Designing responsible AI feedback in higher education: faculty readiness, student judgement, and metacognitive calibration

1 Introduction: why AI feedback requires more than technical readiness

AI is no longer something higher education can postpone. It is already part of everyday academic work. Lecturers use it to prepare classes, support assessment, guide students’ writing, explain concepts, create examples, and provide feedback. This is not surprising. Classes are getting larger, students’ needs are becoming more varied, and academic workloads continue to increase (Zawacki-Richter et al., 2019; Holmes et al., 2023; UNESCO, 2023).

In this article, AI feedback refers mainly to feedback generated or supported by generative AI systems. This includes comments, explanations, suggestions, alternative wording, and guidance produced in response to students’ work. The argument may also have relevance for other forms of AI-supported feedback, but the focus here is not on rule-based automated correction, predictive dashboards, or adaptive tutoring systems.

Still, AI feedback is not automatically helpful. A tool may give detailed comments, but the student may still not know what to change or why. Another student may use AI to polish a draft without really understanding the ideas behind it. In these cases, feedback can make learning look better than it is. Feedback supports learning only when students understand it, judge whether it is relevant, and use it to decide what to do next (Nicol and Macfarlane-Dick, 2006; Carless and Boud, 2018; Zhang and Hyland, 2022).

AI feedback, then, is not good or bad in itself. What matters is how it is used. If it is built carefully into a course, it can help students test ideas, return to their work, and ask for support before they submit an assignment. If it is poorly structured, it can encourage the opposite: over-reliance on the tool, quick fixes, or acceptance of suggestions that students have not really understood. This matters in higher education because feedback is closely tied to assessment quality, integrity, authenticity, authorship, and plagiarism (Dawson, 2020; Perkins, 2023; UNESCO, 2023).

Lecturers, therefore, need to be ready for AI, but readiness cannot mean only knowing how to use the tool. In this article, faculty readiness refers to a set of connected capacities: technical understanding of AI tools, pedagogical design capacity, feedback literacy, faculty evaluative judgement, ethical awareness, assessment competence, and knowledge of institutional policy. Prompt writing and institutional rules are useful, but they do not answer the main teaching question: how should AI feedback be built into learning? Lecturers need to decide when AI feedback is appropriate, what limits are needed, what students should do before they receive it, what evidence of learning should be made visible, and what students should do afterwards (Mishra and Koehler, 2006; Long and Magerko, 2020; Holmes et al., 2023; UNESCO, 2023).

This Perspective article argues that responsible AI feedback should help students judge their own understanding more accurately. The argument draws on self-regulated learning and metacognition, but applies them to the particular challenge of AI feedback. Metacognitive calibration refers to the relationship between students’ own judgements, such as confidence or perceived understanding, and independently evaluated evidence of performance. AI feedback may support this process by encouraging students to compare, question, and revise their work. However, it cannot be treated as the final standard for calibration, because it may be incomplete, inaccurate, or not fully aligned with course expectations (Zimmerman, 2000; Schunk and Zimmerman, 2008; Panadero, 2017).

The contribution of this article is not to claim that feedback interpretation, self-assessment, or revision are new. Instead, the article explains why these processes need to be adapted when feedback comes from AI. Unlike conventional feedback sources, whose educational authority is usually clearer within the institution, generative AI can produce feedback that sounds plausible but may still be inaccurate, irrelevant, or weakly aligned with disciplinary expectations. Students therefore need evaluative judgement not only to use feedback, but also to judge the feedback itself. At the same time, calibration needs to be assessed separately, using independently evaluated evidence of performance.

2 From feedback access to metacognitive calibration

Digital tools are now a normal part of higher education. Polls, quizzes, dashboards, discussion platforms, and student response systems help lecturers see some aspects of what is happening in class. Students can answer questions, vote, submit responses, post comments, and receive feedback. For lecturers, this provides a useful snapshot of how students are responding to a task, question, or concept (Kay and LeSage, 2009; Hung, 2017; Altwijri et al., 2022; Shahba et al., 2023). In this section, digital engagement is used in a limited sense to refer to visible participation traces, such as responses, submissions, comments, and feedback receipt. It is not used to capture the full behavioural, cognitive, emotional, or agentic dimensions of engagement.

However, visibility does not necessarily mean learning. A student may choose the correct answer but still be unable to explain why it is correct. Another student may take part in a discussion while still holding misconceptions. Digital feedback supports learning only when students engage with it, compare it with their own ideas, and use it to guide their next steps (Nicol and Macfarlane-Dick, 2006; Carless and Boud, 2018; Zhang and Hyland, 2022).

This concern is not new. Classroom response systems can help more students take part and can make their answers visible. But that does not mean deeper learning has happened. If the lecturer shows the correct answer and then moves on, the activity has limited value. The more useful step is to ask why students chose that answer, how confident they were, and what the range of responses suggests about their understanding (Kay and LeSage, 2009; Hung, 2017; Shahba et al., 2023).

AI feedback brings the same issue into a new setting. A tool can respond in seconds with comments, explanations, examples, alternative wording, and suggestions. This can be useful, especially when students need support, and lecturers cannot respond to every draft or question. But quick feedback is not necessarily good feedback. Students may receive several comments from AI and still be unsure which ones matter, which ones they can trust, or how they relate to the assessment criteria (Shi and Aryadoust, 2024; Yildiz Durak and Onan, 2025; Guo et al., 2025). What changes with generative AI is not only the speed or amount of feedback. It is also the uncertainty of its authority. AI feedback can sound clear and confident while still being inaccurate, generic, or poorly aligned with the task. For this reason, students need to evaluate the feedback itself rather than simply act on it. This requires evaluative judgement: the ability to make informed judgements about the quality of work and, in this context, about the credibility, relevance, and usefulness of the feedback provided. Evaluative judgement helps students decide whether AI feedback should be accepted, questioned, rejected, or investigated further. It is closely related to feedback literacy and metacognitive calibration, but it is not the same. Feedback literacy concerns how students understand and use feedback, while metacognitive calibration concerns how accurately students’ self-judgements align with independently evaluated evidence of performance.

Students also respond to AI feedback in different ways. Some appreciate it because it is quick, available at any time, and less intimidating than asking a lecturer. Others are more cautious and question whether the feedback is reliable. The problem is that even weak or general feedback can sound convincing. Students may then accept suggestions without checking whether they are accurate or relevant. This is an academic integrity concern, but it is also a learning concern. When students use AI feedback mainly to avoid making decisions, feedback stops doing its educational work (Nazaretsky et al., 2024; Perkins, 2023; Venter et al., 2025).

The shift needed is from simple access to feedback toward calibrated judgement. Access asks whether students received feedback or used a tool. Calibration asks whether students’ own judgements become more closely aligned with independently evaluated evidence of performance. In AI-supported feedback contexts, students may use AI feedback, assessment criteria, and lecturer or peer comments to interpret their work, weigh possible changes, and decide how to revise. These sources can support evaluative judgement and revision, but they should not be confused with the evidence used to assess calibration accuracy. Calibration requires students’ confidence or self-judgement to be compared with independently evaluated performance, such as lecturer-assessed work, validated answers, later unaided performance, or another defensible criterion. Students then need to return to their earlier judgement after feedback and revision, and consider whether later performance shows more accurate self-evaluation over time (Bjork et al., 2013; Carless and Boud, 2018; Mollick and Mollick, 2023).

3 Faculty readiness as pedagogical judgement

Knowing how to use an AI tool, write an effective prompt, or explain what AI is does not, on its own, prepare lecturers to use AI well in education. The broader point is that technology integration depends on pedagogical and disciplinary judgement, not only technical skill. Lecturers also need to know how to design learning around AI, guide students in using AI feedback appropriately, protect academic integrity, and decide when AI should or should not be part of a learning task (Mishra and Koehler, 2006; Long and Magerko, 2020; Holmes et al., 2023; UNESCO, 2023).

This is especially important for AI feedback. Its value does not come only from being fast, detailed, or available. It depends on what students are expected to do with the feedback. If they use AI only to correct errors, improve wording, or polish a submission, the learning process may remain quite limited. Lecturers, therefore, need to design AI-supported feedback in ways that keep students involved in judging, revising, and explaining their own work (Nicol and Macfarlane-Dick, 2006; Carless and Boud, 2018; Zhang and Hyland, 2022).

Faculty readiness, in this context, includes several connected dimensions. Lecturers need technical understanding of how AI feedback tools work and where their limitations are. They also need pedagogical design capacity, so that AI feedback is placed at the right point in the learning process. They need feedback literacy, because students must be helped to interpret and use feedback rather than simply receive it. They need evaluative judgement so that they can assess the quality, credibility, and disciplinary appropriateness of AI-generated feedback and determine whether it aligns with assessment criteria and intended learning outcomes. They also need ethical awareness, assessment competence, and knowledge of institutional policy, because AI feedback raises questions about transparency, accountability, fairness, authorship, and academic integrity (UNESCO, 2024; Kassorla et al., 2024; Burneo-Arteaga et al., 2025).

These forms of readiness are connected, but they are not the same. Pedagogically, lecturers need to decide what students should do before and after receiving AI feedback, and what the feedback should help reveal about their thinking. Pedagogically and ethically, lecturers need to develop students’ evaluative judgement so that they can question feedback that may sound confident but may not be credible, fair, relevant, or aligned with disciplinary expectations. They also need to build students’ feedback literacy, so that students learn to interpret feedback, question it, and decide what to do next.

In the proposed cycle, these dimensions of readiness are practical. They shape the choices lecturers make when designing AI-supported feedback: what students produce before using AI, what kind of feedback is useful, what evidence of learning should be collected, when human judgement is needed, how students should explain their use of feedback, and how responsibility should be recorded. The cycle may focus on students’ actions, but it depends on lecturers’ design decisions. Students take the learning steps; lecturers create the conditions that make those steps meaningful.

Taken together, these dimensions show that faculty readiness is not mainly technical. It is a form of pedagogical judgement: knowing how to design AI feedback so that students reflect on their work, revise it, justify their decisions, and remain accountable for their learning. This judgement is also institutional, because AI feedback affects assessment design, academic integrity, student responsibility, and university policy (Holmes et al., 2023; UNESCO, 2023; Dotan et al., 2024).

4 A responsible AI feedback design cycle

AI feedback is most useful when it is part of the learning process, not something added at the end to tidy up students’ work. The cycle proposed here draws on formative feedback, feedback literacy, evaluative judgement, and self-regulated learning, but adapts these ideas to the specific conditions of generative AI feedback: uncertain epistemic authority, the possibility of product improvement without deeper learning, the need to evaluate rather than simply use feedback, and the need to preserve human responsibility for decisions made after feedback (Nicol and Macfarlane-Dick, 2006; Carless and Boud, 2018; Mollick and Mollick, 2023).

Table 1 operationalises the cycle by linking each stage to student action, lecturer role, evidence, criterion, and intended outcome.

Table 1. Operationalising the responsible AI feedback design cycle.

The cycle presented in Figure 1 follows the same seven stages operationalised in Table 1: initial judgement, initial response, confidence check, AI feedback as one source, human mediation, revision, documentation, and responsibility, and recalibration over time.

The cycle begins by asking students to consider how they understand the task, the criteria, and the knowledge or skills being assessed. This is different from the later confidence check, where they judge the quality or completeness of the work they have produced. Calibration depends on comparing these self-judgements with independently evaluated evidence of performance, not with AI feedback itself. AI feedback and other comments may inform evaluative judgement and revision, whereas calibration concerns whether students’ self-judgements correspond more accurately with what independently evaluated performance shows (Zimmerman, 2000; Panadero, 2017; Bjork et al., 2013; Dunlosky and Rawson, 2012).

In many assessment contexts, students should first produce a response before receiving AI feedback, especially when lecturers need to see what students understand on their own. In other contexts, AI may reasonably come earlier in the process, for example, by offering examples, guiding questions, accessibility support, or initial scaffolding. The important decision is where AI feedback fits in relation to the learning objective, students’ level of expertise, and the stakes of the assessment (Dawson, 2020; Bearman et al., 2020; Perkins, 2023).

AI feedback should then be treated as one source of information, not as the final judgement. Its role in the cycle is therefore to support evaluative judgement and revision, not to provide the independent criterion against which calibration is determined. Students need to decide what to accept, question, reject, or investigate further, using the task goals, assessment criteria, disciplinary standards, and lecturer guidance. Human mediation becomes important when feedback conflicts with the criteria, when the task is high stakes, when feedback appears generic or inaccurate, when students cannot justify their choices, or when disciplinary judgement is needed (Carless and Boud, 2018; Zhang and Hyland, 2022; Guo et al., 2025; UNESCO, 2023; Holmes et al., 2023; Dotan et al., 2024).

The next stage is revision, documentation, and responsibility. Students need to explain what they changed, why they changed it, how AI feedback shaped their decisions, and what remains their own responsibility. The final stage is recalibration over time. Students revisit their earlier confidence and self-judgement after feedback and revision, and compare subsequent judgements with independently evaluated performance evidence, such as later assessed work, an unaided task, or transfer performance. Recalibration is therefore demonstrated not by agreement with AI feedback, but by closer correspondence between later self-judgements and what students subsequently demonstrate. Whether the proposed cycle improves this correspondence remains a conceptual claim requiring empirical testing.

5 Discussion: implications for academic integrity, policy, and faculty development

The use of AI feedback changes how academic integrity needs to be understood. It is not enough to focus only on detecting misuse after submission. Detection tools, restrictions, and penalties may still be necessary, but they do not address the deeper issue: students may use AI feedback in ways that improve the final product without strengthening their own judgement. A more educationally useful approach is to design assignments that require students to show how they worked, what feedback they received, what decisions they made, and how the work changed before submission (Dawson, 2020; Bearman et al., 2020; Perkins, 2023).

For this reason, academic integrity should include more than declaring whether AI was used. Students should also be able to explain how they used feedback, why they accepted or rejected particular suggestions, and what responsibility they took for the final version. Feedback should help students become better judges of their own work; it should not take that judgement away from them (Carless and Boud, 2018; Zhang and Hyland, 2022; UNESCO, 2023).

At the same time, this article focuses on the responsible pedagogical use of AI feedback, not on the full responsible-AI agenda. Universities still need to address wider issues such as data privacy, platform governance, bias, accessibility, equity of access, intellectual property, transparency, and the possible use of student work by commercial systems. These issues matter because even a well-designed learning activity can become problematic if the wider conditions for responsible AI use are not in place.

This work cannot depend on individual lecturers working things out on their own. Universities need policies that do more than state what is allowed and what is not. Institutional policy needs to connect AI feedback with assessment integrity, student transparency, and lecturer judgement. It should help lecturers design tasks where students’ thinking can be seen, help students acknowledge and explain how they used AI, and protect the lecturer’s role in academic judgement (Dawson, 2020; Holmes et al., 2023; UNESCO, 2023; Dotan et al., 2024; Kassorla et al., 2024; UNESCO, 2024).

Faculty development also needs to move beyond short workshops on tools. Sessions on prompts, feedback generation, or platform use can be useful, but they are not enough. Lecturers need time to discuss what is happening in their own classes, review assessment tasks, compare examples, and work through difficult cases with colleagues. They also need space to think about how AI feedback works in their own disciplines, because the issues will not be the same in every subject. Communities of practice, assessment clinics, peer review of AI-integrated tasks, and discipline-specific workshops can help lecturers develop approaches that are more confident, consistent, and grounded in their teaching contexts (Carless and Boud, 2018; Dawson, 2020; Dotan et al., 2024; Kassorla et al., 2024; Burneo-Arteaga et al., 2025).

Human mediation also needs to be realistic. It does not mean that lecturers should review every interaction between students and AI, especially in large classes. Instead, mediation can be designed through checkpoints, sample reviews, peer discussion, short reflection prompts, annotated decision logs, or targeted support when students cannot explain how they used the feedback. This keeps the lecturer judgement in the process without making AI-supported feedback impossible to manage.

Looking ahead, the question is no longer simply whether AI feedback should be allowed. In many higher education contexts, students are already using it, with or without formal guidance. The more urgent issue is how universities can make this use educationally meaningful. Future work should therefore examine assessment designs that require students to show their thinking, not only submit a final product. More specifically, future studies should test whether a structured AI-feedback cycle helps students make better judgments about their own work. This could include examining whether initial responses, confidence checks, feedback records, and revision explanations improve confidence–performance calibration based on independently evaluated performance evidence, feedback uptake, depth of revision, error detection, later unaided performance, retention, transfer, student agency, and lecturer workload. Studies could also compare structured AI-feedback activities with more open-ended AI use to examine whether the cycle reduces uncritical acceptance of AI suggestions, strengthens evaluative judgement, and improves the accuracy of students’ self-judgements relative to independently evaluated performance over time.

6 Conclusion: faculty readiness as pedagogical judgement

This article has argued that AI feedback should not be treated as another technical add-on to assessment. What matters is the learning process built around it. Students need chances to show their thinking before they receive feedback, while they are working with it, and after they revise. Used in this way, AI feedback can help students compare ideas, improve their work, explain their decisions, and remain responsible for what they submit. It can also support metacognitive calibration, but only when students compare their confidence and self-judgement with independently evaluated evidence of performance rather than treating AI feedback as the standard against which their judgement is evaluated.

Faculty readiness, therefore, is not simply a technical skill. Lecturers need to make careful decisions about when AI feedback is useful, how students should work with it, what evidence of learning should be visible, and where human judgement is still needed. They also need ethical awareness, feedback literacy, and assessment design skills, so that AI feedback supports learning rather than just helping students produce a more polished piece of work.

The responsible AI feedback design cycle is proposed as a conceptual framework. It now needs to be tested in real teaching contexts, across different disciplines, assessment types, and levels of student expertise. Future studies should examine whether structured AI-feedback designs improve the correspondence between students’ self-judgements and independently evaluated performance.

AI feedback should not take the thinking out of learning. Its value lies in helping students look more carefully at their ideas, improve their work, and take responsibility for their choices. The task for higher education is not just to introduce AI feedback, but to design and guide it in ways that protect its educational purpose.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI was used to support the preparation of this manuscript. Specifically, it was used to assist with language refinement, clarity, grammar, sentence structure, and overall readability. The authors reviewed, edited, and approved all AI-assisted revisions and take full responsibility for the final content, arguments, citations, and conclusions presented in the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Mohammed Ghazal, Abu Dhabi University, United Arab Emirates

Reviewed by: Arash Javadinejad, Catholic University of Valencia San Vicente Mártir, Spain