Ahmad Ari Aldino, Bhagya Maheshi, Yuheng Li, Ying Zhou, Yi-Shan Tsai, Dragan Gašević & Guanliang Chen (2026) — Assessment & Evaluation in Higher Education 51(5), 892–910. doi:10.1080/02602938.2026.2638920.
đź“„ Full text (Taylor & Francis, OA)
Summary
An empirical study of 21 higher-education teachers using PolyFeed, an AI-powered feedback tool combining (1) a BERT-based ML model (from Aldino et al. 2024) that detects which learner-centered feedback components are missing from teacher-written feedback and suggests them, and (2) ChatGPT-4o mini to rephrase/enhance the teacher's draft. Teachers gave feedback on a simulated student presentation, then used the tool, then were interviewed. The study answers two questions: how teachers interact with AI feedback tools (RQ1) and how they perceive them (RQ2). Framework: Ryan et al.'s (2023) learner-centered feedback dimensions — Future Impact, Sensemaking, Agency.^[raw/papers/tandf-2026-learner-centered-feedback-ai.md]
RQ1 — How teachers interacted with the tools
ML suggestion acceptance (detection → adoption)
- The ML model most often flagged missing Meeting Learning Objective (20 of 21 teachers omitted it; 16 accepted the suggestion) and Student–Teacher Relationship (14 omitted; 12 accepted).
- Affirmation and Encouragement was missing in 6 cases but accepted by all 6 teachers (100%).
- Teachers used professional judgement to accept or reject suggestions — the ML functioned as a diagnostic assistant, not an autopilot.^[raw/papers/tandf-2026-learner-centered-feedback-ai.md]
Revision of ChatGPT-enhanced feedback
- 12 of 21 teachers made further sentence-level revisions; the rest left the text unchanged.
- Most common actions: editing (f = 32) and removing (f = 27); adding was rare (f = 8).
- The dominant pattern was calibrating tone: Edit–Praise (f = 11, e.g. "First of all, I want to congratulate you on completing your assignment!" → "Well done on completing your assignment"), Remove–Suggestion (f = 9), Remove–Encouragement (f = 8), Remove–Praise (f = 6), Edit–Correction (f = 9). Teachers systematically moderated exaggerated praise and generic suggestions to protect authenticity and professional voice.^[raw/papers/tandf-2026-learner-centered-feedback-ai.md]
- Revisions clustered most strongly in the Agency dimension — especially Student–Teacher Relationship (f = 24; encouragement f = 11) — confirming that the relational/affective dimension of feedback resists AI delegation.
RQ2 — Teacher perceptions
- Benefits: promotes reflection (n = 14), improves language and structure (n = 11), identifies missing components (n = 10), saves time (n = 2).
- Challenges: need for human editing (n = 9), inconsistent tone (n = 7), potential misinformation (n = 5), trust issues (n = 5).
- Experience gap: teachers with >5 years of experience reported more challenges (tone, editing burden, trust, misinformation); less-experienced teachers valued scaffolding benefits (identifying missing components, saving time). The authors flag a developmental risk: novice teachers who defer to AI suggestions may build less independent feedback judgement.^[raw/papers/tandf-2026-learner-centered-feedback-ai.md]
Interpretation
- The tools were most valuable as reflective prompts that surface overlooked aspects of feedback — "making invisible gaps visible" — scaffolding teachers' evaluative judgement, rather than as providers of finalised text (an "assist but verify" pattern; cf. human-in-the-loop-ai).
- Design implications: adjustable tone parameters, discipline-specific templates, and transparency controls; without them AI adoption may increase rather than reduce workload and can erode teacher authority if positioned as autonomous providers.
- The authors call for future research on student perspectives of AI-assisted teacher feedback and longitudinal study of whether reliance strengthens or displaces evaluative judgement.^[raw/papers/tandf-2026-learner-centered-feedback-ai.md]
Related Pages
- agency-gap-ai-writing — Calibrating AI initiative to learner/teacher needs
- feedback-futures-genai — Editorial citing this study's "reflective scaffold" reading and workload redistribution
- genai-teacher-feedback-comparison — Student-side complement: perceptions of AI vs teacher feedback
- chatgpt-feedback-engagement-genai — Student-side engagement with ChatGPT feedback
- feedback-loop — Feedback as dialogic process; the framework used here (Future Impact / Sensemaking / Agency)
- teacher-role — Teacher agency and professional judgement in AI-mediated feedback
- human-in-the-loop-ai — "Assist but verify" pattern; teachers accept/reject/edit AI output
- scaffolding — AI as scaffold for evaluative judgement, especially for novice teachers
- faculty-development-genai — Professional development implications (de-skilling risk)
- over-reliance — Novice-teacher deferral risk
- higher-ed — Deployment context
Citation
APA: Aldino, A. A., Maheshi, B., Li, Y., Zhou, Y., Tsai, Y.-S., Gašević, D., & Chen, G. (2026). Enhancing learner-centered feedback with AI: Teachers' practices and perceptions. Assessment & Evaluation in Higher Education, 51(5), 892–910. https://doi.org/10.1080/02602938.2026.2638920