AI feedback can match teacher feedback when assessment design is strong
Updated Jul 12, 2026
Universities do not need another abstract argument about whether AI can write feedback. They need to know whether students improve, whether they trust what they receive, and what makes automated feedback educationally credible in practice. At Student Voice AI, we read that as a student voice question as much as a technology question. That is why Valentina Grion, Beatrice Doria, Daniele Agostini and Giorgia Slaviero's Assessment & Evaluation in Higher Education paper, "Artificial intelligence and feedback in university education: effectiveness and student perceptions", matters for UK universities piloting AI-supported feedback.
Context and research question
The practical attraction of AI feedback is obvious. Universities want faster, more personalised formative support, but staff time is limited and assessment loads keep rising. The harder question is not whether an LLM can produce comments. It is whether those comments help students improve work in ways that are educationally credible, emotionally acceptable, and operationally defensible.
Grion and colleagues tackle that question in a project-based university course by comparing feedback from two large language models, GPT-o4-mini and DeepSeek R1, with feedback from an expert human teacher. The study used a quasi-experimental design in which 47 student groups, covering 238 students, were randomly assigned to one of the three feedback conditions. The researchers then analysed changes in project performance from pre-feedback to post-feedback. Student perceptions were assessed separately through a validated questionnaire completed by 200 students. For UK higher education teams, that makes the paper especially useful because it tests both educational effect and student response, rather than relying on opinion alone.
Key findings
The headline result is that project performance improved significantly across all three feedback conditions. Students who received feedback from the human teacher improved, but so did students who received feedback from GPT-o4-mini and DeepSeek R1. Crucially, the study did not find significant performance differences between the three sources. That matters because many institutional discussions still assume the question is simply whether AI feedback is worse than human feedback. In this study, it was not.
The paper goes further than a simple "no difference" finding. The equivalence analyses suggested practical comparability between GPT-o4-mini and teacher feedback, while DeepSeek R1 met a non-inferiority threshold. For UK teams, that is a stronger and more useful claim than saying the study failed to detect a gap. It suggests that, in the right conditions, AI feedback can produce improvement that is practically close to what an expert teacher achieved in the same setting.
The authors state the broader conclusion clearly:
"feedback effectiveness depends less on its source than on the pedagogical architecture in which it is embedded."
Students' perceptions were also more positive than many institutions may expect. The study found similarly high levels of mastery, emotions, and satisfaction across all three conditions. That should not be misread as proof that students no longer care who gives feedback. Earlier evidence still suggests students use Generative AI for feedback, but trust teachers more when they judge usefulness and trust in broader terms. What this paper adds is a more precise point: when expectations are explicit and the feedback process is well designed, the gap between human and AI feedback can narrow in practice.
The paper therefore shifts the argument from source to design. The study's final line is not that AI can replace teachers. It is that AI-generated feedback can become a credible part of formative assessment when assessment literacy is strong and criteria are explicit. That is a much more disciplined conclusion, and a more useful one for universities deciding where AI belongs in their assessment and feedback system.
Practical implications
For UK universities, the first implication is to stop treating AI feedback as a plug-in efficiency fix. If an institution wants to test AI support credibly, it should start in tasks where criteria are already clear, students understand the purpose of revision, and there is a defined opportunity to act on the feedback. Otherwise, a weak feedback design may be mistaken for an AI problem, or a weak AI workflow may be masked by a well-run module. The benefit is cleaner evidence about what is actually working.
Second, institutions should ask students more precise questions about AI-supported feedback. A single item on satisfaction is too blunt. Universities need to separate usefulness, trust, clarity, emotional impact, and whether students actually changed their work afterwards. That gives Student Experience, Quality, and Digital Education teams a better basis for deciding whether an AI feedback pilot should expand, pause, or be redesigned. The benefit is sharper local evidence before a pilot becomes institution-wide practice.
Third, universities should analyse open comments about AI feedback systematically rather than rely on a handful of striking remarks. This is where a documented process matters. If students say AI feedback was fast but generic, helpful but untrustworthy, or clear on structure but weak on disciplinary judgement, those are not the same problem. A governed approach such as the student comment analysis governance checklist helps teams separate recurring themes, document limitations, and keep review standards consistent. The benefit is that AI feedback decisions become more evidence-led and less anecdotal.
Finally, institutions should keep human oversight visible even when the performance results look promising. This paper suggests AI feedback can be educationally credible in some formative contexts, not that all LLM-based feedback is interchangeable or governance-ready. Universities still need clear boundaries on where AI is used, how output quality is checked, and how students can question or challenge what they receive. That matters especially when institutions are comparing governed local approaches with the looser use of generic LLMs in student feedback workflows. The benefit is better student trust and stronger institutional defensibility.
FAQ
Q: How should a university pilot AI-generated feedback after reading this paper?
A: Start with low- or medium-stakes formative tasks where the criteria are explicit and students have a real chance to revise their work. Compare student improvement before and after feedback, and collect open comments on trust, clarity, usefulness, and emotional response. That gives a university more than an adoption number. It gives evidence about whether the feedback changed work in practice.
Q: What should institutions keep in mind before generalising from this study?
A: This was one quasi-experimental study in a project-based university course, comparing two LLMs with one expert teacher. The study covered 47 student groups, 238 students in the performance analysis, and 200 students in the perceptions questionnaire. That makes it strong evidence about a specific formative context, but not a blank cheque for every discipline, task type, or feedback use case. Universities should be especially cautious about extrapolating from formative project feedback to high-stakes grading or more complex judgement tasks.
Q: What does this change about student voice work on AI and assessment?
A: It makes student voice more diagnostic. Institutions should not only ask whether students "liked" AI feedback. They should ask whether it felt credible, specific, fair, and usable, and whether students would act on it again. Open-text comments are especially valuable here because they reveal whether the issue sits in speed, tone, trust, standards, or the wider feedback process. That gives universities a much clearer basis for action than a headline positivity score.
References
[Paper Source]: Valentina Grion, Beatrice Doria, Daniele Agostini, Giorgia Slaviero "Artificial intelligence and feedback in university education: effectiveness and student perceptions" DOI: 10.1080/02602938.2026.2697962
Request a walkthrough
Book a free Student Voice Analytics demo
See all-comment coverage, sector benchmarks, and reporting designed for OfS quality and NSS requirements.
- All-comment coverage with HE-tuned taxonomy and sentiment.
- Versioned outputs with TEF-ready reporting.
- Benchmarks and BI-ready exports for boards and Senate.
UK-hosted · No public LLM APIs · Same-day turnaround