AI-assisted feedback in higher education: what the evidence says

Feedback helps when it informs

Not all feedback is equal: comments that explain how to improve outperform rewards or bare right-or-wrong corrections.

Feedback is among the most studied influences on learning, and among the most variable. A meta-analysis of 435 studies (994 effect sizes, more than 61,000 learners, from kindergarten to university) found a medium average effect of feedback on learning (d = 0.48), with wide variation between studies: feedback is not one uniform intervention (Wisniewski et al., 2020).

The kind of feedback matters more than the amount. In the same analysis, feedback carrying the most information, including guidance on how to regulate one's own work, showed the largest effects (d = 0.99), while rewards and punishment showed the smallest (d = 0.24) (Wisniewski et al., 2020). An older and broader meta-analysis found that feedback interventions improved performance on average (d = 0.41) but that over a third of them made it worse (Kluger & DeNisi, 1996). More feedback is not the aim; better feedback is.

Feedback only helps if students use it. A systematic review of 195 publications maps the conditions that shape whether learners engage with feedback, from how it is framed to what they are asked to do with it (Winstone et al., 2017), and the idea of student feedback literacy, appreciating feedback, making judgements, managing affect and taking action, frames that engagement as a capability to be developed (Carless & Boud, 2018).

In Marking Assist: Marking Assist drafts rubric-aligned comments that the marker reviews and edits before release, so marking time can go on specific, actionable feedback.

Marking is hard, human work

Time pressure and repetition are real constraints, and early evidence suggests they can affect judgements, not just wellbeing.

The peer-reviewed evidence on marking workload is mostly qualitative. In open responses from 3,807 students and 281 educators at two Australian universities, staff described feedback work as heavily constrained, particularly by time (Henderson et al., 2019).

Early evidence suggests that time pressure and repetition can affect the judgements themselves. In a UK online experiment, 100 adults with at least an undergraduate degree (79 of them with marking experience) each marked ten short essays with errors seeded in; boredom rose across the session, higher boredom was linked to lower marks, and giving markers a rubric did not prevent the effect (Erturk et al., 2022).

That is one study, not a sector-wide estimate. The essays were 200 to 300 words, the setting was artificial and the participants were not module staff. No peer-reviewed figure for the share of marking errors caused by fatigue was found, and none is claimed here.

In Marking Assist: Marking Assist shows the marker, and sends the AI, each script as rendered pages: scripts and slides as the pages the student made, and tables as cleaned text, so judgements rest on the work as the student submitted it.

Criteria help, but judgement varies

Rubrics improve transparency and learning; they do not remove the tacit judgement at the heart of marking.

When 24 experienced UK assessors from four disciplines judged student work, their choice, ranking and scoring of criteria varied (Bloxham et al., 2016). The authors conclude that such decisions are "so complex, intuitive and tacit that variability is inevitable" (p. 466) (Bloxham et al., 2016).

Rubrics, in the sense of criteria with descriptions of each level of performance (Brookhart, 2018), help with transparency and learning. A three-level meta-analysis of 21 controlled studies found that rubrics moderately improved academic performance (g = 0.45), with smaller effects on self-regulated learning (g = 0.23, five studies) and self-efficacy (g = 0.18, four studies) (Panadero et al., 2023). The number of criteria and performance levels made little difference, and the authors note an ongoing debate about whether sharing criteria encourages "criteria compliance" (Panadero et al., 2023). An earlier review of the formative use of rubrics found they may support learning through transparency, reduced anxiety and self-regulation, mostly alongside self- or peer-assessment; it did not address agreement between markers (Panadero & Jonsson, 2013).

Taken together: rubrics are worth having and worth sharing, but the evidence does not show that they remove marker variability. Consistency comes from calibration and moderation, not from the document alone.

In Marking Assist: Marking Assist drafts feedback against your own rubric criteria, while calibration, moderation and the final judgement stay with your marking team.

AI assists; humans should decide

Evidence on AI feedback is promising but mixed; on AI grading of complex work, it is weaker still.

Across 41 studies (4,813 students, 2010 to January 2024), AI and human feedback produced no statistically significant difference in learning performance (g = 0.25, 95% CI [−0.11, 0.60]), with very high variation between studies; the authors recommend a hybrid human–AI approach (Kaliisa et al., 2026). A meta-analysis of automated writing evaluation (20 studies, 2,828 participants, mostly tools that pre-date generative AI) found a medium effect on writing performance (g = 0.55) with significant heterogeneity (Fleckenstein et al., 2023).

Students trust it less. In a survey of 6,960 students at four Australian universities, 90.5% rated teacher feedback somewhat or very trustworthy, against 60.1% for AI feedback; the authors conclude that the two serve different needs, complementary rather than interchangeable (Henderson et al., 2025). Among 457 students at one Swiss university, learning that feedback was AI-generated lowered their ratings of it, which points to source bias as well as quality (Nazaretsky et al., 2024).

On grading complex work the evidence is weaker still. In one study of 463 university exam questions, zero-shot ChatGPT-3.5 gave the same exam grade as the teacher in 30% of exams and an adjacent grade in 45% (Cohen's kappa 0.139), and scored marginally higher (Flodén, 2025). GPT-4o's ratings of 6,482 essays by English language learners showed no substantial gender or socioeconomic bias but significant differences by race and ethnicity, which the author cautions may partly be a statistical artefact of unequal group sizes (Yamashita, 2025). These studies used early models, zero-shot prompting and single contexts; AI evidence ages quickly, in either direction.

In Marking Assist: Marking Assist drafts feedback for the marker to check, and the marker makes the final grading decision; none of the research cited here evaluated Marking Assist.

What AI-assisted marking can and cannot claim

The evidence above sets limits on what any AI-assisted marking tool can honestly say. These are the limits we hold ourselves to on this site.

• We do not say that faster feedback improves learning.

  Earlier syntheses found inconsistent results on the timing of feedback (Wisniewski et al., 2020). Timeliness matters to students, but speed by itself has not been shown to raise attainment. We say "timely feedback students can act on".

• We do not say that feedback always improves performance, or quote an effect size of 0.7 or more.

  Over a third of feedback interventions reduced performance (Kluger & DeNisi, 1996), and the older headline figures were revised to 0.48 by a meta-analysis of primary studies (Wisniewski et al., 2020).

• We do not say that rubrics make marking consistent or objective.

  The formative-use review did not study reliability (Panadero & Jonsson, 2013); experienced assessors varied even with criteria (Bloxham et al., 2016); and a rubric did not prevent boredom-related drift (Erturk et al., 2022).

• We do not say that rubric design details drive learning gains.

  The number of criteria and performance levels made little difference (Panadero et al., 2023).

• We do not say that AI marks as accurately or reliably as a human marker.

  The evidence is mixed and context-dependent: one higher-education study found only slight agreement on grades (Flodén, 2025), and the meta-analytic evidence concerns feedback and learning, not grading accuracy (Kaliisa et al., 2026).

• We do not say that AI feedback is better than human feedback.

  The pooled difference was not significant and heterogeneity was very high (Kaliisa et al., 2026).

• We do not say that AI is unbiased.

  Subgroup differences were found in GPT-4o's ratings (Yamashita, 2025). Absence of bias has to be shown for each context, not assumed.

• We do not say that students do not read their feedback.

  Engagement is conditional on many factors, not absent (Winstone et al., 2017).

• We do not put a percentage on marking errors caused by fatigue.

  No peer-reviewed estimate of that kind was found; one experiment (Erturk et al., 2022) is not a basis for one.

• We do not say that the research proves Marking Assist works.

  None of the studies cited evaluated it. Only an independent evaluation could support product-specific claims.

How Marking Assist applies this

1. The marker grades first. A draft cannot be asked for until the marker has given a grade or band and written their notes; the AI never grades the work and never writes from the script alone.
2. The draft is written from the marker's grade, notes and rubric, in a feedback style the marker chose, and is labelled as a draft throughout.
3. The marker edits the draft, can compare any edit with it, and makes the final decision. Every version is kept.
4. Nothing reaches a student until the marker finalises it. Finalising signs the feedback off in the marker's name and deletes the student's file.

See how it works, pricing and the data-protection summary.

Create an account

References

Studies cited evaluate feedback, marking and AI in general; none evaluated Marking Assist.

1. Bloxham, S., den-Outer, B., Hudson, J., & Price, M. (2016). Let's stop the pretence of consistent marking: Exploring the multiple limitations of assessment criteria. Assessment & Evaluation in Higher Education, 41(3), 466–481. https://doi.org/10.1080/02602938.2015.1024607
2. Brookhart, S. M. (2018). Appropriate criteria: Key to effective rubrics. Frontiers in Education, 3, Article 22. https://doi.org/10.3389/feduc.2018.00022
3. Carless, D., & Boud, D. (2018). The development of student feedback literacy: Enabling uptake of feedback. Assessment & Evaluation in Higher Education, 43(8), 1315–1325. https://doi.org/10.1080/02602938.2018.1463354
4. Erturk, S., van Tilburg, W. A. P., & Igou, E. R. (2022). Off the mark: Repetitive marking undermines essay evaluations due to boredom. Motivation and Emotion, 46(2), 264–275. https://doi.org/10.1007/s11031-022-09929-2
5. Fleckenstein, J., Liebenow, L. W., & Meyer, J. (2023). Automated feedback and writing: A multi-level meta-analysis of effects on students' performance. Frontiers in Artificial Intelligence, 6, Article 1162454. https://doi.org/10.3389/frai.2023.1162454
6. Flodén, J. (2025). Grading exams using large language models: A comparison between human and AI grading of exams in higher education using ChatGPT. British Educational Research Journal, 51(1), 201–224. https://doi.org/10.1002/berj.4069
7. Henderson, M., Bearman, M., Chung, J., Fawns, T., Buckingham Shum, S., Matthews, K. E., & de Mello Heredia, J. (2025). Comparing generative AI and teacher feedback: Student perceptions of usefulness and trustworthiness. Assessment & Evaluation in Higher Education, 51(5), 863–878. https://doi.org/10.1080/02602938.2025.2502582
8. Henderson, M., Ryan, T., & Phillips, M. (2019). The challenges of feedback in higher education. Assessment & Evaluation in Higher Education, 44(8), 1237–1252. https://doi.org/10.1080/02602938.2019.1599815
9. Kaliisa, R., Misiejuk, K., López-Pernas, S., & Saqr, M. (2026). How does artificial intelligence compare to human feedback? A meta-analysis of performance, feedback perception, and learning dispositions. Educational Psychology, 46(1), 80–111. https://doi.org/10.1080/01443410.2025.2553639
10. Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. https://doi.org/10.1037/0033-2909.119.2.254
11. Nazaretsky, T., Mejia-Domenzain, P., Swamy, V., Frej, J., & Käser, T. (2024). AI or human? Evaluating student feedback perceptions in higher education. In Technology enhanced learning for inclusive and equitable quality education: EC-TEL 2024 (Lecture Notes in Computer Science, pp. 284–298). Springer. https://doi.org/10.1007/978-3-031-72315-5_20
12. Panadero, E., & Jonsson, A. (2013). The use of scoring rubrics for formative assessment purposes revisited: A review. Educational Research Review, 9, 129–144. https://doi.org/10.1016/j.edurev.2013.01.002
13. Panadero, E., Jonsson, A., Pinedo, L., & Fernández-Castilla, B. (2023). Effects of rubrics on academic performance, self-regulated learning, and self-efficacy: A meta-analytic review. Educational Psychology Review, 35(4), Article 113. https://doi.org/10.1007/s10648-023-09823-4
14. Winstone, N. E., Nash, R. A., Parker, M., & Rowntree, J. (2017). Supporting learners' agentic engagement with feedback: A systematic review and a taxonomy of recipience processes. Educational Psychologist, 52(1), 17–37. https://doi.org/10.1080/00461520.2016.1207538
15. Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, Article 3087. https://doi.org/10.3389/fpsyg.2019.03087
16. Yamashita, T. (2025). Exploring potential biases in GPT-4o's ratings of English language learners' essays. Language Testing, 42(3), 344–358. https://doi.org/10.1177/02655322251329435

Cite this page

Last reviewed 4 November 2026. Studies cited evaluate feedback, marking and AI in general; none evaluated Marking Assist.