Enhancing learner-centered feedback with AI: teachers'' practices and perceptions

Synthesis: 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.

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 judgment to accept or reject suggestions — the ML functioned as a diagnostic assistant, not an autopilot.

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.
• 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 judgment.

Interpretation

• The tools were most valuable as reflective prompts that surface overlooked aspects of feedback — "making invisible gaps visible" — scaffolding teachers' evaluative judgment, rather than as providers of finalised text (an "assist but verify" pattern; cf. Human-in-the-Loop).
• 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 judgment.

What this means for practice

• Instructors. Use the ML component list as a coverage check before sending feedback: 20 of the 21 teachers omitted "Meeting Learning Objective" and 14 omitted "Student–Teacher Relationship," and those were the two components the model most often surfaced.
• Instructors. Budget time for post-editing instead of expecting the tool to finish the job — 12 of 21 teachers rewrote the ChatGPT text, mostly editing (f = 32) and deleting (f = 27), and several reported that revising the output took as long as writing the feedback themselves.
• Instructors. Expect to strip inflated praise and generic encouragement by hand: revisions clustered in the relational dimension of feedback (Student–Teacher Relationship f = 24; encouragement f = 11), and teachers systematically moderated tone to keep their professional voice.
• Faculty developers. Design training around evaluative judgment, not tool operation: less-experienced teachers valued the tool's coverage and speed, while teachers with more than five years of experience flagged tone, editing burden, trust, and misinformation risk — the authors warn that novices who defer to suggestions may build less independent judgment.
• Faculty developers. Build a review step for relational language into any AI feedback pilot, since teachers treated the affective dimension as the part that cannot be delegated.

Limitations

• Only 21 teachers were studied, and they worked in a controlled environment on a single simulated task — giving feedback on one three-minute recorded student self-introduction, not feedback in their own courses.
• The study measured teacher interaction and perception only; no student outcomes or student perspectives were collected, so the effect of these tool-supported feedback practices on learners is untested.
• Several findings rest on very small counts — the 100% acceptance of Affirmation and Encouragement suggestions reflects 6 cases, and the challenge codes have a denominator of 11 teachers — so the components of the "assist but verify" pattern vary substantially in evidential weight.

Citation

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