Students find AI feedback useful, but not personal enough to trust on its own
Updated May 21, 2026
Faster feedback is easy to promise with Generative AI. Harder is preserving the part students still treat as educational rather than mechanical: the sense that someone knowledgeable has understood their work and can talk them through what to do next. That is why Sarah E. Rose, Louise Taylor, Gary Pheiffer, Zoe Fortune and Natalie Wilde's Assessment & Evaluation in Higher Education paper, "Lacking the ‘personal touch’: students’ perceptions of generative artificial intelligence in assessment feedback", matters. For universities collecting module comments, AI pilot feedback, and wider evidence on students using Generative AI for feedback but still trusting teachers more, it sharpens a practical question: what exactly is missing when AI feedback feels efficient but not fully credible?
Context and research question
Universities are increasingly testing GenAI in assessment support because it appears to solve a familiar problem. Students want feedback quickly, staff time is stretched, and AI can produce apparently polished responses at scale. But feedback only works if students see it as worth acting on. If they read AI-generated comments as generic, inaccurate, or detached from the module context, faster turnaround may create a thinner feedback process rather than a better one.
Rose and colleagues address that issue through a qualitative study with 25 students from five UK higher education institutions, one with a campus in the United Arab Emirates. Participants took part in seven focus groups and two interviews, and the authors used reflexive thematic analysis to examine how students defined useful feedback, whether they thought GenAI could provide it, and how they felt about staff potentially using it. That makes the paper useful for UK Student Experience, Quality, and Market Insights teams because it looks directly at student sense-making rather than assuming that uptake equals trust.
Key findings
Students saw real potential in GenAI feedback, especially around objectivity and individualisation. Participants could imagine AI being useful for quick clarification, low-pressure support, and responses that felt less arbitrary than a rushed human comment. That matters because some students are clearly open to AI in feedback when they need immediate guidance or want help outside staff contact hours, much like recent evidence that students turn to GenAI when assessment support feels private and instant.
The abstract captures that ambivalence clearly:
"students evaluated GenAI as a potentially useful tool for providing objective and individualised feedback, but not necessarily accurate feedback."
Accuracy and context were the real fault line. Students did not judge AI feedback only by whether it sounded supportive or well structured. They wanted to know whether it had actually understood the task, the marking criteria, and the standards of the discipline. In practice, that means AI feedback may sound competent while still feeling risky to act on. For universities, the takeaway is simple: polish is not the same as credibility.
The paper's central contribution is to show that the "personal touch" means more than friendliness. Students used that phrase to describe a combination of human expertise, contextual judgement, accountability, and the possibility of dialogue. They were not only asking for warmer wording. They were asking for feedback that felt connected to a real person who knew the module, could justify the guidance, and could help them interpret it. That is important because it reframes the issue. The gap is not only emotional. It is epistemic and relational.
Students therefore treated AI feedback as support for a process, not a substitute for a relationship. The findings reinforce the idea that effective feedback is dialogic. Students may welcome AI for initial explanation, structure, or rehearsal, but they still want a human when the task becomes more nuanced, consequential, or discipline-specific. That gives universities a more useful design rule than a simple pro-AI or anti-AI position.
Practical implications
For UK universities, the first implication is to evaluate AI-supported feedback across separate dimensions rather than one general satisfaction item. Ask about speed, specificity, accuracy, usefulness, and whether the feedback felt personal enough to trust. That gives teams clearer evidence on whether students are objecting to the source, the wording, the accuracy, or the lack of follow-up, which makes intervention more precise.
Second, institutions should collect open-text comments on what students think is missing from AI feedback. A scaled score cannot show whether the problem was shallow task understanding, weak contextual fit, or the absence of human dialogue. If those comments are being gathered across pilots or modules, they need a defensible workflow for analysing open-text feedback at scale. The benefit is that universities can distinguish convenience from educational value before they standardise a weaker process.
Third, universities should make human oversight visible wherever GenAI is used in feedback workflows. Students need to know whether a tutor wrote the comment, edited AI output, checked it against criteria, or is available to discuss it afterwards. This is also where institutions should be careful about depending on generic LLM workflows for interpreting student comments, especially when reproducibility and traceability matter. The benefit is better trust, clearer expectations, and less avoidable anxiety around authenticity and standards.
Finally, teams should treat AI-feedback comments as governance evidence, not only pedagogy feedback. Concerns about trust, transparency, and contextual accuracy quickly become sensitive when they relate to assessed work. A student comment analysis governance checklist is therefore a practical starting point. Student Voice Analytics can then help institutions group recurring comments about speed, accuracy, trust, and the missing personal touch, so decisions about scaling AI feedback are based on patterns rather than anecdotes. The benefit is a more defensible route from pilot to policy.
FAQ
Q: How should a university pilot AI-generated feedback without weakening student trust?
A: Start with a narrow, low-stakes use case and tell students exactly what the AI is doing and what staff still review personally. Then ask separately about speed, specificity, accuracy, and whether the feedback felt personal enough to trust. That makes it easier to see whether the tool is helping with early clarification or undermining confidence in feedback quality.
Q: What are the methodological limits of this study?
A: This is a qualitative study of 25 students across five UK higher education institutions, including participants in the UK and the UAE. Its value lies in explaining mechanisms and expectations, not in estimating sector-wide prevalence. Universities should therefore use it as a strong interpretive guide for local feedback design, then test the same questions in their own comment and survey data.
Q: What does this change about student voice work on AI more broadly?
A: It suggests that student voice on AI feedback should ask more than whether a tool was useful. Universities need to know whether students still feel seen, whether they trust the source, and whether they can ask follow-up questions when the stakes are real. That makes open comments especially valuable, because they reveal what "personal touch" actually means in practice.
References
[Paper Source]: Sarah E. Rose, Louise Taylor, Gary Pheiffer, Zoe Fortune and Natalie Wilde "Lacking the ‘personal touch’: students’ perceptions of generative artificial intelligence in assessment feedback" DOI: 10.1080/02602938.2026.2658633
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