Editorial
The landscape of education and…
by Sheree Chong (Ministry of Education, Singapore)
This article was inspired by teachers who have, admirably, attempted to use Generative AI to provide feedback for their primary and secondary school students. Their discoveries have deepened our insights into when and how practitioners should use AI to improve student learning.
Many colleagues started experimenting with AI to provide faster and more personalised feedback on students’ work. While the speed and convenience of AI are appealing, teachers have reported that AI is often too eager to assist, and at times might generate feedback which overwhelms students (and sometimes, even the teacher). On the other spectrum, AI feedback is too directive and provides students with immediate answers. Hence, this has caused many in the fraternity to think about a few questions:
- When does AI feedback actually improve learning?
- How can we use AI feedback to develop learners who can think independently and regulate their own learning?
- Does inaccurate AI feedback, intentionally designed, serve a purpose in helping develop critical thinking skills in our learners?
- What are the implications of AI feedback (both accurate and inaccurate) on the Four Levels of Feedback Ambition?
As we can see, the answer is more nuanced than simply asking whether AI feedback is accurate and/or efficient. The real issue should be whether students are able to understand, evaluate, and act on that feedback. Teachers do not seek to merely produce feedback efficiently, but to develop learners who can think independently and regulate their own learning.
Part One: Implications of AI Feedback on the Four Levels of Feedback Ambition
In Part 1 of this article, we will discuss two familiar contexts:
- Context 1: When AI Feedback is accurate
- Context 2: When AI Feedback is inaccurate
For each of these, Tan (2026) suggests 3 possible scenarios as shown in Figure 1:
Figure 1: Scenarios with might arise as a result of AI Feedback
Hattie & Timperly (2007) identified four levels of feedback ambition from their meta review of feedback research. Feedback is categorized into four distinct levels:
- Self Level: relates to the personal attributes/self-esteem of the learner and generally does not help the learner to improve in task performance.
- Task Level: focuses on the correctness and quality of the work done, and results in immediate improvement, but might not result in transfer of knowledge to subsequent tasks.
- Process Level: addresses the methods and strategies used to complete the task, and thus enables learners to apply the learning to new tasks.
- Self-Regulated Level: encourages learners to manage their own learning through goal setting, self-assessment and reflection.
Each level serves a different purpose and can significantly impact student learning and motivation when appropriately applied. Further details are shown in Table 1.
What could be the implications of each scenario in Figure 1 on the Four Levels of Feedback Ambition (Table 1)?
Table 1: Four Levels of Feedback Ambition
Context 1: When AI Feedback is Accurate
Every teacher hopes that AI provides accurate feedback for every student. However, our premise is that accurate feedback alone does not guarantee transferable learning. In designing the AI responses, we need to think about how the students would respond to the feedback and the outcomes we want for the students.
Figure 2: Scenarios for feedback recipience when AI Feedback is accurate.
In Scenario 1, feedback is accurate but students are unable to understand the feedback. As such, they cannot use the feedback, and this would likely lead to high anxiety or disengagement. Such feedback, though accurate, does not help the learner and might have a negative impact. Even perfectly accurate feedback has little value if students cannot interpret it. These are some examples of feedback which may lead to anxiety/frustration:
- feedback filled with technical language/unfamiliar terminology
- feedback which is not related to the success criteria
- too much feedback/too many questions asked
In Scenario 2, feedback is accurate and students can understand the feedback but do not know how to use it. From the students’ perspective, utility value of the feedback is low. This has limited impact because students lack the skills to translate the feedback into action. For groups of students like these, the teacher would need to dive in to diagnose learning gaps and provide further scaffolding and/or differentiation to help students make sense of the feedback. For example, many students understand comments such as “strengthen your analysis” or “develop your argument” but still struggle to translate those comments into concrete revisions. In these situations, the teacher needs to provide examples, modelling, learning progression, or more guided practice.
Finally in Scenario 3, students can understand the feedback and act on it immediately. In this scenario, the impact depends on the kind of feedback which AI has been tasked to provide. This can be illustrated in Table 2 below:
Table 2: Implications of Accurate AI feedback on the Four Levels of Feedback Ambition
Hattie and Timperley (2007) argue that effective feedback should help students answer three questions:
- Where am I going?
- How am I doing?
- What should I do next?
Table 2 above explains that even if AI feedback is accurate, it might not always mean the three questions are well answered. On the surface, it might look like AI has the potential to support all three questions. Students comprehend the feedback, know what to change, and successfully improve their work. However, providing answers alone does not guarantee deep learning. The ultimate goal should extend beyond completing the current task. Effective feedback should also help students apply what they have learned in future tasks.
Context 2: When AI Feedback is Inaccurate
Generative AI is impressive but not infallible. Hallucinations, incorrect reasoning and inappropriate suggestions remain possible. Does this mean, however, that all inaccurate AI Feedback has no utility value? The diagram below highlights three additional scenarios.
Figure 3: Scenarios when AI feedback is inaccurate
In Scenario 4, feedback is not accurate and students are not able to detect it. As such, students confidently accept incorrect feedback, reinforcing misconceptions that become harder to correct later.
In Scenario 5, students suspect the feedback may be wrong. Some learners recognise inconsistencies but lack the confidence to challenge the AI. Without teacher guidance, uncertainty may still lead to incorrect learning and this is potentially harmful. However, if teacher intentionally designs a task which facilitates student evaluation of the feedback, it could lead to powerful student agency and deep learning.
Finally, in Scenario 6, students recognise that the feedback is inaccurate. The task should allow them to question the AI response, provide alternative opinions, and seek the right answers. Interestingly, inaccurate AI feedback can sometimes become a learning opportunity. When students evaluate AI critically, verify evidence, seek clarification and justify their reasoning, they develop valuable metacognitive skills and become more independent learners. This is a potentially impactful approach in designing AI feedback.
Part 2: Intentional Design of AI Feedback
Having mapped out the different scenarios when AI feedback is accurate, and inaccurate, we now consider the design implications for feedback practice. The second part of this article will discuss pertinent issues teachers could keep in mind when designing AI Feedback. Three specific recommendations are suggested in this regard:
- Beginning with the end in mind
- Ensuring optimal cognitive effort
- Facilitating specific gap analysis
Beginning with the End in Mind
Tan’s (2022) “four-box” model below describes how teachers’ professional development in assessment literacy can enhance students’ learning.
Tan recommends that teachers begin with the end in mind (Box 4) and prioritize students’ learning by starting with a clear understanding of what students need to learn. This could be one of the following – simply improving task performance, address specific learning gaps, or, if the teacher has a more ambitious goal, to develop students’ evaluative judgement. This clarity would help the teacher to plan for the appropriate learning tasks/activities (Box 3). For example, if the outcome in Box 4 is to develop students’ acumen in judging their learning, a worthwhile activity would be for students to question the AI feedback. In turn, this would help the teacher to shape their feed-up strategies (Box 2) which include articulating the standards and allowing students to engage with the criteria/standards. With these in mind, teachers would be able to identify further knowledge they require (Box 1).
- Ensuring Optimal Cognitive Effort
Faster AI feedback does not necessarily produce better learning if it bypasses students’ cognitive effort. One teacher I worked with shared that we should spend less time asking AI to “mark this essay” and more time designing prompts that stimulate productive thinking. Faster is not always better, and speed alone does not improve learning.
Students would benefit when teachers spend time designing AI prompts that support thinking. For example, instead of asking AI to rewrite a student’s paragraph, teachers might instruct AI to:
- identify one area for improvement without giving the answer
- ask guiding questions
- prompt students to compare their work with success criteria
- encourage students to explain their reasoning before revealing suggestions
In this way, AI becomes a thinking partner rather than an answer generator.
Teachers have also learnt that sometimes the solution is not to hastily improve the AI feedback but changing how students engage with it. This means that the teacher has clarity about what he/she wants students to learn from using the feedback. The table below illustrates how this helps the teacher decide the level of feedback ambition to aim for:
Table 3: How clarity in learning outcomes for students impact the design of feedback
- Facilitating specific gap analysis
Teachers have discovered that AI is often too willing to lend assistance and might provide feedback which is overwhelming for the learner. Hence, teachers need to have clarity about the specific learning gap they want AI to identify, in relation to the success criteria. Having done that, they could provide AI with the relevant prompts. This will ensure that AI will be disciplined and clear about the gap to detect, leading to feedback which students understand and can act upon.
This can be illustrated by the following example (Primary 5 English Language writing exercise – ‘Lost in the Forest’).
Teacher clearly explains the success criteria for the writing task:
- I can use descriptive vocabulary.
- I can organise my ideas into clear paragraphs.
- I can use correct subject–verb agreement.
Example 1
Teacher gives AI the following prompt:
Teachers who tried this have discovered that AI might identify many things for improvement (ie: vocabulary, sentence variety, paragraphing, punctuation, grammar, dialogue, description, character development and plot) and hence produces too much feedback for a Primary 5 pupil.
Example 2
In an improved attempt, the teacher could focus on one specific learning gap, based on the success criteria. The teacher identifies a group of students in the class that has difficulty using descriptive vocabulary. For this group of learners, the teacher can then give AI a much more focused prompt:
With this, the AI feedback would be more focused, manageable and actionable. It also allows the teacher to differentiate the feedback which AI can give to different groups of students. In these examples, effective AI feedback based on specific gap analysis depends on whether the teacher has clearly defined what the learner needs feedback on.
Conclusion
In conclusion, Generative AI offers exciting opportunities to enhance formative assessment, but its real value depends on thoughtful and very intentional pedagogical design. AI should be programmed to support meaningful cognition, analysis, and transfer of learning rather than simply providing quick answers or generating more comments.
The best AI feedback is not necessarily the fastest or the most detailed. Feedback is most impactful when it helps students think, question, revise, and ultimately become independent learners. With well-designed learning experiences, AI can serve as a powerful tool which provides high quality feedback in a personalised manner to suit the different needs of learners within the same class.
References
Hattie, J., & Timperley, H. (2007). The Power of Feedback. Review of Educational Research, 77(1), 81-112.
Koh, G., & Tan, K. H. K., (2024). The Feedback Pedagogy Cycle. Assessment For All Learners. https://assessmentforall.com/the-feedback-pedagogy-cycle/
Sadler, D. R. (1989). Formative assessment and the design of instructional systems, Instructional Science, 18(2), 119–144, 2-s2.0-0039921137, https://doi.org/10.1007/BF00117714.
Tan, K. H. K. (2024). The Four Boxes of Assessment Feedback Literacy Feedback. Assessment for All Learners. https://assessmentforall.blogspot.com/2022/07/the-four-boxes-of-assessment-feedback.html
Tan, K. H. K. (2024). Four vital questions concerning assessment feedback. Assessment for All Learners. https://assessmentforall.com/four-vital-questions-concerning-assessment-feedback/
Tan, K. H. K. (2026). Supporting Learners with Assessment: A Fundamental Approach. Fundamentals of Assessment – Principles and Practices for the Classroom (pp 152-166). Routledge.