I’ve been arguing for years that AI belongs in classrooms. I’ve also been warning, just as consistently, that pedagogy determines whether it helps or hurts. A new paper by Corbin, Tai, and Flenady (2025) gives that warning some serious philosophical teeth, specifically around something most of us take for granted: feedback.
Their argument is direct and, I think, important. GenAI systems can generate feedback messages. They can identify grammatical errors, flag weak arguments, suggest revisions. What they cannot do is recognize the student. And recognition, the authors argue, is what makes feedback actually work.
Why AI Feedback in Education Needs a Relational Lens
The feedback literature has moved well past the idea that feedback is just information delivery. Contemporary research frames feedback as a process, something that happens when a learner makes sense of information and uses it to improve (Henderson, Ryan, and Phillips, 2019). But Corbin et al. point out something the literature has mostly left implicit: that process depends on a relationship of mutual recognition between teacher and student.
They draw on two post-Hegelian philosophers to make this case. Axel Honneth’s Struggle for Recognition (1996) frames respect and identity as products of mutual acknowledgment between people. Robert Brandom’s A Spirit of Trust (2019) roots trust in a structure where both parties hold each other accountable to shared norms. The convergence is striking. Recognition isn’t a bonus feature of good teaching. It’s the condition under which feedback becomes formative.
When a teacher provides feedback, both parties expose themselves. The student submits work that reflects their developing understanding. The teacher offers a judgment that reflects their expertise and professional identity. Both are vulnerable. That vulnerability, Corbin et al. argue, “is not a weakness to be overcome but rather a necessary condition for the establishment of trusting and respectful relationships and, consequently, for effective feedback” (p. 726).
I’ve covered related dynamics before. Luo (2025) found that GenAI use was already eroding trust in teacher-student relationships, with students afraid to seek help for fear of being suspected of cheating. Corbin et al. approach the same tension from a different angle: it isn’t just that AI disrupts trust. It’s that AI structurally cannot participate in the kind of trust that feedback requires.
Recognitive vs. Extra-Recognitive: The GenAI Feedback Framework
The paper proposes two categories. “Recognitive feedback” happens between agents capable of mutual recognition, where both parties acknowledge each other’s status, vulnerability, and authority. “Extra-recognitive feedback” comes from sources that cannot genuinely recognize the student as a developing scholar. GenAI fits squarely in the second category.
Brandom’s concept of “normative attitudes” and “normative statuses” does useful work here. A student claiming expertise in a discipline implicitly commits to certain norms: engaging with the literature, meeting deadlines, being open to critique. A teacher claiming authority commits to staying current, mentoring carefully, evaluating evidence with rigor. Feedback is the ongoing process of holding each other accountable to those commitments. GenAI can simulate that accountability, but it can’t genuinely participate in it.
Corbin et al. describe extra-recognitive feedback as “primarily a unidirectional transmission of information, lacking the reciprocal nature that characterizes recognition-capable feedback” (p. 726). That description tracks with what Shaw and Nave (2026) have been calling cognitive surrender, the tendency to let AI handle the thinking. If students receive AI feedback without the relational scaffolding that makes it meaningful, they may process it the same way they process a spell-check suggestion: accept, reject, move on. No reflection. No growth.
Fan et al. (2025) documented exactly this pattern. Students using ChatGPT for writing skipped the metacognitive steps, evaluation, orientation, monitoring, that make revision productive. The essay improved. The student didn’t. Corbin et al.’s framework helps explain why: extra-recognitive feedback gives you the information without the relational context that makes you actually wrestle with it.
What This Reveals About Human Feedback Too
One of the strengths of this framework is that it doesn’t only diagnose AI feedback. It also exposes problems in existing human feedback practices.
Consider a teacher who provides cursory tick marks alongside a student report. Technically, that’s recognitive feedback, human to human. But the framework reveals it as defective. The student recognized the teacher’s authority by submitting work in good faith. The teacher failed to reciprocate. Following Honneth, Corbin et al. argue this constitutes a form of misrecognition, a kind of disrespect that can damage student self-esteem.
The same logic applies to peer feedback. Programs that try to replicate teacher-student dynamics often fail because students don’t recognize their peers as having the authority to evaluate their scholarly development. Successful peer feedback, the authors suggest, requires explicitly establishing mutual recognition between participants as co-developing scholars.
This diagnostic power is what elevates the paper beyond a simple “AI can’t do what humans do” argument. The authors write that “like the proverbial fish discovering water, considering feedback systems that can replicate content but not provide recognition has helped us articulate what was always present but often unexamined in human feedback practices. our recognition-theoretical lens helps to identify why certain feedback approaches succeed or fail.” (p. 727). I find that genuinely useful as a lens for thinking about feedback quality across the board.
GenAI as a Feedback Sandbox
Corbin et al. don’t dismiss GenAI feedback, and neither should we. Their most practical contribution is the idea that extra-recognitive feedback is well suited to specific contexts: checking formatting, flagging citation errors, offering initial reactions to draft arguments. These tasks don’t need a recognitive relationship, and AI handles them without taxing a teacher’s time or patience. Students can seek this kind of feedback at any hour, repeatedly, without anxiety about burdening anyone.
The authors also propose that GenAI can serve as a pedagogical “sandbox.” Students can experiment with ideas, test arguments, and receive low-stakes feedback before moving into the higher-stakes recognitive space of teacher or peer interaction. That sandbox function could build confidence and prepare students for the vulnerability that genuine feedback demands.
“The optimal integration of GenAI feedback may actually enhance human feedback relationships,” Corbin et al. write. “By offloading certain extra-recognitive tasks to AI systems, educators can invest more fully in building the recognitive relationships that support deep learning” (p. 728).
I agree with that, and I think it connects to something Corbin and his colleagues argued in an earlier paper (Corbin, Dawson, Nicola-Richmond, and Partridge, 2025) on AI and assessment as a wicked problem. There’s no clean technical fix. The question is always pedagogical: what kind of learning are we trying to support, and what role should each tool play in that process?
The Pedagogical Bottom Line
If this framework is right, then the worst move institutions can make is positioning GenAI as a feedback substitute. The impulse to scale up AI feedback to fix student-to-staff ratios sounds efficient, but it risks replacing the very thing that makes feedback effective.
The better approach is strategic. Use GenAI for the extra-recognitive tasks it handles well. Free up human time and energy for the recognitive interactions that build students into scholars. And teach students the difference between the two, so they know what to expect from each source and how to use both productively.
The technology will keep improving. The relational need won’t go away.
References
- Brandom, R. B. 2019. A Spirit of Trust: A Reading of Hegel’s Phenomenology of Spirit. Cambridge, MA: Belknap.
- Corbin, T., Tai, J., & Flenady, G. (2025). Understanding the place and value of GenAI feedback: A recognition-based framework. Assessment & Evaluation in Higher Education, 50(5), 718–731. https://doi.org/10.1080/02602938.2025.2459641
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
- Henderson, M., T. Ryan, and M. Phillips. 2019. “The Challenges of Feedback in Higher Education.” Assessment & Evaluation in Higher Education 44 (8): 1237–1252. doi:10.1080/02602938.2019.1599815.
- Honneth, A. 1996. The Struggle for Recognition: The Moral Grammar of Social Conflicts, edited by J. Anderson. Cambridge, MA: MIT Press.
- Luo, J. (2025). How does GenAI affect trust in teacher-student relationships? Insights from students’ assessment experiences. Teaching in Higher Education, 30(4), 991–1006. https://doi.org/10.1080/13562517.2024.2341005
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646