Instructors are looking at AI to help with marking, one of the slowest parts of the job. A new study suggests students have a clear view on that, and that ChatGPT doesn’t judge good feedback the way students do.
The paper, by Carlton J. Fong and four co-authors, was published in October in the journal Computers and Education Open. It combines two studies with American undergraduates: one compared ChatGPT’s judgments of feedback with students’ judgments, and the other used focus groups to ask students how they would feel if their instructors used AI to mark their work.
Students and ChatGPT ranked feedback differently
The first study used 45 feedback comments on a generic college essay, written by experienced instructors to range from unhelpful to very constructive. Earlier work by the same research group had already had hundreds of students rate the comments. One group of 48 students scored each comment for how positive and how specific it was. A second group of 392 compared the comments in pairs and picked the more constructive one, which produced a ranking from best to worst.
The researchers then asked the free version of ChatGPT (GPT-3.5) to define constructive feedback and to rank the same 45 comments.
The two rankings only partly agreed. The rank correlation between them was 0.34, meaning they shared about 11 per cent of their variation. What mattered to students was clear: how specific a comment was explained most of the differences in their ranking, while praise and length made no significant difference. None of those features explained ChatGPT’s ranking, including specificity, even though specificity appeared in its own definition of constructive feedback. The authors say it’s unclear what ChatGPT was actually basing its ranking on.
Praise from a machine “doesn’t really mean anything”
The second study, in fall 2024, put 20 undergraduates from a large public university in the U.S. Southwest into five online focus groups. Most were first- or second-year students and 35 per cent were international students.
The students were shown ChatGPT’s definition of constructive feedback. It listed six qualities: specific, actionable, balanced, focused on behaviour rather than the person, timely and encouraging. The students agreed on the first two. Feedback had to say exactly what was wrong and how to fix it. “I’m not the Van Gogh of English writing, so I know I need to fix something. Can you tell me what?” one participant said.
They mostly didn’t care about encouragement, and some rejected praise from AI outright. “Especially if it’s coming from AI — I feel like a lot of it is fake,” one student said. “It’s not like a real person telling you that, so it doesn’t really mean anything.”
Speed, often seen as AI’s biggest advantage for marking, didn’t matter much to them either. Several students sympathized with busy instructors who teach large classes. Others said they preferred some time between handing in an assignment and getting it back, so the relief of finishing wasn’t cut short by a list of changes.
The researchers also showed each group 10 feedback comments and asked which ones were written by AI. Only near the end of the session did they reveal that ChatGPT had written all 10. The authors report that participants generally couldn’t tell AI feedback from an instructor’s.
“I can do that at home”
None of the participants had yet had an instructor use AI to mark their work, but most disapproved of the idea. They gave three main reasons:
- Connection: students saw an instructor reading their work and responding to it as an act of care. Instructor feedback can also take account of how a student has improved over time.
- Expertise: they were paying to learn from experts. “If I want feedback from here, I can just ask AI myself,” one student said. “The reason I’m attending this university … is because I want to learn from the instructor, not through AI, because I can do that at home.”
- Fairness: students are often limited in how they can use AI for their own assignments. Several called it hypocritical for instructors to use it to mark the same work, especially major projects that took students weeks.
They did accept some uses. Most were fine with AI flagging grammar and spelling, and many didn’t mind it on discussion posts or small assignments graded on completion. Most were also open to a hybrid approach, where AI drafts the feedback and the instructor checks it, edits it and adds their own comments. One student suggested AI might even reduce bias when an instructor disagrees with a student’s argument.
Most students said they would want instructors to tell them if AI was involved. Because students in the study couldn’t tell the two apart, the authors caution instructors against quietly using AI feedback.
Caveats
The ranking test used ChatGPT 3.5, an older and free version of the model, and ran once in a single session. The authors say they have since seen some evidence that a newer version, which they call “GPT-6 Astra,” lines up more closely with students. Even so, they say it still falls well short of how strongly specificity predicted the students’ rankings. The focus groups were small and drawn from one university, and they asked about a situation the participants hadn’t yet experienced.
The title refers to HAL, the computer in 2001: A Space Odyssey. The authors write that when the last focus group asked whether instructors might start using AI to give feedback, the researchers’ tentative “yes” met with “a sense of dread and disbelief.”
Source: Fong, C. J., Zarei, P., Schallert, D. L., Mahdavivand Fard, S., & Dezhkameh, A., “‘I’m Sorry, Dave. I’m Afraid This Feedback is not Constructive’: Human Perspectives Versus ChatGPT Evaluations of Feedback in an Era of Artificial Intelligence,” Computers and Education Open, 2026. doi:10.1016/j.caeo.2026.100429 (open access, CC BY-NC-ND 4.0)
