Teacher–AI Collaboration to Support Assessment and Feedback: A Case Study in Norwegian Secondary Education

Trust

Several teachers described GenAI as a helpful tool, but emphasized the need to always verify the output before using it with students.

“I don’t fully trust it, no. But it’s a bit like, we shouldn’t blindly trust it either, but actually use it as a tool.” (T1)

This shows a healthy amount of skepticism when introduced to a new technological system. Wanting to verify the contents of generated feedback, teachers also felt more confident using GenAI in areas they knew well, and expressed uncertainty when the content went beyond their expertise, being unable to verify it themselves. This shows that trust in GenAI seems to be conditional.

“I trust it to the extent that it’s things I already know. But if it starts including things I don’t know, then I can’t give that feedback to the students. I have to kind of verify what it says first. If I can’t do that, then I can’t use it either, I think.” (T10)

The sense of responsibility is closely tied to teachers’ professional role and their obligation to justify and explain feedback to students. The need to verify content may, however, become problematic when introducing semi-autonomous systems, as their use presupposes a certain level of trust and a willingness to relinquish some degree of control. Because teachers ultimately remain accountable for the feedback and assessments provided to students, their threshold for trusting GenAI systems may be higher than for less critical applications, such as administrative reporting or real-time student recommendation systems, particularly when the uncertainty of such recommendations is clearly communicated.

Teacher Autonomy and Agency

A strong recurring theme in the interviews was the importance of maintaining professional autonomy and a sense of ownership when using GenAI. Teachers consistently expressed that their role is not only to deliver content but to actively make, validate, and ensure good quality in the feedback they provide to students.

“I’ve heard some people say, just click and send, but I really disagree. You don’t have ownership over it. […] I think it’s important that I look at it and agree and know what feedback I’m sending to my students.” (T20)

This shows teachers’ desire to be involved in the final feedback by making decisions on what content should be included and not. However, it also shows that this is not equally important for all teachers, and that some seem more interested in achieving efficiency than in maintaining control. At the same time, some teachers worried that transferring control to GenAI might undermine the teacher’s role in the learning process, as well as introduce anchoring bias, where the AI’s initial interpretation shapes the teacher’s final judgment. These teachers stressed that GenAI should be seen as a supplement and not a substitute, and emphasized the importance of teacher–student interaction and following the student’s development throughout.

“I’m a bit afraid that if students only receive feedback from the AI, I lose an important part of my role as a teacher regarding the interaction with the students. I’d miss out on the process and maybe just be a product. For us, it’s important to follow the student’s process and be there with them along the way.” (T10)

Teachers view GenAI as a useful assistant but stress the importance of professional autonomy. They emphasize actively evaluating and refining AI-generated feedback to align with their judgment and knowledge of individual students. Autonomy is considered essential for maintaining educational quality, teacher identity, and meaningful teacher–student interactions. Teachers also note that initial GenAI feedback may bias their assessments, reinforcing the need for teacher ownership and agency. While GenAI support is valued, teachers firmly reject the idea of relinquishing some degree of control to AI.

Adaptation, Integration, and Adoption of GenAI

Several teachers expressed that learning to use GenAI tools came with a steep curve and some initial frustration. This was related both to the fact that learning the new GenAI platform came on top of an already busy teacher schedule and that this decision came from management, and not from the teachers themselves. Therefore, the need for collaboration, training, and collective structures was a recurring topic, and teachers suggested allocating time at the start of the school year to experiment with GenAI tools as a team by trying it out in class before coming back and discussing what worked and did not. Others emphasized the importance of actually using the tools to become more confident.

“Regular coursing, and maybe that we have some common program, discuss and share experiences. That we don’t sit by ourselves, because it’s a lot of that around here.” (T19)

As noted earlier, teachers adopt diverse strategies and have varying needs; therefore, a one-size-fits-all approach is unlikely to effectively support teacher–GenAI complementarity. At the same time, several teachers pointed to concrete technical barriers and a lack of platform integration that made adoption harder than necessary. With many teachers being concerned about managing their time effectively, such frustrations might negatively impact the integration and adoption of GenAI tools. Nevertheless, most teachers expressed a fundamentally positive attitude toward the potential of GenAI, especially when framed as a long-term and human-centered shift in education. This openness was often coupled with curiosity and a desire to explore its usefulness, showing that teachers’ resistance to the integration could be more closely linked to a lack of time than to the technology itself.

“I think it will definitely help us a lot in society. Seeing patterns, helping with everyday things. […] If we can use it to help plan teaching, then great. Use it as a model and then adapt. That would be perfect.” (T14)

Overall, the findings indicate that adopting GenAI tools requires time, effort, and institutional support for teachers to gain confidence as users. Navigating multiple platforms can be challenging, particularly for older teachers. The data also reveal a sense of limited agency and a need to better understand and learn how to use GenAI effectively. Teachers emphasize collaboration as essential and value opportunities to share experiences through courses or workshops. Despite these challenges, they remain largely positive about GenAI’s potential and stress that successful adoption depends on the ability to make it part of their professional practice (including the needed training and support).

Timely Feedback

One of the most appreciated benefits of GenAI feedback was its immediacy. Teachers frequently highlighted how difficult and time-consuming it is to help all their students in the classroom, and that GenAI enables students to receive feedback exactly when they need it. The following comments focus on the part of the system where students receive generated feedback directly from GenAI and not the version where teachers review the feedback before giving it to them.

“That students have access to it and can ask Learny for some feedback, instead of me running around. It helps me a lot. […] It saves me time, and they don’t have to sit and wait.” (T20)

Some teachers highlighted how this helps students to improve their work while their ideas are still fresh, rather than waiting days or weeks for teacher comments. Others pointed to how this timely support enhanced students’ independence and learning agency, while also allowing teachers to support students who needed more in-depth help.

“It can be a tool to support them to maybe become more independent and less dependent on asking the teachers about little things all the time.” (T13)

While this might be the perspective of one or more teachers, a critical question to ask is whether students actually become more independent or not. If they use GenAI to handle many of the smaller questions they may have (and questions in general), their dependency might transition partly to AI, making teachers experience their students as more independent. While this might be a trade-off, it will free up time for teachers to support students with greater needs and to answer more cognitively demanding questions.

High-Quality Feedback

Several teachers emphasized that the overall quality of GenAI feedback was surprisingly good and described it as well-structured, clear, and aligned with pedagogical principles. Others were impressed by how specific and rich the feedback was compared to what they would typically have time to write, and reflected on how combining human and GenAI feedback could result in feedback of high quality.

“I think it’s very good at giving feedback that follows what we’ve learned about formative assessment: two or three positives, followed by a suggestion for improvement. […] It’s mostly focused on what the student does well. […] I think it seems very qualitatively good.” (T15)

Some teachers even pointed out that the GenAI often highlighted the same issues they would have addressed themselves, and even phrased things in ways similar to their own:

“It’s like I wrote it myself. […] I have experienced that AI has quite a similar mindset to mine. Sees the same things and stuff. And then it formulates it a bit like I might have said it myself. […] It’s actually kind of amazing how often it hits the mark based on what the assignment asks.” (T19)

This shows how GenAI feedback often identifies the same strengths and areas for improvement that teachers would themselves, making them experience it as both accurate and specific. Some teachers also found that the GenAI tool can provide more extensive feedback than they typically have time to write, showing how a complementarity between teachers and GenAI might be beneficial. In this way, GenAI was, by some teachers, viewed on as a “second teacher,” enhancing the overall quality of feedback given to students.

Time-Saving Potential

Several teachers emphasized how receiving GenAI feedback drafts could save them time. While most teachers wanted to personalize the feedback, they described the drafts as useful starting points that can be adjusted rather than writing the feedback from scratch. Others described how the time-consuming part of assessment was often not grading itself, but formulating meaningful feedback in a way that students could understand. Here, GenAI offered a valuable shortcut.

“I feel like I save some time, it’s a draft for my feedback. Then I go in and adjust and edit it a bit, maybe add something more personal. So I think it does something, I’m probably saving some time on this.” (T11)

Some teachers also stated that this kind of support could reduce the cognitive load of repetitive feedback tasks, freeing up time for other teaching tasks.

“For me, it’s about efficiency. If this tool helps me become more efficient at my job and takes away some work, then I’ll use it, without a doubt. That means I get more time for the things GenAI can’t do, […] like calling parents or resolving conflicts, more interpersonal things.” (T8)

Taken together, these findings indicate that GenAI can meaningfully enhance teachers’ efficiency in feedback processes, especially when used to generate initial drafts. Teachers state that formulating clear and constructive explanations for students is often time-consuming, and that GenAI could free time for other relational and pedagogical tasks that it cannot replace. However, a critical question that remains concerns how teachers work with these drafts. Specifically, it is important to understand whether teachers tend to use the drafts largely as generated, or whether they invest additional time in revising, personalizing, and contextualizing them, and how this balance will evolve as teachers become more experienced with the system.

Inspiration and Scaffolding

Another frequently mentioned quality of GenAI was its role as a cognitive and linguistic support tool. Many teachers described how the GenAI drafts helped them get started more easily when writing feedback. Some teachers also emphasized how this starting point supported reflection and gave them ideas for how to formulate better feedback.

“It’s great to have a little pointer and then build on that yourself. It gets me started much more easily than starting with a blank page.” (T9)

This indicates that GenAI feedback drafts not only have the potential to increase teachers efficiency, but also create opportunities for critical reflection. Although this support was primarily discussed from a teacher perspective, a few teachers also reflected that the tool might help students discover new angles or gain inspiration for what to revise or improve.

“The strong students loved Learny. They got feedback all the time. […] ’Oh, I can write about that!’ So it’s a good starter.” (T21)

In this way, GenAI can also be viewed as a source of inspiration and scaffolding for students. By generating initial feedback drafts, it offers ideas that both teachers and students might take advantage of, in addition to encouraging reflection and improvement in teachers feedback practices.

Motivation

Several teachers noted that GenAI feedback can enhance student motivation by aligning with formative assessment principles. Along with emphasizing shortcomings, it highlights strengths and offers constructive suggestions for improvement. This positive framing is seen as especially beneficial for students with lower self-confidence and self-esteem.

“And then there’s both what you’re going to do to improve it, and what you’ve done that’s good that you have to continue with. Because that’s also just as important, I think, especially for someone who may not be that super strong academically.” (T12)

The uplifting tone and availability of GenAI to give feedback were also viewed as a strength in itself, making students feel seen and encouraged. Some teachers pointed to how this positivity could reinforce a sense of competence and success.

“There’s something good here, someone actually bothers to praise the students, it is perfect!” (T16)

Overall, teachers believe GenAI feedback enhances student motivation. Its positive tone and accessibility make students feel recognized and supported—especially those who may struggle with self-confidence. Teachers therefore note that GenAI serves as a kind of personal motivator, providing encouragement they may not always have time to give. They also observe that this support can improve students’ work quality, fostering a greater sense of achievement.

Challenges in Feedback Quality

While several teachers praised the overall structure and usefulness of GenAI-generated feedback, others raised concerns about variations in clarity, relevance, and accuracy. The perceived quality of the feedback often depended on the specific assignment, the individual student, and the teacher’s expectations. Moreover, the language used by GenAI was too advanced or unnecessarily complicated for many students.

“It’s an okay draft, but I have to be honest, I wouldn’t press send right away, it remains quite much work. Firstly, there’s a lot of feedback on quite little work. […]. I can’t always understand how the student is supposed to use that feedback well. You manage to pick out some things that are correct and such, but there’s a lot of text, there’s difficult language, and there’s some imprecise wording.” (T5)

This finding indicates that some teachers experience GenAI feedback drafts as misaligned with task goals or student needs, requiring substantial revision before they can be shared with students. Some teachers also pointed out how feedback could be too vague, long, repetitive, or misleadingly positive.

“A weakness can be that it is very positive, maybe too positive. I think it can trick the students a little into thinking that they’ve done better than they actually have.” (T2)

Overall, teachers noticed several shortcomings when assessing the quality of the GenAI feedback, such as vague or generic content, unnecessarily difficult language, and misleadingly positive phrases. These qualities of generated feedback can make it less approachable and actionable for students, making teacher revisions necessary.

Lack of Relational and Contextual Awareness

One of the most frequently mentioned limitations of GenAI feedback was its lack of relational and contextual awareness. Teachers described how GenAI did not possess knowledge about their students and their individual needs, and feared that feedback, therefore, could feel distant or depersonalized if they did not contribute to it.

“I have to adjust it so that the student recognizes me and what we have worked on. […] You have to write a little to make it a little more personal, so that the students do not feel that it is just a machine that is assessing their work and such.” (T19)

Others emphasized that meaningful feedback must motivate students and support them in reaching long-term goals and underscored how their knowledge about their students was necessary to achieve this.

“Learny doesn’t know the student the way I do. […] When it comes to motivation, only I know how to motivate the student based on the long-term goal I’ve set. […] We won’t reach it unless the feedback is targeted.” (T4)

Overall, teachers underscored the importance of having relational and contextual awareness, including familiarity with the student’s learning history, emotional state, language proficiency, motivation, and personal goals. Furthermore, teachers have a good understanding of students’ cognitive and metacognitive capabilities and use this information when giving feedback. Therefore, GenAI’s lack of this awareness can cause situations where feedback is not appropriate, and teachers emphasize that effective feedback is not only about correcting content but also delivering it in a way that is supportive to the individual student. Because students might experience GenAI feedback as distant and depersonalized, they highlighted the need to adjust feedback to make it more personal and human-like.

Limited Understanding of Students’ State

Teachers were also concerned about GenAI’s limited understanding of students’ states, and that GenAI feedback might confuse, mislead, or even stress students. One common issue was the lack of consistency and connection across repeated feedback generations, resulting in feedback containing different comments on the same work even when nothing had changed, which caused confusion and uncertainty.

“I think you will get students who will ask for feedback very often. And as I understood, it gives different feedback, regardless of whether a single improvement has been made or not. That you just stand in the same place and press feedback, and then you get five different feedback on the same work. […] That will only confuse the student.” (T2)

This finding indicates that teachers worry the endless stream of suggestions might make students feel their work is never good enough. Teachers reported that some students, especially high-achieving ones, kept requesting new feedback and became overwhelmed. Others highlighted the risk of students misinterpreting overly positive feedback as an indicator of high academic achievement, which could cause confusion later in the formal assessment process.

“It’s never good enough. […] It doesn’t work, it’s awful for the high-performing kids who are never satisfied, and they keep getting told that you can do this and that better. […] They ask more and more questions, and in the end, they are almost outside of the task. […] They need to learn to say: ‘What I did is good enough. Now I can stop’.” (T21)

Therefore, teachers emphasize that GenAI’ limited understanding of the student’s current state was problematic, where inconsistencies in generated feedback could confuse students, an overload of suggestions may lead to more stress among high-performing students, and overly positive feedback can give students a false impression of their academic level.

Teachers’ Lack of Insight Into GenAI Feedback Students Receive

Several teachers reflected on the implications of not knowing the content of feedback provided to students by GenAI (when using the immediate feedback version). One teacher described the discomfort of being responsible for evaluations that were influenced by unseen GenAI feedback, while others noted how this made it difficult to assess students’ independent capabilities.

“The ethical part of it I find difficult. One thing is that the responsibility lies with the school owner, but I am the one who is assigned the task of doing it. […] Let’s say that a student gets a bad grade. Then the student, and maybe parents, can come back later and say ‘Yeah, but I did what the feedback told me’.” (T14)

This raises a larger question about teacher responsibility when integrating technological tools into education. Some teachers also pointed out that conflicting feedback from GenAI and the teacher could lead to confusion about whom to trust, and make it difficult for teachers to build on or align with the GenAI input.

“It can be a challenge for the teacher to continue that guidance afterwards, if you don’t know what the student has been guided on beforehand. Because then I think the student will probably be a bit unsure too, should I listen to what the teacher gives in feedback, or should I listen to Learny? […] at least if there is a bit of a discrepancy […] It can actually hinder learning.” (T15)

In more serious cases, teachers raised concerns about the potential for harmful or inappropriate suggestions in generated feedback, especially in sensitive subjects such as religion or politics:

“Let’s say that for some strange reason, the student has received feedback about something that is ethically unacceptable. It could be a political message, it could be things that are not within the freedom of speech, and then, in a way, the student is stimulated through AI to just continue [working on] their product. […] What kind of control do we have on the quality [of feedback] in these cases?” (T10)

Taken together, teachers experience discomfort in not knowing what kind of feedback students receive. Because teachers are ultimately responsible for evaluating and guiding their students, they worry about being questioned about assessments they cannot fully justify. Teachers express that unseen GenAI feedback makes it harder to evaluate students’ work and that conflicting advice from teachers and GenAI can confuse students. Moreover, unfiltered or ethically problematic feedback on sensitive subjects poses serious concerns about the appropriateness and integrity of the learning process. All these concerns about the automated feedback highlight the potential remedial role the teacher can have by incorporating an HI approach.

Challenges With Student Self-Regulation

Several teachers voiced concern about students’ ability to self-regulate their use of GenAI feedback tools. In contrast to traditional classroom dynamics, where teachers control the timing, amount, and frequency of feedback, the GenAI system allows students to generate feedback whenever and however much they want. While this can promote greater independence and autonomy, it also places increased responsibility on students to manage their learning process effectively.

“There should be a way to limit how often they can press ‘Help me’ […] It should be a maximum of 2-3 times. […] Feedback should be available after working for at least an hour with a task, not just five minutes. […] If you want to train them to think critically, you can’t give them too much help.” (T18)

Others noted that frequent feedback requests could reduce opportunities for students to work independently, reflect, and develop endurance over time, making them dependent on constant input.

“I’m not sure, in the long run, if students learn to have perseverance in their work or not. That you can then kind of get feedback all the time about what needs to happen. Because there is something in learning to keep working on a task over time, alone, without getting feedback. Which is also valuable.” (T14)

Overall, teachers expressed ambivalence about giving students full autonomy in how they engage with GenAI tools. On one hand, they saw potential for these tools to strengthen students’ independence and ownership of learning; on the other, they worried that too much freedom could lead to over-reliance or superficial engagement. Some described how students struggled to regulate their use of GenAI and tended to seek feedback repeatedly rather than reflecting on their work. These concerns reveal a tension between promoting agency and ensuring productive regulation, highlighting the need for scaffolds that help students use GenAI feedback purposefully and critically.

Risk of Cognitive Laziness

Teachers expressed concern that while GenAI provides valuable support, it may also reduce the critical engagement and cognitive effort of both students and teachers. Although the tool can scaffold learning and streamline feedback, several teachers feared that excessive reliance on it could lead to passivity and diminish reflection and critical thinking.

“If someone always tells you how to move forward, then you’ve lost a part of your critical sense on what you are doing.” (T18)

For students, this risk was associated with over-scaffolding, receiving too much ready-made guidance that could weaken their ability to think independently. For teachers, the concern was that habitual use of GenAI would make them stop thinking for themselves, such that GenAI would replace teachers’ professional judgment and creativity in feedback practices.

“We have to be careful not to fall into the trap of thinking that this was easy […] thinking that we can just press a button and then our job is done and that we don’t need the teacher anymore.” (T21)

Overall, teachers emphasized that while GenAI can enhance efficiency, it should not replace human reflection or professional responsibility. Maintaining teacher oversight and pedagogical involvement was seen as essential not only for ensuring the accuracy and appropriateness of feedback but also for preserving cognitive engagement and agency in the feedback process.

Design Implications for HI-Aligned Feedback Systems

In this study, design principles refer to actionable sociotechnical configurations that shape how responsibility, decision-making, and expertise are distributed between teachers and GenAI across the feedback workflow. Rather than evaluating GenAI performance in isolation, these principles foreground complementarity—how GenAI augments, supports, or temporarily substitutes specific instructional tasks while preserving teachers’ professional judgment. For example, context anchoring through teacher rubrics and examples ensures that GenAI outputs reflect pedagogical intent, while configurable learner-level settings and transparency dashboards make GenAI behavior interpretable and adjustable within real classroom conditions. In this way, HI becomes an analytical lens for understanding how automation reshapes teacher cognition and decision-making, highlighting that effective collaboration depends less on technological capability alone and more on how systems are designed to sustain human agency and instructional reasoning.

Aligned with emerging research priorities, these principles also outline concrete directions for the development of educational AI systems. First, HI-aligned tools require AI affordances that explicitly support educational complementarity, including transparent models, editable feedback histories, and uncertainty signaling that allow teachers to calibrate trust without relinquishing control. Second, effective systems must be co-designed with stakeholders, embedding teachers’, students’, and school leaders’ perspectives into system functionality so that GenAI tools reflect the realities and constraints of classroom practice rather than functioning as external technical add-ons. Third, the principles highlight the need for professional learning models for hybrid systems that move beyond general AI literacy toward pedagogical reasoning with AI-generated data—supporting teachers in interpreting feedback analytics, monitoring student regulation, and making instructional decisions grounded in learning theory. Finally, the design implications acknowledge broader shifts in professional roles and identity: as GenAI redistributes cognitive and creative labor, systems must be intentionally designed to reinforce teachers’ roles as pedagogical orchestrators rather than passive overseers. Taken together, these implications position HI not merely as a conceptual framing but as a design-oriented research agenda in which educational AI evolves through iterative human–AI co-learning, sustained teacher involvement, and carefully balanced automation that enhances efficiency without undermining professional responsibility.

Teachers’ Professional Practice

The integration of GenAI into feedback practices presents both opportunities and challenges for supporting teachers’ formative assessment and feedback. Our findings indicate the importance of GenAI not being a replacement for teachers’ feedback; at the same time, it highlights the benefits of responsibly incorporating it as a supplementary tool that requires active engagement, adaptation, and pedagogical oversight. Our study highlights the importance of empowering teachers to maintain their professional autonomy and the responsibility to critically assess and revise AI-generated feedback before sharing it with students. This underscores the importance of protecting and supporting teacher agency when introducing GenAI into educational contexts. This includes giving them adequate time and space to experiment with GenAI, reflect on its role, and learn when and how it can best support student learning. Additionally, professional development should not only focus on giving teachers technical and practical skills, but also on increasing AI literacy and agency, as teachers need the ethical, pedagogical, and critical understanding of GenAI’s role. To facilitate meaningful and responsible use of GenAI, both pre-service and in-service teacher education should therefore include structured training and exposure to GenAI tools, which will provide teachers with safe opportunities to experiment, develop the necessary competencies, and integrate such tools into their practice.

School Development and Professional Learning

The findings also highlight a strong need for structured collaboration and shared professional learning around the integration of GenAI tools. Teachers expressed a desire for dedicated time to try out tools together, reflect as a team, and share both challenges and best practices. School leaders and decision-makers should therefore consider facilitating this through common and safe spaces (e.g., workshops), collaborative pilot projects, and integrating reflection around GenAI into existing professional development structures. This is essential to prevent isolated practice and uneven adoption, as noted as a concern in the interviews in this study. Furthermore, schools and school governing bodies (municipalities in our case) should be aware of the additional time demands placed on teachers during the early phases of GenAI adoption and allocate resources accordingly. When GenAI tools are introduced without support or structure, the risk increases that teachers will either reject the technology or use it uncritically.

Technology Design and Implementation

The challenges identified in this study point to several directions for improving the design of GenAI feedback tools. First, GenAI systems must be flexible enough to accommodate diverse subjects, educational goals, and classroom contexts. In this study, GenAI was used to support writing, language learning, and subject-specific feedback in areas such as math and science. These varied applications highlight the need for either broadly capable models or smaller, fine-tuned systems tailored to educational domains. Future systems should avoid “one-size-fits-all” approaches and integrate these insights by fine-tuning models based on teacher edits or enabling customizable feedback profiles by subject, level, or tone. Moreover, teachers expressed concern about the lack of transparency and explainability regarding the feedback students received independently. Improving the traceability, transparency and explainability of GenAI feedback-through the use of proper dashboards, logs, or interfaces that show why particular suggestions were made-would strengthen teacher oversight and trust. Integrating explainable AI (XAI) techniques, such as SHAP or LIME, could also help visualize GenAI reasoning and enhance teacher–AI collaboration.

Further, teachers emphasized the need to manage students’ use of GenAI to prevent over-reliance. Feedback systems should incorporate mechanisms that encourage reflection and responsible use, such as limits on feedback iterations or reflective prompts between revisions. Involving teachers in the design process, via participatory design approaches, is key to ensuring classroom relevance, pedagogical soundness, and sustainable use. Finally, feedback quality can be enhanced through improved prompting and model refinement. Research shows that prompts specifying role, tone, and feedback style can substantially improve output (Steiss et al., 2024). Leveraging teacher-edited feedback as training data and incorporating student feedback ratings on clarity and usefulness could create adaptive systems that better emulate teachers’ pedagogical standards.