The evidence
What the research says about feedback
Researchers have studied feedback for decades. The evidence supports some things firmly, others only in part, and the research on feedback written by AI is still young. This page sets out what the studies found, how strong each finding is, and where it stops. Every claim carries its source so that you can check it.
How to read this page
- Effect sizes are given exactly as the authors report them, with the caveats they state.
- Where a study was run with school pupils and not university students, we say so.
- None of these studies tested Gradevane. They are the reasons it is built the way it is.
Last reviewed 11 November 2026.
In brief
Six findings, in one sentence each
On average feedback improves performance, but its effects vary a great deal and a sizeable minority of feedback makes performance worse (Kluger and DeNisi, 1996; Wisniewski, Zierer and Hattie, 2020).
Feedback that carries information about the task and how to improve it does far more than praise, reward or a bare judgement (Hattie and Timperley, 2007; Wisniewski, Zierer and Hattie, 2020).
Feedback helps only when you know the standard, can see how your work compares with it, and still have the chance to act (Sadler, 1989; Nicol and Macfarlane-Dick, 2006).
Explicit criteria make expectations clear and support self-assessment, although the evidence that a rubric alone raises performance in higher education is mixed (Jonsson and Svingby, 2007; Reddy and Andrade, 2010).
A grade can draw attention away from the comments beside it, and showing the comments first has helped in university courses (Butler, 1988; Kuepper-Tetzel and Gardner, 2021).
AI feedback has helped students revise in controlled trials, trained human readers still gave better feedback in a direct comparison, and its effect on lasting learning is not settled (Meyer et al., 2024; Steiss et al., 2024; Weidlich et al., 2025).
Timing
Feedback works best when it can still be used
Sadler (1989) set out three conditions for improvement. You need a concept of the standard you are aiming at, the ability to compare your current work with that standard, and a way to act that closes the gap. Feedback that arrives after the work has been handed in can meet the first two conditions and still fail the third.
Nicol and Macfarlane-Dick (2006) built this into their seven principles of good feedback practice for higher education. One principle is that good feedback gives students the opportunity to close the gap between current and desired performance. Boud and Molloy (2013) go further and argue that feedback should be judged by what the learner does with it, and that courses should be designed so that there is a next task where it can be applied.
In a systematic review of 195 publications, Winstone and colleagues (2017) found that the structure of courses is itself a barrier: when assessments come at the end of a module, students have little chance to apply what the comments say, and long waits reduce use further. Shute (2008) lists timeliness among the qualities of useful formative feedback, while noting that experiments on immediate against delayed feedback give mixed results.
What this means in Gradevane
Feedback on the draft, before the deadline
Gradevane reads a draft, the stage at which every comment can still change the work. Your lecturer's feedback on the submitted piece arrives after that chance has passed.
- You upload the draft with the brief and the marking criteria.
- Feedback comes back while you still have time to revise.
- You can upload a later version and have the two compared.
Content
Good feedback answers three questions
Where am I going?
The goal: what the task and its criteria ask for.
How am I going?
Progress: how the work compares with that goal now.
Where to next?
Action: what to do to move closer to it.
These are the three questions in Hattie and Timperley's model (2007). They also separate feedback about the task, about the process used to do it, and about how the learner manages their own work, from feedback about the person. Their review found that comments aimed at the person, including general praise, carry little information and do least for learning.
The numbers need care. Hattie and Timperley reported an average effect of 0.79 across 12 earlier meta-analyses. When Wisniewski, Zierer and Hattie (2020) went back to the primary studies (435 of them, with more than 61,000 participants) the average was d = 0.48, and the results were so varied that the authors concluded feedback cannot be treated as one consistent treatment. What mattered was the content. High-information feedback averaged d = 0.99 across 42 effect sizes, while reinforcement or punishment averaged d = 0.24.
Kluger and DeNisi (1996) reached a similar conclusion from 607 effect sizes: an average gain of d = 0.41, with more than a third of feedback interventions reducing performance. Their explanation is that feedback helps when it keeps attention on the task and can harm when it turns attention to the self. Shute's review (2008) draws the same practical guidance: specific, about the task, in manageable amounts, without comparison to other people.
What this means in Gradevane
Priorities and actions, tied to your own text
An Enhanced or Deep review is organised around the three questions.
- Where am I going: the review is read against your brief and marking criteria.
- How am I going: strengths, a table that takes each criterion in turn, and a report on each section.
- Where to next: a short list of priorities and five actions that say what to change and why.
Comments point to a place in your draft. There is no praise for its own sake and no comparison with other students.
Look at a sample reviewCriteria
Criteria you can see are criteria you can meet
The first of Nicol and Macfarlane-Dick's principles is that good feedback helps to clarify what good performance is. Jonsson and Svingby (2007) reviewed 75 empirical studies of rubrics and concluded that their main benefit for learning comes from making expectations and criteria explicit, which in turn supports feedback and self-assessment. Panadero and Jonsson (2013) later reviewed the formative use of rubrics and described several routes by which they can improve performance and self-regulation.
The evidence in higher education is thinner than the enthusiasm for rubrics suggests. Reddy and Andrade (2010) found only 20 studies at this level. Students consistently said rubrics clarified what was expected and made marking feel fairer. The findings on performance were inconclusive: some studies found gains, one small study found none, and more than half used weak designs.
Their reading of that pattern is useful. Handing out a rubric may not be enough. The benefit seems to come when students actively use the criteria to judge and revise their own work, which is also what Sadler argued in 1989.
What this means in Gradevane
Read against your own criteria, and checked with you first
Gradevane does not apply a generic standard. It works from the brief and the rubric your module uses.
- Before any feedback is written, the assignment profile shows "Here is what we understood" about your task and criteria, and you confirm or correct it.
- The review then reports on each criterion separately, so you can see where the draft is strong and where it falls short.
- If the brief or criteria are missing, the free checks tell you before you spend a credit.
Grades and comments
Comments get read when they are not buried under a grade
The classic experiments are Ruth Butler's. In one (Butler, 1988), 132 pupils in the fifth and sixth grades of school received comments, grades, or both. Black and Wiliam's review (1998) summarises the result: pupils given comments alone kept their gains, while those given a grade with the comments did not, as if the grade had cancelled the benefit of the comment. A companion study with 200 pupils found that comments led to higher interest and performance than grades or standard praise (Butler, 1987).
These studies deserve their caveats. They involved schoolchildren in the 1980s, the grades ranked pupils against each other, and the pattern was not uniform across pupils of different attainment. They show that a comparative grade can pull attention from the task to the self. They do not show that grades are harmful in themselves.
Evidence closer to university practice points the same way. Winstone and colleagues (2017) report that many students attend to the grade and pass over the comments. Kuepper-Tetzel and Gardner (2021) tested a simple remedy in psychology courses at one university: release the feedback first and the mark a little later. They report better engagement with feedback and better performance on a later assignment. One of their two experiments compared cohorts from different years, so it should be read as suggestive.
What this means in Gradevane
An estimate at the top, the substance underneath
A draft has no mark yet, which removes much of the problem. Gradevane shows a predicted band as a range, with a brief summary beside it, and labels it as an estimate. Only your institution awards marks.
- The band describes your work against your criteria. It never ranks you against other students.
- Most of an Enhanced or Deep review is the priorities, actions and comments.
- You can choose "Feedback only, no predicted band" when you set up a submission.
- Grades are not exaggerated.
Revision
Feedback pays off when you revise with it
Receiving feedback and using it are different things. Carless and Boud (2018) describe feedback literacy as four capacities: appreciating feedback, making judgements about quality, managing the feelings it provokes, and taking action. Winstone and colleagues (2017) catalogue what stops students acting: comments that are vague or full of jargon, too many comments at once, and no later task to apply them to.
The strongest pooled evidence on feedback during writing comes from school. Graham, Hebert and Harris (2015) analysed studies of pupils in grades 1 to 8 and found that feedback on writing improved its quality, with average weighted effects of 0.87 for feedback from adults, 0.58 from peers, 0.62 from self-assessment and 0.38 from computers. We have not found an equivalent meta-analysis for university students, so this is evidence by analogy.
A recent experiment with 70 graduate students is a useful warning (Farrokhnia et al., 2026). AI feedback produced with careful prompting was rated higher in quality than teacher feedback, yet it did not lead to greater improvement in the revised essays. The authors conclude that the quality of feedback is not enough: what students do with it decides the outcome.
What this means in Gradevane
A second version, compared honestly
Gradevane is built around the loop of draft, feedback, revision and second look. The revising is yours.
- Upload a revised draft and the feedback compares the two versions.
- It says plainly whether the work progressed or regressed, criterion by criterion. A revision can improve the argument and weaken the structure, and you will be told.
- A point is described as addressed only where the text has changed.
- Each review names a few priorities first, so the comments can be acted on and do not overwhelm.
AI feedback
What we know, and do not yet know, about AI feedback
Before the current AI models. Automated feedback on writing has been studied for two decades. A meta-analysis by Fleckenstein, Liebenow and Meyer (2023) pooled 20 studies with 2,828 learners and found a medium average effect on writing performance (g = 0.55, 95% confidence interval 0.19 to 0.91), with a similar estimate for university students (g = 0.58). The spread was wide: their prediction interval runs from −1.02 to 2.12, so in some settings the true effect may be negative. A systematic review of school studies found the systems worked best alongside a teacher (Nunes et al., 2022).
Compared with human feedback. Steiss and colleagues (2024) compared 200 pieces of feedback from trained human evaluators with 200 from a widely used AI chatbot on the same secondary-school essays. The human feedback was rated higher on four of five measures: clear directions for improvement, accuracy, prioritising what matters most, and supportive tone. The one exception was how closely the feedback kept to the criteria. The authors still judged AI feedback useful for early drafts, or where a trained reader is not available. In two small university studies of students writing in English as a new language, AI feedback led to outcomes no different from a human tutor's and students were split on which they preferred (Escalante, Pack and Barrett, 2023). A small case study found AI feedback more detailed than instructors' (Dai et al., 2023).
Effects on revision. In a randomised study of 459 upper-secondary students, those who revised an essay with AI-generated feedback improved it more than those who revised with none, a small effect (d = 0.19), and reported more motivation (d = 0.36) and more positive emotions (d = 0.34) (Meyer et al., 2024). The comparison was with no feedback at all, on one task.
Effects on learning. This is where the evidence is weakest. A meta-analysis covering 69 articles found that using an AI chatbot improved measured performance and also reduced mental effort (Deng et al., 2025). Weidlich and colleagues (2025) audited a sample of those experiments and found that only a small minority described the treatment, the control group and a valid measure of learning well enough for the gains to be attributed to the AI. In a field experiment with nearly 1,000 school mathematics students, an unrestricted AI assistant raised practice scores by 48 per cent, yet once it was taken away those students scored 17 per cent lower than students who had never had it; a version with safeguards designed to protect learning largely avoided the harm (Bastani et al., 2025).
AI and marks. When two versions of an AI chatbot marked 50 undergraduate bioscience essays, they generally gave higher marks than the human markers and bunched marks towards the middle, raising weaker essays and lowering stronger ones. The authors conclude that “current LLMs do not reliably reproduce human judgement” (Kay et al., 2026).
What this means in Gradevane
Feedback only, never writing
The research points to where AI help turns into harm: when it does the work for you. Gradevane is designed to stay on the other side of that line.
- It never writes text you could submit: no rewrites, no model answers. Comments say what to change and why.
- A comment is attached to your draft only where the words it refers to are found there.
- The predicted band is a range and an estimate. Grades are not exaggerated, and only your institution awards marks.
- It is a second reader. It does not replace your lecturer, your supervisor or your own judgement.
The numbers
Effect sizes, as their authors report them
| Source | What was pooled or tested | Reported effect | Read it with this in mind |
|---|---|---|---|
| Kluger and DeNisi (1996) | 607 effect sizes, 23,663 observations, feedback interventions of many kinds | d = 0.41 | More than a third of the interventions reduced performance. |
| Hattie and Timperley (2007) | 12 earlier meta-analyses of feedback | 0.79 | An average of averages. The 2020 re-analysis of primary studies was lower. |
| Wisniewski, Zierer and Hattie (2020) | 435 studies, 994 effect sizes, more than 61,000 participants | d = 0.48 overall; 0.99 high-information; 0.24 reinforcement or punishment | Very high variation between studies. Signs of publication bias in journal articles. About 17 per cent of effects were negative. |
| Graham, Hebert and Harris (2015) | Feedback on writing, pupils in grades 1 to 8 | 0.87 adults; 0.58 peers; 0.62 self; 0.38 computer | School pupils. We found no equivalent pooled estimate for university students. |
| Fleckenstein, Liebenow and Meyer (2023) | 20 studies of automated writing feedback, 2,828 learners | g = 0.55 (0.19 to 0.91); 0.58 at university level | Prediction interval −1.02 to 2.12. Says little about the newest AI models. |
| Meyer et al. (2024) | 459 upper-secondary students revising one essay with AI feedback or none | d = 0.19 revision; 0.36 motivation; 0.34 positive emotions | Compared with no feedback. Essay quality was scored automatically. |
Limits
What the research does not show
A page like this can make a product sound proven. Gradevane is new, and the honest position is narrower than that.
If you are a lecturer or researcher and think we have misread a study or missed one that matters, tell us and we will correct this page.
None of these studies is about Gradevane
They studied teachers, tutors, peers and other software. Gradevane has not been evaluated in a published study. The research explains our design choices. It is not evidence that the product works.
We make no claim that Gradevane raises your grade
The mark you receive depends on what you do with the feedback, on the assignment and on your markers. The predicted band is an estimate, and only your institution awards marks.
Averages hide wide variation
Every large review on this page reports that effects differ sharply between studies. An average benefit across thousands of learners does not promise a benefit for one person on one assignment.
Some feedback makes work worse
Kluger and DeNisi found that more than a third of feedback interventions reduced performance. Feedback can be wrong, badly aimed or discouraging, and that includes ours. If a comment conflicts with your brief or your lecturer's guidance, follow the brief and your lecturer.
Much of the evidence comes from schools
Butler's experiments, the meta-analysis of feedback on writing, and several of the AI studies involved school pupils. The higher-education evidence on rubrics and on engagement with feedback relies heavily on small studies and on what students say they do.
The evidence on AI feedback is recent and mixed
Most studies are short, use a single task, and measure the revised text and not what the student can do unaided later. Trained human readers gave better feedback in the most careful comparison so far. AI models also change quickly, so a finding about one may not hold for the next.
Relying on AI can cost you the learning
Several studies report less mental effort or weaker unaided performance when an AI tool does the thinking. Gradevane does not write for you, but it cannot make you do the revising. That part of the learning remains with you.
Sources
The full reference list
29 sources. Each link opens a record of the publication that we checked on 11 November 2026: the publisher, the DOI, the ERIC database, or the authors' university repository. Some full texts need a library login.
Titles are given as published, so some name the AI products that the researchers studied.
Feedback and formative assessment
- Black, P. and Wiliam, D. (1998) ‘Assessment and classroom learning’, Assessment in Education: Principles, Policy and Practice, 5(1), pp. 7 to 74. doi.org/10.1080/0969595980050102 (opens in a new tab)
- Hattie, J. and Timperley, H. (2007) ‘The power of feedback’, Review of Educational Research, 77(1), pp. 81 to 112. ERIC record EJ782448 (opens in a new tab)
- Kluger, A. N. and 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), pp. 254 to 284. Hebrew University of Jerusalem record (opens in a new tab)
- Sadler, D. R. (1989) ‘Formative assessment and the design of instructional systems’, Instructional Science, 18, pp. 119 to 144. link.springer.com/doi/10.1007/BF00117714 (opens in a new tab)
- Shute, V. J. (2008) ‘Focus on formative feedback’, Review of Educational Research, 78(1), pp. 153 to 189. Author's copy, Florida State University (opens in a new tab)
- Wisniewski, B., Zierer, K. and Hattie, J. (2020) ‘The power of feedback revisited: a meta-analysis of educational feedback research’, Frontiers in Psychology, 10, article 3087. Open access on PubMed Central (opens in a new tab)
Feedback in higher education
- Boud, D. and Molloy, E. (2013) ‘Rethinking models of feedback for learning: the challenge of design’, Assessment and Evaluation in Higher Education, 38(6), pp. 698 to 712. University of Technology Sydney record (opens in a new tab)
- Carless, D. and Boud, D. (2018) ‘The development of student feedback literacy: enabling uptake of feedback’, Assessment and Evaluation in Higher Education, 43(8), pp. 1315 to 1325. University of Technology Sydney record (opens in a new tab)
- Nicol, D. J. and Macfarlane-Dick, D. (2006) ‘Formative assessment and self-regulated learning: a model and seven principles of good feedback practice’, Studies in Higher Education, 31(2), pp. 199 to 218. University of Strathclyde record (opens in a new tab)
- Winstone, N. E., Nash, R. A., Parker, M. and Rowntree, J. (2017) ‘Supporting learners' agentic engagement with feedback: a systematic review and a taxonomy of recipience processes’, Educational Psychologist, 52(1), pp. 17 to 37. Aston University record (opens in a new tab)
Grades and comments
- Butler, R. (1987) ‘Task-involving and ego-involving properties of evaluation: effects of different feedback conditions on motivational perceptions, interest, and performance’, Journal of Educational Psychology, 79(4), pp. 474 to 482. Hebrew University of Jerusalem record (opens in a new tab)
- Butler, R. (1988) ‘Enhancing and undermining intrinsic motivation: the effects of task-involving and ego-involving evaluation on interest and performance’, British Journal of Educational Psychology, 58(1), pp. 1 to 14. ERIC record EJ380489 (opens in a new tab)
- Butler, R. and Nisan, M. (1986) ‘Effects of no feedback, task-related comments, and grades on intrinsic motivation and performance’, Journal of Educational Psychology, 78(3), pp. 210 to 216. Hebrew University of Jerusalem record (opens in a new tab)
- Kuepper-Tetzel, C. E. and Gardner, P. L. (2021) ‘Effects of temporary mark withholding on academic performance’, Psychology Learning and Teaching, published online 9 March 2021. University of St Andrews record (opens in a new tab)
Rubrics and criteria
- Jonsson, A. and Svingby, G. (2007) ‘The use of scoring rubrics: reliability, validity and educational consequences’, Educational Research Review, 2(2), pp. 130 to 144. ScienceDirect (opens in a new tab)
- Panadero, E. and Jonsson, A. (2013) ‘The use of scoring rubrics for formative assessment purposes revisited: a review’, Educational Research Review, 9, pp. 129 to 144. doi.org/10.1016/j.edurev.2013.01.002 (opens in a new tab)
- Reddy, Y. M. and Andrade, H. (2010) ‘A review of rubric use in higher education’, Assessment and Evaluation in Higher Education, 35(4), pp. 435 to 448. doi.org/10.1080/02602930902862859 (opens in a new tab)
Feedback on writing and automated feedback
- Fleckenstein, J., Liebenow, L. W. and Meyer, J. (2023) ‘Automated feedback and writing: a multi-level meta-analysis of effects on students' performance’, Frontiers in Artificial Intelligence, 6, article 1162454. Open access at Frontiers (opens in a new tab)
- Graham, S., Hebert, M. and Harris, K. R. (2015) ‘Formative assessment and writing: a meta-analysis’, The Elementary School Journal, 115(4), pp. 523 to 547. Australian Catholic University record (opens in a new tab)
- Nunes, A., Cordeiro, C., Limpo, T. and Castro, S. L. (2022) ‘Effectiveness of automated writing evaluation systems in school settings: a systematic review of studies from 2000 to 2020’, Journal of Computer Assisted Learning, 38(2), pp. 599 to 620. University of Porto record (opens in a new tab)
Feedback from large language models
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025) ‘Generative AI without guardrails can harm learning: evidence from high school mathematics’, Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122 (opens in a new tab)
- Dai, W., Lin, J., Jin, H., Li, T., Tsai, Y.-S., Gašević, D. and Chen, G. (2023) ‘Can large language models provide feedback to students? A case study on ChatGPT’, Proceedings of the 2023 IEEE International Conference on Advanced Learning Technologies, pp. 323 to 325. University of Hong Kong record (opens in a new tab)
- Deng, R., Jiang, M., Yu, X., Lu, Y. and Liu, S. (2025) ‘Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies’, Computers and Education, 227, article 105224. doi.org/10.1016/j.compedu.2024.105224 (opens in a new tab)
- Escalante, J., Pack, A. and Barrett, A. (2023) ‘AI-generated feedback on writing: insights into efficacy and ENL student preference’, International Journal of Educational Technology in Higher Education, 20, article 57. Open access at SpringerOpen (opens in a new tab)
- Farrokhnia, M., Latifi, S., Papadopoulos, P. M., Hogenkamp, L., Gijlers, H., Khosravi, H. and Noroozi, O. (2026) ‘Generative AI offers more, but students revise less: comparing the effects of teacher and AI feedback on student essay revisions’, International Journal of Educational Technology in Higher Education, 23, article 6. University of Twente record (opens in a new tab)
- Kay, W., Rutherford, S., Smith, D., Shore, A. and Francis, N. (2026) ‘Can GenAI be trained to mimic human markers of extended written assignments in higher education?’, Assessment and Evaluation in Higher Education, in press. Cardiff University record (opens in a new tab)
- Meyer, J., Jansen, T., Schiller, R., Liebenow, L. W., Steinbach, M., Horbach, A. and Fleckenstein, J. (2024) ‘Using LLMs to bring evidence-based feedback into the classroom: AI-generated feedback increases secondary students' text revision, motivation, and positive emotions’, Computers and Education: Artificial Intelligence, 6, article 100199. Leibniz Institute for Science and Mathematics Education record (opens in a new tab)
- Steiss, J., Tate, T., Graham, S., Cruz, J., Hebert, M., Wang, J., Moon, Y., Tseng, W., Warschauer, M. and Olson, C. B. (2024) ‘Comparing the quality of human and ChatGPT feedback of students' writing’, Learning and Instruction, 91, article 101894. doi.org/10.1016/j.learninstruc.2024.101894 (opens in a new tab)
- Weidlich, J., Gašević, D., Drachsler, H. and Kirschner, P. (2025) ‘ChatGPT in education: an effect in search of a cause’, Journal of Computer Assisted Learning, 41(5), article e70105. Open University of the Netherlands record (opens in a new tab)
Try it on a draft of your own
Every new account gets three free credits to spend on any level. Read the limits above first, then judge the feedback for yourself.