood why use was permitted but reliance was discouraged. Second, an alternate assignment let AI write song lyrics while students described how they generated and edited them and reflected on the result; the rationale was that lyric writing was a hurdle, not the core learning of the assignment. Third, a discussion of the YouTube channel There I Ruined It, which uses AI vocal models to blend performers with unsuitable songs, connected copyright and ethics to contemporary music practice.

What this means for practice

• Explain the policy, not just the rule. Students experience an unexplained ban as control; a rationale plus a chance to experiment supports autonomy.
• Use AI simulation as a contrast case. Asking students to compare AI answers with a real interview exposes what the model flattens, which is itself a lesson about the technology.
• Treat feedback scarcity as the reason AI feedback appeals. The demand for AI comments is a signal about workload and response time, not proof that AI feedback is good.
• Distinguish shortcuts from thinking. Name which parts of a task are busy work and which carry the learning, and design AI use around that distinction.
• Adapt the assignment, not just the syllabus language. The later interventions changed what students produced and reflected on, which is where the control–support distinction becomes concrete.

Limitations

• This is a single-course reflective case study with 11 students, so the themes describe this group rather than a population; no coding scheme, inter-rater check or member validation is reported.
• The course was an Honors seminar of juniors and seniors at one university, and students self-selected into a course about grading, which likely sensitized them to the control–support frame.
• Data are student reflections and instructor observation; there were no pre/post measures of learning, no comparison section and no follow-up on whether the later interventions changed student behavior.
• The AI tools examined were those available in fall 2023, and the generic, upbeat answers students described may reflect that generation of models more than a stable property.
• Guberman is also the instructor and the analyst, and the article does not describe how reflections were selected for quotation.

Citation

Guberman, D. (2026). Control or Support: Early Student Perceptions of Generative AI in a Course on Grading. The Interdisciplinary Journal for Research on Pedagogy and Learning Design, 1(1).