← शोध समाचार पर लौटें
Editorial cover for a higher-education experiment with GPT-4 feedback
जर्नल पेपरPeer-reviewed study202617 जुल॰ 2026· 3 min

GPT-4 feedback increases student activation and learning outcomes in higher education

Stephan Geschwind, Johann Graf Lambsdorff, Deborah Voss, Veronika Hackl

International Journal of Artificial Intelligence in Education

500-शब्द सार

Editorial cover for a higher-education experiment with GPT-4 feedback

Geschwind, Lambsdorff, Voss, and Hackl investigate whether generative AI can make individualized feedback more scalable in large university courses. Their 2026 open-access IJAIED article reports a lab-in-the-field experiment conducted in undergraduate macroeconomics tutorial classes over one semester. Students answered eight open-ended questions and experienced one of three feedback arrangements: classroom-level lecturer feedback, additional individual feedback from peers, or additional individual feedback generated by GPT-4. The study asks whether the feedback changes student activation and the quality of subsequent answers, rather than only whether students say they like AI.

All groups received lecturer discussion of a sample solution and adaptive feedback on two selected student answers. In the peer-feedback condition, students anonymously rated another student's answer for content and style and wrote suggestions for improvement. In the AI-feedback condition, GPT-4 provided individual scores and qualitative suggestions in a comparable structure. Both individual-feedback formats were designed around three familiar feedback functions: looking back at current performance, clarifying the goal through a sample solution, and pointing forward to improvement.

The researchers operationalized activation in two ways. First, they tracked voluntary participation across the eight tasks; tutorial participation did not count toward the final examination, although students who completed at least seven tasks entered a raffle. Second, they used answer length as an indicator of the effort invested in a task. For learning outcomes, GPT-4 rated the content and style of answers on five-point scales. To make treatment comparisons more consistent, the model rated answers from all three conditions after the course in three separate iterations, and the researchers averaged those ratings.

The AI-feedback condition produced the clearest activation pattern. Participation remained higher over time than in the lecturer-only condition, while peer feedback performed only marginally better than lecturer feedback. Students receiving AI feedback also wrote the longest answers. Because attrition was voluntary and not random, the authors did not rely only on simple averages: their task-to-task analysis used 479 repeated comparisons in which the same student participated in the same condition on consecutive relevant tasks.

For learning outcomes, AI feedback was associated with the strongest improvement in content ratings. Peer feedback did not produce the same advantage, which the authors connect to differences in reliability and completeness: peers sometimes failed to provide feedback, while the AI returned it consistently. The study found no comparable treatment effect on writing style. That distinction matters because it suggests that the intervention supported substantive engagement with macroeconomic answers without automatically improving every dimension of academic writing.

The paper is promising but should be interpreted carefully. The setting was one subject area and one university course, participation was voluntary, attrition differed across conditions, and the outcome measure was improvement in open-ended answers rather than a broad examination of long-term retention or transfer. GPT-4 generated feedback in one condition and also served as the common post-course rater, even though the authors used repeated ratings and prior reliability work to strengthen the measure. These choices make the study more informative than a satisfaction survey, but they do not establish that AI feedback will outperform expert human feedback in every context.

For AIEDHK, the practical lesson is to focus on feedback system design. Individual AI feedback may sustain participation when it is timely, structured, connected to a sample solution, and followed by opportunities to revise. Educators should preserve lecturer oversight, test feedback validity, measure learning with independent assessments, and compare the intervention with realistic alternatives. The contribution is not a claim that GPT-4 replaces teachers. It is evidence that carefully structured AI feedback can extend the reach of tutorial support while keeping course goals and human judgment visible.

संबंधित पेपर

A teaching assistant, instructor and university student discuss two proposed feedback comments beside a probability tree for two coin tosses.
कॉन्फ्रेंस पेपर2025
कॉन्फ्रेंस पेपर 210

Improving Generative AI Student Feedback: Direct Preference Optimization with Teachers in the Loop

Juliette Woodrow, Chris Piech, Sanmi Koyejo

Proceedings of the 18th International Conference on Educational Data Mining

Teaching assistants' choices can help adapt AI feedback to a particular course. In a university probability course, researchers trained a model using comments that staff selected, edited, or wrote themselves. Experts preferred its feedback to GPT-4o in a blind comparison, while classroom staff still frequently wrote their own responses. The study demonstrates a feasible feedback workflow and stronger teacher preference, but it did not test whether students learned more or used the feedback more effectively.

Teacher AgencyAI FeedbackHigher Education
500-शब्द सार पढ़ें →
A student drafts an essay at a window ledge while a peer points to an argument map and a lecturer observes.
जर्नल पेपर2026
जर्नल पेपर 202

Teacher feedback VS AI-assisted peer feedback in L2 writing: A quasi-experimental study in a Chinese University

Jinfeng Tang, Ping Li, Rong Luo

PLOS One

An eight-week study compared teacher comments with peer feedback supported by DeepSeek in two university English classes. Both approaches supported better revised essays, but their improvement patterns and feedback focus differed. An initial vocabulary-diversity advantage for the AI-peer class did not persist into the second task, and sentence-complexity results were limited. Because intact classes received different teaching packages, the findings cannot separate AI's contribution from peer discussion or other class differences.

Peer LearningWriting EducationAI Feedback
500-शब्द सार पढ़ें →
A university student revises an essay using an evidence-to-claim sketch while a lecturer and peer discuss the feedback.
जर्नल पेपर2025
जर्नल पेपर 190

Teacher, peer, or AI? Comparing effects of feedback sources in higher education

Joshua Weidlich, Flurin Gotsch, Kai Schudel, Claudia Marusic-Würscher, Jennifer Mazzarella, Hannah Bolten, Dario Bütler, Simon Luger, Bettina Wohlfender, Katharina Maag Merki

Computers and Education Open

Students can welcome feedback without producing stronger revisions. A randomized university study compared teacher, peer, and language-model feedback while concealing its source from recipients. AI and peer comments were perceived as fairer than teacher comments, but teacher feedback supported better scientific argumentation than AI feedback. The findings separate the experience of receiving comments from the quality of the work that follows, with implications for evaluating automated feedback tools.

Higher EducationAI FeedbackPeer Learning
500-शब्द सार पढ़ें →